Emotionally disturbing images can significantly impair people’s ability to stay focused, according to a new study published in Behavior Research Methods researchers found that negative emotional distractions — such as images of distressed people or threatening animals — disrupted participants’ sustained attention more than neutral or positive visuals These negative images also led to more negative feelings and were more likely to be remembered afterward Sustained attention is a basic mental function that allows people to maintain focus over time It’s central to everyday activities such as reading or working and is often impaired in conditions like depression While past research has explored how internal factors like motivation or fatigue can interrupt sustained attention far less is known about how external distractions — especially emotional ones — influence this process researchers lacked a reliable way to test how these emotional distractions affect the ability to concentrate over extended periods and associate professor at Boston University Chobanian & Avedisian School of Medicine “While upsetting thoughts and experiences can disrupt one’s ability to focus attention while performing everyday tasks translating this phenomenon to the laboratory has remained elusive We were inspired to design a paradigm that could capture this experience in the lab so we could better study it.” The researchers developed a new version of a well-established attention task called the gradual-onset continuous performance task (gradCPT) dubbed the “emogradCPT,” involved a neutral digit-based attention task overlaid on emotional images The goal was to see how the emotional content of these background images affected people’s ability to focus Participants in the first experiment included 64 individuals who were asked to perform a 9.6-minute task where they had to respond to a stream of digits pressing a button for every number except the digit “3.” These digits were presented one after the other with gradual visual transitions task-irrelevant background images were continuously displayed the images were not related to the task and were not supposed to be attended to but they remained visible throughout the task duration the researchers ensured that the images significantly differed in how negative or arousing participants perceived them to be Negative images included scenes such as injured animals or distressed people while positive images showed smiling babies or cute pets Neutral images were devoid of emotional content In addition to tracking accuracy and reaction time the researchers incorporated self-report thought probes that asked participants about their focus and emotional state after each block of images participants completed a surprise memory test to assess how well they remembered the images even though they were told during the task to ignore them The researchers found that negative background images impaired attention more than positive or neutral ones Participants made more errors and had slower reaction times during blocks with negative images They also reported feeling worse and being more distracted in those blocks average accuracy during negative image blocks was significantly lower than during positive or neutral blocks and reaction times were slower by around 20 to 30 milliseconds These effects weren’t just fleeting impressions Participants were more likely to remember negative images in the surprise recognition test conducted 30 minutes after the task even though the images were never directly relevant to their goal The fact that these images left a stronger memory trace suggests they captured attention more effectively “This is the first study to show that when people are sustaining attention distractions that are upsetting or unpleasant are most likely to disrupt that focus,” Esterman told PsyPost “These kinds of emotionally negative distractions also make us feel worse and we are more likely to remember them later,” The researchers also found that people’s subjective experience matched their performance When participants reported being more distracted or feeling worse These associations held across multiple blocks of the task providing further validation for the emogradCPT as a tool to measure the impact of emotional distraction “I was genuinely surprised that the more upsetting distractors were more likely to be remembered even when we told participants to actively try to ignore the distractors,” Esterman said To further confirm the reliability of the findings a second experiment was conducted with 50 new participants the researchers replaced the neutral images with blank backgrounds to test whether any visual image — even a neutral one — was distracting negative images disrupted attention more than either positive or blank backgrounds The results from this second group largely matched those from the first experiment a subset of participants performed the task again inside a brain scanner Performance on the emogradCPT remained consistent providing evidence that the task could be used in future neuroimaging studies to understand how emotion and attention interact in the brain these results highlight that emotionally negative distractions have a measurable and lasting impact on attention Unlike many tasks that measure short bursts of attention the emogradCPT captures how focus is maintained over time — a more naturalistic reflection of real-world demands such as staying attentive during a long drive or focusing at work while surrounded by emotional news or social media The researchers argue that their new task offers a useful tool for understanding how emotional content disrupts focus in both healthy individuals and clinical populations People with anxiety or trauma-related conditions often show stronger reactions to emotional material and this task could help researchers investigate whether they are particularly vulnerable to distraction by emotionally negative cues The set of images used was specific to the current experiments and may not generalize to other visual materials or contexts Participants might have become aware of the blocked structure of the task which could influence both performance and self-reports it could be improved with additional testing sessions or more focused experimental designs is to determine if distractibility as measured by this new task (the emogradCPT) is related to clinical conditions (like PTSD and ADHD) as well as distractibility in more real-world settings,” Esterman noted the emogradCPT offers a new way to investigate how emotion and attention interact in a sustained manner It may prove especially valuable for clinical research aimed at understanding emotional biases and attentional control in people with psychological disorders Future studies could also adapt the task to explore personalized distractions — such as substance-related cues in addiction — or use eye-tracking and brain imaging to pinpoint the exact cognitive processes involved and our phone lighting up all day long,” Esterman said “This study represents a new tool that could help characterize the ability to resist these types of everyday distractions.” “We believe this study will help scientists measure how distractible a person is and whether those distractions intrude in their memories We also believe it can open new windows to studying attention in clinical populations and their neural mechanism alongside brain imaging both of which are directions we are currently pursuing.” we hope these findings will assist in characterizing and treating anxiety and posttraumatic stress disorders,” he explained “We are currently pursuing a VA-funded study using these tasks to better understand PTSD.” The study, “Characterizing the effects of emotional distraction on sustained attention and subsequent memory: A novel emotional gradual onset continuous performance task,” was authored by Michael Esterman, Sam Agnoli, Travis C. Evans, Audreyana Jagger‑Rickels, David Rothlein, Courtney Guida, Carrie Hughes, and Joseph DeGutis. An online survey of adults in the U.K. found that frequent earworms were linked to a broad range of mental and motor habits. These findings hint that earworms might be mental echoes of a habit-prone brain. New research suggests that the shift from handwriting to digital tools in early education may come at a cost. In an experiment with 5-year-olds, those who practiced writing by hand showed better letter naming, spelling, and word reading than those who used keyboards. Taller students tend to score slightly higher on standardized tests than their shorter classmates, according to a new study of New York City public schools. A new study finds preschoolers’ brains respond differently when hearing a story read aloud versus hearing it from a screen, highlighting how live reading engages social brain networks more strongly than solitary screen time. A major study suggests that widely used medications like paracetamol and ibuprofen may influence cognitive performance in subtle ways. The research introduces a “cognitive footprint” model to estimate how small effects scale across populations. Scientists have discovered that moving to a beat may prepare your brain to hear speech more clearly in chaotic environments. Training the brain and body together may help older adults perform better, even when mentally drained, new research shows. Please enter your username or email address to reset your password. 2025 10:41 PM EDTOlivia “Livvy” Dunne didn’t just show up to her first Kentucky Derby — she made it a fashion moment One day after rocking an all-pink ensemble for the Kentucky Oaks the former LSU gymnast switched up her look with a vintage-inspired polka dot dress ahead of the 151st running of the Kentucky Derby at Churchill Downs paired the black-and-white outfit with an oversized black hat fit for any race-day red carpet Dunne posted behind-the-scenes photos of her Derby look to Instagram and documented her day at the track "Suddenly I’m a Kentucky Derby fan." Another got into the spirit of the evening "You won the derby," another viewer proclaimed It was a full-circle weekend for the social media star who just capped her collegiate gymnastics career in April when LSU placed third in the NCAA semifinals she served as the official “riders up!” announcer for the Kentucky Oaks — a first for the athlete “Your girl’s first derby :)” With a look this polished — and timing this smart — it’s clear Dunne is already thinking several steps ahead of the competition. By Rachel Dillin is a trending news writer for Men's Journal She's a lifelong journalist who covers entertainment and celebrity news Rachel Wolf is CEO of Public First and an author of the 2019 Conservative Party manifesto Feel like I’ve been gaslit and lied to and now I’m thinking I’m going Reform because they haven’t had a shot they can’t do any worse than what the past governments have done.” – Woman I hope no one reading this is surprised by Reform’s results in last week’s local elections They are the predictable result of the failures of the last decade There is no new magic Reform voter and no new problem politicians must figure out how to tackle They are the same people who “surprised” us in Brexit and the change they wanted didn’t happen they were the reason the Conservatives softened economic policy and abandoned austerity why Tees Valley Mayor Ben Houchen supported the nationalization of his airport why Labour had no problems with promising rail nationalization and why Nigel Farage is advocating to nationalize British steel They are the people we have been writing about for the last decade They are not protest voters – they have a very reasonable case for not wanting the incumbent parties – but they are understandably anti-political.  In the 2024 election for just a moment — bolstered by Tory failures — but it then immediately turned on them The problems these voters have are not with the Conservatives But let’s also have a reality check – those who think this is all about core economic policy or general “disillusionment” are kidding themselves They have consistently voted for a party that promised to lower it in every election since 2010 – and they have instead seen higher net migration is the only person who has any claim to consistency left (and it is Farage one of the reasons they dislike immigration is economic – they think it depresses wages – but it’s not the only reason and making them better off is not going to make this problem disappear What we mean by that is that they now dictate the opposition narrative If you want to change the terms of the debate It is also likely the case that voters will increasingly think of them as the opposition – which will in turn affect voting has come close to owning opposition to the government – Cut immigration and go after easy wins (such as international students) – Backtrack more on DEI and other perceived “woke” initiatives – Feel under even greater pressure on any investment for climate (including the electricity grid) – Try and reboot the Boris Johnson-era policy of “leveling up” – and they’ll focus on small-level improvements people will notice – Worry a lot about any pro-EU stances Reform will make the running on a much wider playing field than the government Farage is a gifted politician who will jump on any row — any inklings that the civil service and he will turn it into a national story that will run and run Reform will start facing more scrutiny.  Let’s not overdo this – Labour did just fine without much policy before the election – but any clear insanity will be noticed Reform will have to inch left on economics and avoid straying from net zero realism to climate denialism.  although we shouldn’t underestimate how unserious neither Labour nor the Conservatives can get anything done.  Nevertheless, Reform must balance being anti-establishment underdogs while leading the polls (and as of last week several councils) for nearly half a decade They’re not going to have it entirely easy.  The immigration white paper is just the start of a summer of pivots that this government will make – on energy you should pay attention to exactly the same people you needed to understand for the last decade The problem of the last few years isn’t that we made the wrong promises — it’s that we haven’t delivered on them In the face of huge challenges posed by coronavirus and Brexit the British government mustn’t forget why it won last year’s election the 43-year-old not only left her fans speechless but also caught the attention of fellow Olympian Simone Biles In a recent Instagram update the 23-time Grand Slam champion set new fashion goals as she showcased an ethereal look that drew a wave of positive reactions from fans Williams’ outfit featured a white off-the-shoulder long gown The cinched detailing accentuated her curves while the over-the-shoulder sash added a touch of elegance and drama to the ensemble straight hair that perfectly complemented her sophisticated look “Highlights of my night are on my TikTok!” she wrote The comment section was flooded with followers gushing over her pre-Met Gala look You look absolutely BEAUTIFUL,” a commenter added Stunning as always,” a social media user wrote Besides the fire and smiling face with heart-eyes emojis emojis others expressed their excitement to see the tennis legend in the annual fashion event Can't wait to see you tomorrow,” a fan commented Biles also chimed in on the comments and echoed a sentiment similar to what fans were saying playfully suggesting her enthusiasm to see Williams on the red carpet The 2025 Met Gala is set to feature top-tier athletes, including WNBA star Angel Reese Olympic track champion Sha'Carri Richardson and Biles The three are also part of the hosting committee Meanwhile, the event, scheduled for Monday, will be led by honorary chair LeBron James of the Los Angeles Lakers Formula 1 legend Lewis Hamilton is among the co-chairs Serena Williams during the halftime show of Super Bowl LIX this was not the first time Serena Williams attended the fashion event Williams strutted down the red carpet in a metallic gold Balenciaga gown styled with black gloves 2 overall pick of the NFL Draft and passing on Travis Hunter in the process while also moving up in the fifth round taking Shedeur Sanders the Browns were also in the news circuit quite a bit earlier in the offseason when they finally agreed to a contract extension with Myles Garrett But in all of the chatter surrounding Cleveland this offseason some very intriguing players have gotten lost in the fog and perhaps the most severe case of that is defensive end Isaiah McGuire The Browns selected McGuire in the fourth round of the 2023 NFL Draft add after a quiet rookie campaign in which he did not play very much 2.5 sacks and three forced fumbles in 16 games last seaosn McGuire's playing time increased following the trade of Za'Darius Smith at the deadline to the point where he actually played in 90 percent of Cleveland's defensive snaps in Week 17 The 23-year-old also registered an 83.3 overall grade at Pro Football Focus in 2024 which ranked 11th among 211 qualifying edge rushers McGuire seems likely to take on more of a full-time role although Alex Wright—who appears to be recovered from a torn triceps injury—may push him for snaps upon his return McGuire seems like an obvious breakout candidate He will be playing on a defensive line that includes Garrett and incoming rookie Mason Graham which means he may have some free lanes to get into opposing backfields and rack up some sacks The Browns felt like they landed a steal when they nabbed McGuire on Day 3 of the draft two years ago and based on his production over the final two seasons of his collegiate career at Missouri not even Cleveland fans are talking about McGuire who displayed plenty of potential during the back half of 2024 and has very clear raw physical talents that could make him quite the threat in the trenches in 2025 and beyond Perhaps it's because of all the hoopla surrounding the Browns' first-round draft pick their quarterback situation and all of the Garrett news from months ago Maybe McGuire will begin to pick up steam as a sleeper candidate as we get deeper into the offseason and into training camp you just can't help but feel that Cleveland may have a defensive star in the making here MORE: Insider Clears the Air on Browns, Shedeur Sanders Conspiracy Theory MORE: Cleveland Browns LB Arrested on Sunday MORE: Cleveland Browns Star Tabbed Trade Candidate Amid Murky Future MORE: NFL Insider Drops Notable Intel on Browns' Shedeur Sanders Decision MATTHEW SCHMIDT The content on this site is for entertainment and educational purposes only Betting and gambling content is intended for individuals 21+ and is based on individual commentators' opinions and not that of Sports Illustrated or its affiliates All picks and predictions are suggestions only and not a guarantee of success or profit If you or someone you know has a gambling problem crisis counseling and referral services can be accessed by calling 1-800-GAMBLER activists called attention to the deadliest infectious disease in the world — tuberculosis Yes, tuberculosis is still around, and it’s killing more than 1.25 million people each year according to the World Health Organization Author and tuberculosis activist John Green said that’s more than malaria Tuberculosis is both preventable and curable, and active cases in the U.S. are relatively uncommon — though infection rates are climbing quickly So why do many people believe we got rid of tuberculosis already Many of the tuberculosis activists who met with their legislators for “TB Hill Day” in early April also weren’t aware of how much havoc tuberculosis still wreaks around the world — including some of those who survived tuberculosis themselves said that she has spent a lot of time explaining — and justifying — her illness to family I wish somebody famous could get TB so that they could speak out like we’ve seen other celebrities speak out for other diseases,’” Skaggs said John Green didn’t have a tuberculosis diagnosis “We got our prayers answered,” Skaggs said Green’s popularity surged in the 2010s with his hit young adult novels such as “The Fault in our Stars.” He is also known for his educational “Crash Course” videos which have become a mainstay in high schools and colleges across the country He’s been fundraising for Partners in Health for years — but his fascination with tuberculosis began in 2019 during a trip to Sierra Leone in Africa when he was brought to a tuberculosis clinic and met Henry Reider Green and Reider’s friendship took off the more Green saw the gap in treatment quality between Sierra Leone and the U.S Reider couldn’t afford accurate testing and struggled through trial-and-error treatment that dragged on for so long that his lymph nodes swelled large enough to puncture the skin in his neck Green processed everything he was learning about tuberculosis the best way he knew how — by writing what would eventually become “Everything is Tuberculosis,” which was released in March He also started talking about it in his YouTube videos and on his podcast as “an illness that walks the trails of injustice and inequity that we blazed for it,” as he would later describe it in his book Green is known for the large online community he and his brother Hank gained from their YouTube channel His followers call themselves “Nerdfighters.” When Green started talking about tuberculosis in his YouTube videos and on his podcast he said he had no idea that people would respond so generously They soon became a part of the work to raise awareness about tuberculosis In 2023, some of the most dedicated activists started organizing themselves and formed “TB Fighters.” Some of these “TB Fighters” had very little experience with activism Organizer Hannah Kenny described the early TB Fighters as “a lot of very confused but very excited people” and “a bunch of bees in a trench coat.” hundreds of TB Fighters were attending regular phone banking sessions and writing to their representatives Green took things public in 2023, with videos calling out companies Danaher and Johnson & Johnson for what he calls “price gouging” for tuberculosis testing and medication Hundreds more Nerdfighters began calling and writing to their representatives Within days, Johnson & Johnson announced that it would not renew its patent on a key antibiotic — allowing more affordable generic versions into the market Danaher agreed to lower the prices of its cheapest tests but fell far short of what Doctors Without Borders has campaigned for The companies did not cite what prompted their moves Danaher did not respond to WTOP’s request for comment Tests and medications themselves aren’t the only barriers to treatment war and other diseases have ravaged communities and health care systems in poor countries He also delves into how tuberculosis has been imagined — at times romanticized — and how those imaginings have influenced how we have collectively understood the disease He explores how all these historical forces have come together to shape the current state of affairs and the way the wealthy world addresses tuberculosis today — namely “We know how to live in a world without tuberculosis,” Green writes “But we choose not to live in that world.” Tuberculosis is both preventable and curable — and it’s not nearly as difficult or expensive as one might think by 2030 we could drive down mortality 90%,” said Vincent Lin associate director of health and policy at Partners in Health That amounts to about $130 per year per taxpayer “So we’re not talking about some moonshot very difficult thing to achieve,” Lin said “This is very much within our grasp if we’re able to mobilize the resources to do so.” “Hundreds of thousands of people have their treatment interrupted,” Green told WTOP, citing research from the World Health Organization “Most of the people who’ve had their treatment interrupted will die of tuberculosis.” Green encourages readers to empathize with the people whose lives and problems are physically far away from our own But he also makes a domestic case for treating tuberculosis abroad pointing to COVID-19 as a recent example of how quickly a disease can spread “So we have to be able to deal with these infectious diseases Tuberculosis cases are rising quickly in the U.S. and Lin said those numbers will only climb if action isn’t taken both at home and abroad Even after her four-month isolation period She had to go back on medical leave and was fired soon after “I am literally the best-case scenario,” Skaggs told WTOP I had disability insurance to make sure I could still pay my bills So many people have it so much worse than I do.” Drug-resistant strains are even harder to treat Lin said patients with drug-resistant tuberculosis could end up in treatment for years sometimes taking dozens of pills a day and getting “toxic injectable drugs with a lot of side effects.” An internal memo by U.S. Agency for International Development, obtained by the New York Times estimated that cuts made to its programming could lead to a 28-30% increase in tuberculosis cases over one year globally “TB anywhere is a threat to people everywhere,” Green said Over 200 more activists attended this year’s Hill Day than last year meeting with their representatives from 49 states “One of the congressional staff that we met with told one of the volunteers ‘You guys are as prepared for this meeting as folks who represent some of the biggest defense contractors in the nation,'” Lin said TB Fighters are already planning for the next Hill Day — working to push Congress to renew its pledge to the Global Fund to Fight AIDS, Tuberculosis and Malaria. “That replenishment pledge is this fall and it sets up the funding for the next three years,” Lin said “That is a critical thing that we … have been talking to members of Congress about — to say that this funding is important that we think this is a very worthwhile use of our taxpayer dollars.” Green will be on tour for several more weeks using his platform to continue raising awareness of this issue “How we spend our attention matters,” Green said “So to see people using their attention to focus on tuberculosis has been very moving for me and very encouraging we can accomplish things that we simply can’t accomplish alone.” Get breaking news and daily headlines delivered to your email inbox by signing up here This website is not intended for users located within the European Economic Area and especially with the feeling I had in my shoulder' says time trial world champion and double Olympic champion “I can look back on these last couple of weeks with satisfaction I took two wins since returning to competition and found the rhythm again," Evenepoel said.  "Not everything went as I had wanted and hoped in some of the races but it’s only normal after so much time off the bike I will now recover a bit and then continue my preparations for the summer." The 24-year-old was forced to take almost six weeks off the bike to recover after being doored by a driver in training on December 3.  He returned to racing with a revised schedule at De Brabantse Pijl and took the victory before finishing third at Amstel Gold Race and ninth at Flèche Wallonne.  the World and Olympic time trial champion ended the Tour de Romandie with the stage 5 victory as we didn’t have the rain that was initially predicted The first four kilometres of the race were rather technical and then everything was pretty straightforward until the finish line," Evenepoel said He covered the 17.1km course in nearly 50 kph in a time of 20:33 beating overall winner João Almeida (UAE Team Emirates-XRG) by 11 seconds in Genève he said he was pleased with how his shoulder felt during the effort and especially with the feeling I had in my shoulder in the corners and on the straight lines," Evenepoel said Evenepoel is expected to compete at Critérium du Dauphiné and then the Tour de France a race he finished in third overall in 2024 Kirsten has a background in Kinesiology and Health Science She has been involved in cycling from the community and grassroots level to professional cycling's biggest races She began her sports journalism career with Cyclingnews as a North American Correspondent in 2006 Kirsten became Women's Editor – overseeing the content strategy race coverage and growth of women's professional cycling – before becoming Deputy Editor in 2023 you will then be prompted to enter your display name Metrics details Skin lesions remain a significant global health issue with their incidence rising steadily over the past few years Early and accurate detection is crucial for effective treatment and improving patient outcomes This work explores the integration of advanced Convolutional Neural Networks (CNNs) with Bidirectional Long Short Term Memory (BiLSTM) enhanced by spatial and temporal attention mechanisms to improve the classification of skin lesions The hybrid model is trained to distinguish between various skin lesions with high precision the CNN (original architecture) with BiLSTM and attention mechanisms model achieved the highest performance and Matthews Correlation Coefficient (MCC) of 91.55% The proposed model was compared to other configurations including CNN with Gated Recurrent Units (GRU) and attention mechanisms to highlight the efficacy of the proposed approach This research aims to empower healthcare professionals by providing a robust diagnostic tool that enhances accuracy and supports proactive management strategies The model’s ability to analyze high-resolution images and capture complex features of skin lesions promises significant advancements in early detection and personalized treatment This work not only seeks to advance the technological capabilities in skin lesion diagnostics but also aims to mitigate the disease’s impact through timely interventions and improved healthcare outcomes ultimately enhancing public health resilience on a global scale characterized by abnormal growths on the skin that can lead to severe health consequences if not detected early The incidence of skin lesions has been steadily rising in recent years highlighting the critical need for effective detection and treatment strategies Early and accurate diagnosis is essential for enabling timely intervention and personalized treatment plans which in turn improve patient outcomes and reduce the overall healthcare burden particularly through advanced Convolutional Neural Network (CNN) techniques has emerged as a promising tool in medical imaging and diagnostics These models have the ability to analyze high-resolution images of skin lesions distinguishing between benign and malignant tissues with remarkable accuracy This capability is vital for early detection and holds the potential to revolutionize the way skin lesions are diagnosed and treated ensuring that patients receive the care they need at the earliest possible stage A recent research3 indicates a significant expansion in the global dataset for skin lesion images used for training ML models Datasets from repositories such as the International Skin Imaging Collaboration (ISIC) and DermNet NZ have provided researchers with a wide array of skin lesion images covering diverse types and pathological conditions researchers have developed sophisticated CNN architectures which have shown outstanding performance in classification tasks This work aims to advance the field of skin lesion classification by leveraging a hybrid model that integrates CNN with Bidirectional Long Short Term Memory (BiLSTM) This innovative approach is designed to capture the complex features and variations in skin lesion images more effectively than traditional models By elucidating the nuances of different skin lesion types this research endeavors to empower healthcare professionals with robust tools for accurate diagnosis and proactive management strategies this research holds promise in mitigating the impact of skin lesions through timely intervention and personalized care By improving diagnostic accuracy and enabling more effective treatment plans it aims to enhance individual health outcomes and bolster public health resilience on a global scale Integrating advanced machine learning techniques with comprehensive datasets represents a significant step forward in the fight against skin lesions offering hope for better management and control of this widespread health issue The motivation to develop a model that combines CNNs with BiLSTM networks and multiple attention mechanisms stems from the complex nature of the dermatological diagnosis Skin lesions can vary greatly in appearance making it essential to capture both spatial and temporal information CNNs excel at identifying detailed spatial features are adept at handling temporal dependencies BiLSTMs enable the model to consider the sequence of features within each image preserving contextual information across spatial dimensions Integrating these architectures allows for a more comprehensive analysis of skin lesions capturing their intricate details and progression attention mechanisms enhance the model’s interpretability by highlighting the most relevant regions and features providing clinicians with insights into the decision making process This approach seeks to reconcile the high accuracy of AI models with the necessity for transparency and trust in medical applications ultimately enhancing diagnostic precision and fostering better clinical outcomes The major contributions of this work are listed below An integrated model using Convolutional Neural Networks (CNNs) with Bidirectional Long Short Term Memory (BiLSTM) enhanced by spatial and temporal attention mechanisms to improve the classification of skin lesions has been proposed An exhaustive experimentation of various approaches along with conventional CNN was performed and reported using several performance metrics The manuscript structure is organized into different sections to provide a unique approach towards skin lesion classification comprising 2656 clinical images across six distinct conditions Their investigation highlighted the efficacy of transfer learning with Inception-ResNet-v2 demonstrating superior performance in classifying skin diseases challenges such as dataset imbalance and clinical overlap among disease categories underscored the need for improved data management strategies and refined model training approaches to mitigate classification inaccuracies specifically ResNet152 and InceptionResNet-V2 This approach aimed to enhance the discriminative feature learning capabilities of the models challenges associated with dataset specificity and the computational demands of fine-tuning deep networks were acknowledged as potential limitations highlighting the need for further advancements in algorithmic efficiency and model generalizability integrating ResNet-18 with the LIME framework to provide visual justifications for predictions thereby enhancing trustworthiness in clinical settings the reliance on a single dataset and occasional inconsistencies in LIME’s explanations pose limitations these studies underscore the potential and challenges of deploying DL models in dermatological diagnostics emphasizing computational efficiency and accuracy enhancement through optimized feature processing Despite achieving high performance metrics on the HAM10000 dataset limitations related to dataset scarcity and computational resource demands were highlighted as areas necessitating further exploration and refinement emphasizing the critical role of diverse and expansive datasets in improving model robustness and accuracy and model interpretability were identified as crucial avenues for future research to enhance the clinical applicability and reliability of AI-driven diagnostic tools in dermatology addressing dataset imbalance and achieving superior performance metrics on the challenging HAM10000 dataset the study acknowledged computational complexities associated with ensemble learning and the need for further research to optimize model scalability and generalizability across diverse clinical scenarios combining lesion boundary segmentation with deep learning-based classification performed by a full-resolution convolutional network (FrCN) enhances feature extraction for classifiers like Inception-v3 the framework improved classification performance with Inception-ResNet-v2 achieving notable F1-score gains for benign and malignant cases Weighted accuracies reached up to 89.28% across datasets This highlights the value of integrating segmentation and classification for accurate skin lesion analysis 17 addresses the growing prevalence of skin diseases which pose psychological and physical risks Due to limitations in visual resolution and the subjectivity of manual diagnosis a computer-aided diagnostic framework is proposed using ResNet152 and InceptionResNet-V2 with a triplet loss function The framework maps input images into Euclidean space to compute L-2 distances for learning discriminative features These features are then used to classify skin disease images effectively The dataset consists of human facial skin disease images collected from a hospital in Wuhan Experimental results demonstrate the framework’s ability to improve the accuracy and efficiency of skin disease diagnosis which trains deep ensembles on 153 non-dermoscopic lesion images PECK combines convolutional neural networks using hierarchical learning to iteratively correct prediction errors the SCIDOG segmentation algorithm is proposed to detect lesion contours in noisy non-dermoscopic images without relying on data-intensive training SCIDOG enables precise lesion feature extraction enhances melanoma and benign nevi diagnosis on the MED-NODE dataset The proposed methods achieve superior diagnostic performance using 10-fold cross-validation surpassing prior state-of-the-art techniques 27 presents a fully automated method for classifying skin lesions from dermoscopic images addressing the challenge of distinguishing malignant melanomas from benign lesions The approach ensembles deep features from multiple pretrained convolutional neural networks (CNNs) Prediction probabilities from various models are fused to enhance classification accuracy the method achieves an AUC of 87.3% for melanoma and 95.5% for seborrheic keratosis surpassing top-ranked methods in simplicity and performance This demonstrates a reliable and robust solution for dermoscopic image analysis and skin lesion diagnosis a custom CNN architecture is employed in combination with BiLSTM units and Attention Mechanisms for skin lesion classification The rationale behind using CNNs in conjunction with RNNs lies in their complementary strengths: CNNs excel at capturing intricate spatial features in images are adept at analyzing sequential data and temporal patterns This combination ensures that both spatial and sequential aspects of the data are effectively utilized leading to more accurate and robust classification there were no prior experiments that combined CNN and RNN for this specific application and Attention Mechanisms in tandem is crucial as it highlights the potential of these advanced techniques to revolutionize medical diagnostics attention mechanisms are incorporated to further enhance the performance of the CNN-RNN combination Attention mechanisms allow the model to focus on the most relevant parts of the input data ensuring that critical features are given more weight during the analysis This not only improves the model’s accuracy but also its interpretability making it easier for clinicians to understand and trust the model’s predictions The techniques implemented in the proposed model along with a comparison against other pretrained models and their respective counterparts This comparison underscores the superiority of the proposed model in terms of accuracy validating the effectiveness of the proposed work To capture sequential dependencies in the data Recurrent Neural Networks (RNNs) like Long Short-Term Memory (LSTM) units These RNNs allow the proposed model to maintain and transfer information from previous steps of convolution layers enhancing its ability to understand sequential patterns The proposed approach integrates RNNs with a custom CNN and attention mechanisms The proposed model’s performance is evaluated against several other models including pre-trained networks like ResNet-50 as well as configurations using only custom CNNs and custom CNNs combined with RNNs which merges a custom CNN with RNNs and attention mechanisms proved to be the most effective among the tested configurations The custom Convolutional Neural Network (CNN) is designed to efficiently extract spatial features from input images The model starts with an input layer that accepts images of size (224 corresponding to 224 × 224 pixels with 3 color channels (RGB) It includes four convolutional layers: the first layer has 32 filters of size (3 3) with ReLU activation and ‘same’ padding followed by the second layer with 64 filters all using the same kernel size and activation function Each convolutional layer is followed by a MaxPooling2D layer with a pool size of (2 which reduces the spatial dimensions by half effectively downsampling the feature maps while preserving essential features After the final convolutional and pooling layers the output is flattened and passed through a fully connected (dense) layer with 512 units and ReLU activation This dense layer combines the features learned by the convolutional layers into a more abstract representation a Dropout layer with a 0.5 dropout rate is included randomly setting 50% of the input units to zero at each update during training The final layer is a dense layer with units equal to the number of classes This layer provides the probability distribution over the classes for each input image making it suitable for classification tasks RNNs28 are a type of neural network designed to handle sequential data by capturing temporal dependencies But since the dataset comprises of images the following RNN techniques are only used in combination of the custom CNN model These models created with the combination of the custom CNN and RNN are used for demonstrating the superiority of the proposed model The LSTM29 is a specialized variant of RNNs tailored for sequential data processing while circumventing the vanishing gradient problem often encountered in conventional RNNs Its architecture is engineered to address this challenge Every LSTM device has the basic components: a cell The cell is capable of retaining information across extended time steps while the three gates regulate the flow of data into and out of the cell LSTM architecture17 GRU architecture17 Bidirectional LSTMs31 extend the traditional LSTM models by applying two LSTMs to the input data an LSTM processes the input sequence in the forward direction (forward layer) the input sequence is reversed and fed into another LSTM (backward layer) This dual application of LSTMs enhances the model’s ability to learn long-term dependencies The bidirectional nature allows for the integration of both past and future contextual information a BiGRU processes this input in two directions computing forward hidden states and backward hidden states leveraging their learned representations and adapting them to the specific classification task These pre-trained models were used as fixed feature extractors to compare with the proposed model The input image was passed through the pre-trained model’s convolutional layers and the resulting output features were used for classification a dense layer with 1024 units is added to further process the extracted features This approach leverages the pre-trained weights without further fine-tuning allowing the extracted features to capture meaningful information from the images By comparing the performance of these transfer learning models with the proposed model this work is aimed to evaluate the effectiveness and efficiency of the proposed approach in diagnosing skin lesions primarily used for object detection and classification tasks It enhances the capabilities of DL models with its 16 layers each employing ReLU as the activation function enabling effective feature extraction through nonlinear transformations of input data an adaptive learning rate method that accelerates computation and reduces the need for extensive parameter tuning the SoftMax activation function is employed to facilitate quick convergence and provide probability estimates for multi-class classification tasks These modules utilize parallel convolutions with different kernel sizes allowing the network to effectively capture multi-scale features This approach enhances the model’s ability to extract intricate patterns and details from images contributing to its high performance in diverse image classification tasks Having been pre-trained on extensive datasets InceptionV3 showcases robust capabilities in handling complex visual data Its architecture not only promotes efficient feature extraction through its inception modules but also incorporates advanced techniques to optimize training and enhance generalization InceptionV3 has consistently delivered impressive results underscoring its reliability and effectiveness in the field of DL for image analysis and classification Xception36 short for “Extreme Inception,” is a deep learning model architecture introduced by François Chollet in 2017 This model aims to enhance computational efficiency by using a technique called split convolution Split convolution divides the processes of spatial convolution and dimensionality reduction which helps decrease the number of parameters and boosts efficiency The Xception comprehends and extracts intricate feature representations from image data Known for its high computational efficiency and strong performance in image recognition and classification tasks Xception is frequently employed in image processing applications especially on devices with limited resources the spatial relationship between features is leveraged to generate a spatial attention map which emphasizes the spatial features and complements the channel attention average pooling and max pooling operations are performed along the channel axis resulting in two 2D maps: \(\:{F}_{{S}_{avg}}\in\:{R}^{1\times\:H\times\:W}\)and \(\:{F}_{{S}_{max}}\:\in\:{R}^{1\times\:H\times\:W}\) These maps represent the average-pooling and max-pooling features across channels The two maps are then concatenated along the channel dimension to create a combined feature descriptor This concatenated feature descriptor is processed through a convolution layer with a kernel size of 7 padding set to ‘same,’ and a sigmoid activation function wherever spatial attention is used in combination with a model kernel size and padding have been kept the same The spatial attention map \(\:{M}_{S}\) encodes the locations to emphasize or suppress within the input feature map This map is element-wise multiplied with the input tensor to highlight or diminish specific spatial locations based on the generated attention map The process involves reducing the input tensor’s mean and max values across channels and applying a convolution operation to obtain the attention map this map is used to reweight the input tensor enhancing the network’s focus on important spatial features This attention mechanism effectively captures and emphasizes relevant spatial information improving the model’s performance in tasks involving spatial relationships where \(\:GAP\) and \(\:GMP\) denote global average pooling and global max pooling These pooled features are then processed through a shared multi-layer perceptron (MLP) as depicted in (13) Here, \(\:{W}_{1}\:\in\:\:{R}^{(C/r\times\:C)}\) and \(\:{W}_{2}\:\in\:\:{R}^{(C\times\:C/r)}\) are the weights of the two dense layers, r is the reduction ratio, δ represents the ReLU activation function, and σ denotes the sigmoid function. The channel attention weights are computed by combining the MLP (Multi Layer Perceptron) outputs are depicted using (14) Finally, the attended feature map X’ is obtained by: (depicted by Eq. 15) where ⊙ represents element-wise multiplication it not only enhances performance but also provides valuable insights into the model’s feature prioritization process employing a trainable context vector to compute attention weights across the temporal dimension Given an input sequence \(\:X\:\in\:{R}^{T\times\:D},\) where T represents the number of time steps and D the feature dimensionality the mechanism first projects the input into a hidden representation: \(\:W\:\in\:{R}^{D\times\:U}\) is a weight matrix and U is the number of units in the attention layer The attention weights are then computed using a context vector \(\:u\:\in\:\:{R}^{U}\): where \(\:{a}_{it}\) represents the importance of each time step The attended representation is obtained by weighting the input: This mechanism allows the network to assign varying importance to different time steps potentially capturing long-range dependencies more effectively than standard recurrent architectures By returning both the attended output X’ and the attention weights \(\:{a}_{it}\) this approach not only enhances the model’s ability to focus on relevant temporal information but also provides interpretability allowing researchers to analyze which time steps the model deems most crucial for its predictions contributing to stable training and effective temporal feature extraction The data’s pixel values was rescaled by dividing them by 255 to standardize the pixel intensities between 0 and 1. Such normalization steps ensure that the model retains consistency and compatibility across all stages and also enhances computational efficiency. The proposed method does not include a denoising step, as different denoising algorithms can lead to significant information loss if not applied carefully. Moreover, the model employs attention mechanisms to highlight important areas, which helps in reducing noise without significant information loss. Data augmentation techniques enhance the diversity and variance in the training dataset, allowing the model to learn from a broader range of image transformations and distortions. By training the model with augmented images, it becomes more resilient and better prepared to manage various lighting conditions, orientations, and distortions. The images were resized to 224 × 224 pixels, processed in batches of 32, and labeled using categorical class mode with shuffling enabled to ensure random sampling. Model architecture of the proposed model The proposed model architecture is illustrated in Fig. 3 The proposed model integrates CNNs with BiLSTMs (refer Section III.B.1) and incorporates custom spatial and temporal attention mechanisms to effectively capture spatial enhancing the robustness and reliability of disease diagnosis The model begins with a series of convolutional and max-pooling layers which are essential for extracting spatial features from the input images The input layer accepts images of size 224 × 224 × 3 which corresponds to the dimensions and color channels of the images The initial layers consist of Convolutional 2D (Conv2D) Layer 1 with 32 filters and a kernel size of 3 × 3 using the Rectified Linear Unit (ReLU) activation function This is followed by a MaxPooling 2D layer that reduces the spatial dimensions aiding in the extraction of low-level features like edges and textures This pattern continues with Conv2D Layer 2 Conv2D Layer 3 and Layer 4 increase the filter size to 128 and 256 respectively progressively capturing more complex features as the layers deepen The architecture of the CNN layers has been already discussed in the custom CNN model (refer Section III.A) Custom layers for spatial, channel, and temporal attention are defined and utilized in the model. The spatial attention mechanism enhances spatial features by focusing on important regions within the images. A sample of images and their Spatial Attention maps is given Fig. 4 The channel attention mechanism assigns importance to different feature maps or channels identifying the most critical aspects of the data This is achieved by performing global average and max pooling operations passing the results through shared dense layers and combining the outputs to form a channel-wise attention map The temporal attention mechanism is particularly useful for sequence data enabling the model to focus on significant time steps within the sequences It uses a dense layer to compute attention scores which are then used to weigh the input features the model reshapes the extracted feature maps into a sequence format to be fed into the BiLSTM layers This transformation is achieved using a Reshape layer which converts the 2D spatial features into a 1D sequence The BiLSTM layers then process this sequence data to capture temporal dependencies and patterns The use of BiLSTMs is particularly beneficial for skin lesion classification as it allows the model to maintain and leverage contextual information across the spatial dimensions of the image are capable of learning long-term dependencies and patterns within sequences This property enables the model to better understand and classify complex patterns in skin lesions which may involve subtle variations in texture and structure over different regions of the image The final layers of the model consist of a Dense layer with 512 units and ReLU activation The output layer uses a softmax activation function to produce probability distributions for each class with the number of units equal to the number of classes in the dataset Where\(\:\:\sigma\:\:\)is the softmax function \(\:{e}^{{z}_{i}}\) is the standard exponential function of the input tensor The upper limit of the summation is 7 because of the 7 number of classes \(\:{e}^{{z}_{j}}\) is the standard exponential function of the output tensor The model is then trained on the augmented training data, with the validation data used to monitor its performance and adjust the learning process through the callbacks. Original images and their spatial attention maps Comparison of performance metrics for various models. Radar chart depicting the performance of the proposed model The Table 6; Fig. 5 provide a detailed evaluation of various deep learning models across multiple performance metrics offering valuable insights into their strengths and weaknesses demonstrating superior performance in almost all metrics compared to its counterparts it proves its ability to handle both false positives and false negatives effectively Its unmatched specificity of 98.79% and AUC of 99.42% further reinforce its capability to distinguish between positive and negative classes with remarkable precision Secondary metrics such as the JAC score (87.08%) and MCC score (91.55%) also highlight its well-rounded and robust performance making it the most reliable model in this comparative study fall short of the proposed model’s capabilities The CNN + LSTM with Attention Mechanisms achieves high scores but consistently lags behind the proposed model in crucial metrics like JAC (86.04%) and MCC (90.80%) while competitive with an accuracy of 90.66% and precision of 90.80% is unable to match the robustness of the proposed model Its JAC score of 83.71% and MCC score of 89.12% reveal that it struggles to generalize as effectively underlining the benefits of BiLSTM over GRU in capturing complex bidirectional dependencies although achieving respectable metrics such as 90.21% accuracy and 90.18% precision performs significantly lower on JAC (82.93%) and MCC (88.60%) indicating its relative inferiority when compared to models with BiLSTM layers Non-attention-based models show a marked decline in performance highlighting the critical role of attention mechanisms in enhancing model effectiveness The CNN + BiLSTM achieves 88.96% accuracy and 88.82% precision benefiting from the BiLSTM’s ability to capture sequential dependencies the absence of an attention mechanism restricts its ability to focus on the most important features resulting in lower scores such as a JAC score of 81.13% and MCC score of 87.13% The CNN + LSTM and CNN + BiGRU models exhibit similar limitations the CNN + GRU performs significantly worse with an accuracy of 86.98% and a steep drop in JAC (78.37%) and MCC (84.99%) making it less effective in capturing intricate relationships in the data With an accuracy of 80.06% and an F1 score of 80.85% it is clear that the lack of recurrent layers limits its ability to learn temporal and sequential patterns also show subpar performance in comparison InceptionV3 achieves an accuracy of 81.63% and a JAC score of 70.51% while VGG16 lags further with only 73.79% accuracy and the lowest JAC score of 59.72% These results underscore the limitations of these architectures for the given task as they fail to match the performance of hybrid models combining CNNs with RNNs and attention mechanisms the CNN + BiLSTM with Attention Mechanisms consistently outperforms its competitors across all metrics Its ability to integrate the strengths of convolutional layers and attention mechanisms enables it to capture complex patterns and focus on critical features making it the most effective and robust model in this evaluation This comparative analysis clearly illustrates the superiority of the proposed model and its ability to deliver state-of-the-art performance for this task Overall, the radar charts vividly showcase the dominance of the CNN + BiLSTM with Attention Mechanism, reflecting its unparalleled ability to capture intricate patterns and focus on relevant features while minimizing errors. This model’s comprehensive and consistent outperformance across all metrics solidifies its position as the most efficient and reliable among the architectures evaluated. (a) Accuracy plot of the proposed model The provided plots in Fig. 7(a) (b) illustrate the training and validation performance metrics for the CNN with BiLSTM and Attention Mechanisms model over 20 epochs showcasing how the model’s accuracy and loss evolve during the training process The Fig. 7(a) shows the accuracy progression for both training and validation datasets the model starts with an accuracy of approximately 40% both training (blue line) and validation (orange line) accuracies show a steady upward trend the accuracies for both datasets exceed 90% The close alignment of the training and validation accuracy curves indicates that the model is effectively learning from the training data while maintaining consistent performance on the validation data This suggests that the model is not overfitting as there is no significant divergence between the training and validation accuracy a common issue where the model performs well on training data but poorly on unseen data The Fig. 7(b) illustrates the loss values for both training and validation datasets Loss measures how well or poorly the model’s predictions match the actual results with lowervalues indicating better performance the loss for both training and validation is around 1.4 there is a sharp decline in loss for both datasets indicating rapid learning during the initial epochs the training loss (blue line) and validation loss (orange line) closely track each other which further indicates good generalization by the model The parallel descent of both curves towards lower loss values without significant gaps demonstrates that the model is efficiently capturing the patterns in the data without overfitting Confusion matrix achieved for the proposed model. Confusion matrices for all other models with attention mechanisms (a) CNN with BiGRU and attention mechanisms (b) CNN with LSTM and attention mechanisms In Fig. 8, the proposed model stands out as the top performer among the other three models depicted in Fig. 9 Its confusion matrix displays strong diagonal elements indicating high classification accuracy across all seven classes This model demonstrates a remarkable ability to distinguish between similar categories with minimal off-diagonal elements suggesting low misclassification rates The BiLSTM architecture appears to capture long-range dependencies in the data effectively incorporating CNN with BiGRU and Attention Mechanisms though with a slight increase in misclassifications compared to the BiLSTM model This subtle difference suggests that while BiGRU is highly effective the BiLSTM might have a marginal edge in capturing complex patterns within this particular dataset Across all four models, two classes consistently achieve near-perfect classification, suggesting these categories have highly distinctive features easily recognized by the neural networks. Conversely, two other classes show persistent confusion across all models, indicating a challenging similarity that even these sophisticated architectures struggle to fully differentiate. The matrices reveal subtle but important differences in how each model handles the nuances of the classification task. For instance, the BiLSTM model appears particularly adept at correctly classifying two specific categories that the other models misclassify more frequently. This could indicate its superior ability to capture subtle, long-term dependencies crucial for distinguishing these particular classes. The graph illustrates the model’s exceptional performance across seven distinct classes: AKIEC each represented by a different coloured curve All curves demonstrate remarkably high classification accuracy which is indicative of an almost ideal classifier The area under the curve (AUC) for each class ranges from 0.98 to 1.00 underscoring the model’s robust discriminative power representing the performance of a random classifier further emphasizing the model’s effectiveness the curves for different classes are tightly clustered suggesting consistent performance across all categories This consistency is particularly impressive given the challenges often associated with multi-class medical image classification tasks The proposed model’s ability to maintain such high accuracy across diverse classes showcases its potential as a powerful tool in automated medical diagnosis potentially aiding in the early detection and classification of various skin conditions This comparative analysis emphasizes the evolution of methods and performance measures across state-of-the-art studies This work has introduced a pioneering model that integrates Convolutional Neural Networks (CNN) with Bidirectional Long Short-Term Memory (BiLSTM) units and Attention Mechanisms demonstrating remarkable performance in medical diagnostics This model achieves an impressive accuracy of 92.73% underscoring its robustness and reliability in classifying various skin conditions The addition of attention mechanisms enhances the model’s capability to focus on the most critical aspects of the data ensuring even subtle variations in medical images are meticulously examined This approach paves the way for reliable and early detection of skin conditions effectively captures the spatial features necessary for accurate diagnosis further improvements to the CNN architecture itself could significantly enhance its performance when combined with recurrent neural networks (RNNs) the synergistic effect with RNNs can be amplified leading to even more accurate and robust models and Matthews Correlation Coefficient (91.55%) underscores its potential to revolutionize clinical practices High accuracy means the model correctly identifies most skin conditions Precision shows that diagnoses are usually correct The F1 score balances precision and recall Recall reflects the model’s ability to identify most actual cases The Jaccard Index and Dice Coefficient measure how well predictions match real conditions with higher values indicating better performance The Matthews Correlation Coefficient provides a balanced measure of the model’s performance considering true positives the exploration of additional hybrid models and improvements in the CNN architecture could further elevate the capabilities of machine learning in medical imaging there are a few challenges in implementing the same in real-world scenario One of the major challenges could be the change in illumination while capturing the images Another challenge could be the impact of heterogeneous background scenario while capturing the images that can dampen the accuracy Future work could be focussed towards enhancing this model towards addressing these challenges using other vision and image processing techniques future work will be directed towards using this automated skin lesion classification model across patients with suitable guidance and support from medical professionals in this field this work not only presents a new model but also marks a significant step towards integrating advanced machine learning techniques into healthcare This work promises a future where technology and medicine converge to save lives and enhance patient care demonstrating the profound potential of machine learning in transforming medical diagnostics The dataset used for this work is available in https://www.kaggle.com/datasets/shawon250/ham10000-balanced-128 × 128 Skin cancer classification with deep learning: a systematic review AI HAM 10000 database to assist residents in learning differential diagnosis of skin lesions In 2022 IEEE 5th Eurasian Conference on Educational Innovation (ECEI) (pp Segmentation and Classification Techniques for Detection of Skin Lesions In 2023 2nd Edition of IEEE Delhi Section Flagship Conference (DELCON) (pp Detection of skin cancer using Cnn algorithm Skin lesions diseases classification using deep convolutional neural network with transfer learning model Decision support system for detection and classification of skin lesions using CNN In Innovations in computational intelligence and computer vision: proceedings of ICICV 2020 (pp A deep-ensemble-learning-based approach for skin lesions diagnosis Artificial intelligence-driven enhanced skin lesions diagnosis: leveraging convolutional neural networks with discrete wavelet transformation Skin cancer classification using convolutional capsule network (CapsNet) Classification of skin lesions in dermatoscopic images with deep Convolution network An enhanced technique of skin lesions classification using deep convolutional neural network with transfer learning models Skin lesion classification using convolutional neural network with novel regularizer Studies on different CNN algorithms for face skin disease classification based on clinical images Skin lesions classification into eight classes for ISIC 2019 using deep convolutional neural network and transfer learning Discriminative feature learning for skin disease classification using deep convolutional neural network Deep semantic segmentation and multi-class skin lesion classification based on convolutional neural network Skin lesion segmentation in dermoscopic images with ensemble deep learning methods A deep learning approach based on explainable artificial intelligence for skin lesion classification Skin lesion classification by multi-view filtered transfer learning FCN-based densenet framework for automated detection and classification of skin lesions in dermoscopy images Malignant melanoma classification using deep learning: datasets Skin lesion classification by ensembles of deep convolutional networks and regularly spaced shifting Multiple skin lesions diagnostics via integrated deep convolutional networks for segmentation and classification Deep learning from limited training data: novel segmentation and ensemble algorithms applied to automatic melanoma diagnosis Fusing fine-tuned deep features for skin lesion classification A review of recurrent neural networks: LSTM cells and network architectures A deep recurrent neural network with bilstm model for sentiment classification In 2018 International conference on Bangla speech and language processing (ICBSLP) Sentiment classification using a single-layered BiLSTM model Email spam detection using multi-head CNN-BiGRU network In International Conference on Advanced Network Technologies and Intelligent Computing Transfer learning based plant diseases detection using ResNet50 In 2019 4th International conference on electrical information and communication technology (EICT) In 2017 2nd international conference on image Rice disease recognition using transfer learning Xception convolutional neural network An empirical study of spatial attention mechanisms in deep networks In Proceedings of the IEEE/CVF international conference on computer vision (pp Image segmentation of skin lesions based on attention mechanism and deep convolutional neural network In 2021 International Conference on Intelligent Transportation Melanoma skin lesions detection using EfficientNet and channel attention module In 2021 IEEE 16th International Conference on Industrial and Information Systems (ICIIS) (pp Skin lesion segmentation based on multi-scale attention convolutional neural network Spatio-temporal attention networks for action recognition and detection Temporal attention network for action proposal In 2018 25th IEEE International Conference on Image Processing (ICIP) Activation functions in deep learning: A comprehensive survey and benchmark Shawon, B. HAM10000 Balanced 128x128 [Dataset]. Kaggle. (2021). https://www.kaggle.com/datasets/shawon250/ham10000-balanced-128x128 The effectiveness of data augmentation for melanoma skin lesions prediction using convolutional neural networks In 2020 IEEE 2nd International Conference on Artificial Intelligence in Engineering and Technology (IICAIET) (pp A survey on image data augmentation for deep learning MobileNet-GRU fusion for optimizing diagnosis of yellow vein mosaic virus Deep learning using rectified linear units (relu) The most used activation functions: Classic versus current In 2020 International Conference on Development and Application Systems (DAS) Comparative analysis of optimizers in deep neural networks Detection and identification of healthy and unhealthy sugarcane leaf using Convolution neural network system The advantages of the Matthews correlation coefficient (mcc) over f1 score and accuracy in binary classification evaluation Multi-label classifier performance evaluation with confusion matrix Generating ROC curves for artificial neural networks International Skin Imaging Collaboration. ISIC 2019: Skin lesion analysis towards melanoma detection. ISIC Archive. https://challenge.isic-archive.com/data/#2019 (2019) International Skin Imaging Collaboration. ISIC 2018: Skin lesion analysis towards melanoma detection. ISIC Archive. https://challenge.isic-archive.com/data/#2018 (2018) https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T Download references The authors would like to thank Centre for Cyber Physical Systems and School of Computer Science and Engineering Chennai for giving the support and encouragement to proceed with the research and produce fruitful results Open access funding provided by Vellore Institute of Technology Ayesha Shaik & Ananthakrishnan Balasundaram School of Computer Science and Engineering Shreyas Kumar & Ananthakrishnan Balasundaram did the implementation and manuscript writing The authors declare no competing interests Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations Download citation DOI: https://doi.org/10.1038/s41598-025-00025-2 Anyone you share the following link with will be able to read this content: a shareable link is not currently available for this article Sign up for the Nature Briefing newsletter — what matters in science Out-of-home advertising is grabbing attention across several industries with 70% of adults recall seeing ads and 83% took action after viewing the ads according to a study from The Out of Home Advertising Association reported purchasing a personal care product after viewing an OOH ad and ads featuring pricing and product benefits have more impact Get the latest news and resources from Digital Signage Today By going public now Benjamin Netanyahu hopes to squeeze Hamas for concessions and please the far right The announcement of Israel’s plan to launch imminently a new, expanded offensive in Gaza and to retain the territory it seized is a significant moment Israeli troops have carried out large and frequently bloody operations that have covered all except central parts of Gaza but they have largely restricted their permanent presence to a buffer zone about 1km deep along the devastated territory’s perimeter and two relatively narrow east-west corridors This now seems to have changed. Once “Operation Gideon’s Chariots” is under way, Israel will send its troops across much – if not all – of Gaza and will seek to establish a “sustained presence” there All this will confirm many people’s long-held fears of Israel’s intentions in Gaza and prompt international outrage The idea of a major new offensive in Gaza has been discussed and debated within the government and the upper ranks of the military for some months So why has Israel’s government announced this plan so loudly A key factor is the indirect talks being held with Hamas about a new ceasefire The government of the Israeli prime minister hopes that the Israel Defense Forces’ call-up of tens of thousands of reservists the threat of the new offensive and the prospect of Israel seizing swaths of territory will force Hamas’s leaders to make concessions then physical possession of terrain will offer useful leverage in future negotiations and allow Hamas to be squeezed further in the meantime Israel’s twin war aims – to crush Hamas and free the 59 hostages it still holds – remain unchanged though Netanyahu has signaled the former is the priority Trump is due to visit the Middle East in 10 days and Israeli officials said the offensive would start after the leader of their country’s most important ally had enjoyed the hospitality of Saudi Arabia Images of destruction and death from Gaza would make the president’s stay that much more diplomatically delicate the complex logistics necessary to move and mobilise additional troops in Israel is likely to mean an even longer delay Israel has also now gone public with its plan to allow some aid into Gaza which has been brought to “the brink of catastrophe” by two months of Israel’s tight blockade of food The scheme involves creating big distribution sites run by private contractors in the south of Gaza to which vetted representatives of each Palestinian family would travel to pick up food parcels likely to be situated in a vast zone up to 5km wide now being cleared along the border with Egypt The scheme has been dismissed as unworkable dangerous and potentially illegal under international law by leading humanitarian organisations There has been no mention either of who might provide healthcare fuel and everything else necessary for life in the territory Policy papers outlining and advocating the imposition of a military administration on Gaza have been circulating among senior officials in Israel for more than a year Netanyahu continues to dismiss out of hand the possibility of the Palestinian Authority which exercises partial authority in the occupied West Bank Nor has he outlined any other kind of future political settlement in Gaza would be that Israeli troops end up the de facto rulers of much of Gaza and its 2.3 million inhabitants The Israeli prime minister also offers little to the majority of Israelis who call for a ceasefire deal to secure the release of the hostages His coalition still depends heavily on the support of far-right parties who are very happy with the prospect of the new offensive and the prospect of a “sustained” Israeli presence in Gaza Netanyahu now appears likely to remain in power for the 15 months or so until the next elections in the Hamas surprise raid into Israel in 2023 that triggered the war have died in the Israeli offensive there which followed Any inquiry should wait until the war has ended Metrics details Remote monitoring of transmission lines plays a vital role in ensuring the stable operation of power systems especially in regions with weak or unstable network signals where efficient data transmission and storage are essential traditional image compression methods face significant limitations in both quality and efficiency when applied to high-resolution imagery in such scenarios.To address these challenges this paper proposes a deep learning–based image compression approach incorporating an Efficient Channel-Temporal Attention Module (ETAM) The ETAM module integrates Efficient Channel Attention (ECA-Net) and a Temporal Attention Module (TAM) to jointly enhance the extraction of spatial and temporal features thereby improving compression efficiency and reconstruction quality.Experimental results demonstrate that the proposed method consistently outperforms both traditional and state-of-the-art deep learning–based compression techniques across multiple evaluation metrics evaluations on the STN PLAD dataset show that ETAM better preserves fine-grained details and textures resulting in reconstructions that closely resemble the original images.These findings underscore the practical potential of the ETAM method for efficient high-quality image compression in real-world applications such as transmission line monitoring under constrained network conditions making real-time monitoring significantly more difficult Traditional approaches to monitoring primarily rely on manual ground inspections or UAV-based image and video acquisition for equipment status assessment with the growing number of transmission components and the continuous expansion of monitoring coverage the efficient transmission and processing of collected data—especially in signal-limited or disconnected environments—has emerged as a major challenge improving compression efficiency without compromising image quality has become a key challenge to ensure the effectiveness and safety of transmission line monitoring This paper proposes a deep learning-based image compression method utilizing the Efficient Channel-Temporal Attention Module (ETAM) The method integrates Efficient Channel-Net and the Temporal Attention Module (TAM) to optimize both spatial and temporal representations of images significantly improving compression efficiency while maintaining high-quality reconstruction The ETAM method is particularly suited for monitoring areas with weak signals as it reduces both transmission bandwidth and storage costs by minimizing data volume while ensuring high-quality image recovery under constrained conditions Image compression technology is a key research area in computer vision with the goal of reducing data size while maintaining image quality the rapid advancement of deep learning has sparked increasing interest in integrating traditional image compression techniques with learning-based approaches to improve both compression efficiency and perceptual quality Previous research has primarily focused on traditional compression techniques and the application of attention mechanisms in image processing Traditional image compression methods—such as JPEG and WebP—achieve data reduction through a combination of transformation all governed by hand-crafted and fixed coding rules these methods can maintain relatively satisfactory image quality; however detail loss and blocking artifacts become more pronounced JPEG and JPEG2000 are prone to significant artifacts when processing edges and high-frequency regions This method divides the image into regions of interest (ROI) and non-ROI areas: lossless BPG encoding is applied to the ROI while lossy BPG encoding is used for the non-ROI areas and the two parts are then merged to reconstruct the final image While such approaches offer moderate compression efficiency their performance degrades significantly when applied to high-resolution images or visually complex scenes often resulting in suboptimal rate-distortion trade-offs and noticeable visual artifacts This algorithm improves the efficiency of block transform coefficients through pre-filtering and higher-order null-frequency context modeling thereby enhancing coding performance and reconstruction quality Although this method improves distortion control at high compression ratios its adaptability and flexibility for complex and diverse image data are still limited This method is notable for requiring a single training process enabling the network to dynamically adjust compression rates in real-world applications without retraining separate models for different bit rates This feature significantly reduces operational and maintenance costs paving the way for the broader application of deep learning in image compression Although autoencoders perform well in static image compression they struggle to effectively handle temporal data this paper incorporates the Efficient Channel-Temporal Attention Module (ETAM) into deep learning-based image compression The channel attention mechanism optimizes feature extraction along the channel dimension while the temporal attention module captures temporal variations and emphasizes critical information by adjusting the weights of key time points These mechanisms enable joint optimization of spatio-temporal features enhancing the model’s feature representation and its applicability to video and dynamic image processing The channel attention mechanism improves the feature representation of key components in surveillance images by capturing dependencies between channels thereby enhancing the accuracy of transmission line state identification TAM) focuses on identifying critical change points in temporal signals which is crucial for detecting line anomalies (e.g. line breaks or overloads) in dynamic environments an attention mechanism that effectively integrates spatial and temporal features can optimize image compression and reconstruction significantly reducing the data volume and ensuring the stable operation of the remote monitoring system is a lightweight attention mechanism designed to enhance feature representation in image processing tasks Unlike traditional channel attention mechanisms ECA-Net improves computational efficiency by using an adaptive one-dimensional convolutional approach to learn inter-channel correlations thereby avoiding the dimensional compression problem common in conventional methods the input feature maps undergo Global Average Pooling (GAP) to capture global statistical information for each channel by adaptively selecting the convolutional kernel size (k) a weighting operation is applied to the channel dimensions allowing the model to learn the importance of each channel There is a direct relationship between the convolutional kernel size (k) and the number of channels (C) in the input feature map which enables ECA-Net to flexibly adjust the importance of each channel thereby emphasizing critical information relevant to the task the weights for each channel are transformed into normalized coefficients through a sigmoid activation function which are then used to adjust the intensity of each channel feature in the original feature map and σ denotes the activation function (sigmoid) k refers to the coverage for local cross-channel interactions and there exists a mapping relationship between k and C: The corner symbol ⊙ denotes the closest odd number to the value inside∣⋅∣The main process of the ECA attention mechanism is as follows: After the input feature map enters the ECA attention mechanism the Global Average Pooling (GAP) operation is first applied to average pool the feature values of each channel producing the global average value for each channel this pooled value is passed through a channel-wise Fully Connected Layer (FC) which outputs the weight coefficients for each channel These coefficients are then used to adjust the feature weights of different channels the ECA attention mechanism applies these weight coefficients to rescale the original feature map thereby weighting the features of different channels and obtaining the weighted feature map The Temporal Attention Module (TAM), proposed by Wu et al.15 aims to capture important features in signals over time and highlight critical signal changes through a weighting mechanism Unlike traditional spatial attention modules TAM focuses on signal variations in the time dimension allowing it to effectively extract time-related information The application of the Temporal Attention Module (TAM) is particularly crucial in remote monitoring scenarios of transmission lines especially in regions with no signal coverage TAM operates on the feature maps generated by the preceding Efficient Channel Attention Module (ECA-Net) applying both max pooling and average pooling to produce two complementary temporal descriptors These maps represent the global importance of line monitoring data in the time dimension and the local critical features These feature maps are then merged and converted into a single-channel feature map via convolution which is subsequently combined with a sigmoid activation function to generate temporal attention weights that highlight critical changes at different time points The key role of this temporal attention mechanism in transmission line monitoring lies in its ability to adaptively adjust the focus of the model’s attention on temporal features based on the dynamics of the input signals TAM generates two feature maps through maximum pooling and average pooling from the output of the channel attention module These two feature maps are then merged into a single-channel feature map via a convolution operation The temporal attention feature map is obtained using a sigmoid function the output is element-wise multiplied with the original feature map The formula for temporal attention is as follows: Where Ms(Y) denotes the convolution result of TAM and \(\:{Y}_{avg}^{s}\) and \(\:{Y}_{max}^{s}\) denote the output features of different pooling layers under the time-attentive module As shown in Fig. 3 the Efficient Channel-Temporal Attention Module (ETAM) is a dual-attention mechanism proposed in this paper for the image compression task Its goal is to address the challenge of optimizing both spatial and temporal dynamic features of an image thereby achieving more efficient feature representation and better reconstruction quality in compression and reconstruction tasks The ETAM module integrates the Efficient Channel Attention Module (ECA-Net) and the Temporal Attention Module (TAM) which not only identifies spatially important features in an image but also captures key temporal dynamics injecting new vitality into deep learning-based image compression methods ECA-Net addresses this limitation by first applying Global Average Pooling (GAP) to capture global contextual information across each channel followed by an adaptive one-dimensional convolution that directly models local cross-channel interactions and generates channel-wise attention weights This approach avoids dimensional compression and redundant parameters enabling the model to flexibly capture key spatial features and highlight the most relevant information for the task the weights of each channel are normalized using a sigmoid activation function achieving precise optimization of the spatial features of the input image ECA-Net’s lightweight design improves computational efficiency while effectively enhancing the spatial feature representation of the image Efficient channel-temporal attention module This module allows information to pass directly across multiple layers helping to overcome the common problem of gradient vanishing in traditional deep networks and accelerating the training process the residual module enables the network to retain feature information learned in the shallow layers without adding additional computational burden by simply adding the input features to the convolved features the input is first passed through a series of convolutional layers followed by nonlinear activation functions to extract enhanced feature representations These new features are then added to the input signals through skip connections to form the final output This “additive” operation allows the model to learn the relative “residuals,” or the differences between the inputs and outputs rather than learning the complete mapping from scratch This design enables the network to better capture the core information of the original data and refine it in subsequent layers Where X and Xout represent the input and output of the residual module, respectively, and Y denotes the total output. The features learned by the shallow network can be passed to the deeper layers through the residual connection module, thus preventing network degradation. The target loss of the proposed framework in this paper consists of four components: bit rate loss perceptual loss and adversarial loss.For distortion loss we use mean square error (MSE) loss.The calculation formula is: where yi represents the ground truth pixel values and N is the number of pixels in the image Where \(\:{{\varnothing}}_{l}\left(x\right)\) and \(\:{{\varnothing}}_{l}\left(\widehat{x}\right)\) represent the feature maps extracted by the l-th layer of the neural network Hl and Wl denote the height and width of the feature maps and wl represents the weighting coefficient for each layer which adjusts the importance of features from different layers The overall objective function is calculated as follows:\(\:{L}_{total}=\alpha\:{L}_{r}+{\lambda\:}_{1}{L}_{d}+{\lambda\:}_{2}{L}_{per}\) The image compression framework proposed in this paper comprises two primary components: a core autoencoder sub-network and an entropy coding sub-network These two modules work in tandem to efficiently perform image compression and reconstruction the core autoencoder sub-network compresses the input image into a latent representation which is subsequently quantized to produce a quantized latent representation the input image x is transformed into a latent representation y by the encoder ga: the potential representation y undergoes a quantisation operation to obtain the quantised representation ŷ: The sub-network incorporates a Residual Block (RB) to expand the receptive field and enhance the learning capability The Residual Block helps mitigate the gradient vanishing problem in deep networks enabling smooth information flow across multiple layers an Efficient Channel-Time Attention Module (ETAM) is introduced in the encoder This module combines the Efficient Channel Attention Module (ECA-Net) and the Temporal Attention Module (TAM) aiming to optimize both spatial and temporal feature extraction ECA-Net improves channel correlations by adaptively selecting convolution kernel sizes to efficiently weight spatial features while TAM leverages maximum and average pooling to capture temporal dynamics thus enhancing the model’s performance in processing temporal signals The latent representation processed by the encoder is passed to the decoder (gs) which shares a similar structure to the encoder utilizing both a residual module and an efficient channel-time attention module The decoder’s task is to decode the quantized latent representation \(\:\hat{y}\) into a reconstructed image \(\:\widehat{x}\): the decoder can not only extract spatial features but also address temporal variations in the image thereby enhancing the quality and accuracy of image reconstruction µ and σ are the mean and standard deviation obtained from hypernetwork learning representing the probability distribution of the potential representation By entropy coding the potential representation information loss can be minimized while ensuring compression efficiency The training objective of the entire network is to optimize compression performance by minimizing the trade-off between distortion and bit rate: Where \(\:{\mathcal{L}}_{distortion}\) measures the distortion between the compressed image and the original image \(\:{\mathcal{L}}_{rate}\) measures the number of bits used in entropy coding and λ is a hyperparameter that controls the balance between distortion and the number of compressed bits The proposed network enhances spatial and temporal feature extraction during compression by integrating both the Residual Module and the Efficient Channel-Temporal Attention Module (ETAM) thereby facilitating efficient feature learning and accurate information reconstruction the entropy coding sub-network improves compression performance by refining the probability distribution of the quantized latent representations through context modeling This architecture achieves a well-balanced trade-off between compression efficiency and reconstruction quality. The overall network structure is depicted in Fig. 5. PSNR (Peak Signal-to-Noise Ratio) is a crucial metric for image quality assessment22 used to evaluate the similarity between the reconstructed and the original images A higher PSNR value indicates that the quality of the reconstructed image is closer to that of the original The formula for PSNR is defined as follows: Where R is the maximum possible pixel value of the image (255 for an 8-bit image) and MSE is the Mean Square Error Where H and W are the height and width of the image respectively \(\:{x}_{i,j}\) and \(\:{\widehat{x}}_{i,j}\) denote the pixel values of the original and reconstructed images at position (i PSNR is mainly used to quantify the quality of signal reproduction in reconstructed images and can reflect the performance of images in tasks such as noise processing compression and reconstruction.Higher PSNR values usually indicate lower distortion and better visual quality of the compressed image SSIM (Structural Similarity Index) is used to evaluate the similarity of an image in terms of brightness the more structurally similar the image is to the original The MS-SSIM metric aligns more closely with subjective quality assessments where higher values indicate better image quality To better visualize the variability of MS-SSIM across different results MS-SSIM-dB represents the MS-SSIM metric in decibels and is calculated as follows: LPIPS (Learned Perceptual Image Patch Similarity) is an important metric for assessing the perceptual similarity between images as it more closely aligns with human visual perception Unlike traditional pixel-level metrics (e.g. LPIPS evaluates perceptual differences by extracting multi-layer features from an image using a neural network and calculating the weighted distances in the feature space ∅1(x) and ∅1(x̂) represent the feature maps extracted by the l-th layer of the neural network H1 and W1 denote the height and width of the feature maps while w1 refers to the weighting coefficients of each layer used to adjust the importance of features from different layers The advantage of LPIPS lies in its ability to capture perceptual differences by focusing on semantic content and visual quality rather than low-level pixel-wise discrepancies LPIPS measures the distance between original and reconstructed images in a deep neural network feature space allowing it to more accurately assess detail preservation and semantic consistency in scenarios where pixel-level differences are high but visual appearance remains similar LPIPS better reflects the subtle variations that are perceptually significant to the human eye The training set used in this paper comprises approximately 30,000 high-resolution images collected from the Internet These images were resized to a random size between 500 and 1000 pixels and then randomly cropped to 256 × 256 the model’s performance is evaluated and compared using the STN PLAD dataset All experiments were conducted on workstations equipped with NVIDIA Tesla V100 GPUs to ensure efficiency and stability The specific hardware environment is as follows: Processor: Intel Xeon Gold 6248R; GPU: NVIDIA Tesla V100 32GB; Memory: 256GB DDR4; OS: Ubuntu 18.04; Deep Learning Framework: PyTorch 1.12; CUDA version: 11.2 The training and testing durations were 24 h in total with the training phase lasting approximately 16 h and the testing phase lasting about 8 h the Adam optimiser is used to train the network The initial learning rate is set to \(\:1\times\:{e}^{-4}\) λ1 is set to \(\:2\times\:{e}^{-3}\) and λ2 is set to 1 the approach used in this paper is to keep each λ relatively fixed and adjust α 2.8] for constraining the low code rate at around 0.1 1.6] for constraining the medium code rate at around 0.25 0.8] for constraining the high code rate at around 0.4 the trained ETAM model is evaluated on the designated test set compression is performed and corresponding evaluation metrics—including PSNR To assess the effectiveness of the proposed method its performance is compared against both traditional and state-of-the-art deep learning–based compression methods across all metrics The overall experimental design aims to comprehensively assess the superior performance of ETAM in image compression tasks through a variety of experimental configurations and a wide range of evaluation metrics The STN PLAD dataset is designed to support intelligent inspection and maintenance of power lines, with important applications in identifying line assets and detecting potential defects. It is particularly suitable for fault detection on transmission lines in signal-free areas. The model was trained by adjusting parameters to produce several models at different code rates, and the Rate-Distortion Curve and Rate-Perception Curve were plotted. Bit rate-distortion curves for different compression models As shown in Fig. 6 the PSNR and MS-SSIM metrics of the proposed method on the STN PLAD dataset are significantly higher than those of traditional compression methods The PSNR curve demonstrates that the proposed method provides higher quality reconstructed images at the same bit rate further confirming the advantages of ETAM in efficiently extracting image features and minimizing information loss The MS-SSIM curve indicates a significant improvement in structural similarity with the proposed method Bit rate-perception curves for different compression models The experimental results show that the ETAM method proposed in this paper significantly outperforms mainstream compression methods in terms of both bit rate-distortion performance and perceptual performance This confirms the effectiveness of the ETAM module in optimizing spatial and temporal feature extraction through the dual attention mechanism and highlights its potential for application in image compression tasks As shown in Table 1 to validate the generalization of the proposed method and LPIPS metrics are tested around 0.1 Bpp using the Kodak dataset it can be observed that the experimental results follow the same pattern as those obtained on the STN PLAD dataset lack the ability to adaptively model image content and are prone to significant block effects Modern image compression standards such as BPG and WebP still rely on rule-driven coding frameworks and are unable to optimize feature extraction strategies through learning mechanisms thus presenting bottlenecks in preserving the structural integrity and perceptual quality of images Some deep learning-based compression methods introduce nonlinear modeling capabilities but their feature extraction modules usually focus on modeling spatial information only ignoring the contextual evolution and dynamic relationships of features which can easily lead to insufficient detail recovery especially when dealing with complex image scenes some of the deep attention mechanisms improve the expression ability but the high computational complexity and large number of parameters are not favorable to be deployed in practical applications ETAM module realizes the joint optimization of spatial and dynamic features by introducing the cascade mechanism of channel attention and temporal attention.ECA-Net can accurately mine the salient features of the image in the channel dimension at a very low parameter cost so as to enhance the model’s ability to pay attention to the spatial structure; and the TAM module simulates the change rule of the features in the temporal or contextual dimension so that the model has the ability to model dynamically The TAM module simulates the change pattern of features in time or context dimension so that the model has the ability of dynamic modeling which can effectively make up for the shortcomings of the traditional model in semantic expression ETAM can greatly improve the expression efficiency and information density of compressed features while keeping the network structure lightweight thus improving the overall compression rate and reconstruction quality the introduction of perceptual loss further optimizes the model’s ability to restore texture and details which makes the reconstructed image closer to the subjective experience of the human eye in terms of visual effect Comprehensive experimental results show that the ETAM method outperforms the comparison methods on several typical datasets and exhibits good generalization ability which proves that it not only has advantages at the index level but also effectively overcomes the key limitations of the existing image compression techniques in the method design The experimental results demonstrate that ETAM significantly enhances both image reconstruction quality and compression efficiency and perceptual quality (LPIPS) decreases from 0.52 to 0.42 Adding the residual module further boosts PSNR to 29.2 and perceptual quality (LPIPS) improves with a reduction to 0.37 with the inclusion of the entropy coding sub-network and perceptual quality achieves its highest level These results indicate that the dual-attention mechanism in ETAM effectively enhances image compression performance by optimizing the extraction of both spatial and temporal features the addition of the residual module and entropy coding sub-network further refines both compression efficiency and image quality ultimately achieving an optimal performance balance Comparison of image reconstruction effect under different compression methods The results demonstrate that ETAM outperforms all other methods across all images which is significantly higher than JPEG (27.3 dB) significantly higher than the other methods These results indicate that the ETAM method significantly improves image quality in terms of both PSNR and MS-SSIM while further reducing perceptual differences as measured by LPIPS aligning more closely with human subjective perception ETAM not only effectively minimizes information loss at high compression ratios but also greatly enhances the visual quality of image reconstruction confirming its superior performance in image compression tasks and offering an efficient and reliable solution for real-world applications such as remote monitoring of transmission lines Visual example of the limitations of the ETAM approach It is worth stating that the main application scenario of the method in this paper is remote monitoring and inspection image compression of transmission lines In typical transmission datasets such as STN PLAD most of the images have clear structural boundaries and have a relatively homogeneous background while there are individual images in some of the Kodak24 datasets where ETAM performance gains are not evident these failures do not represent a performance deficiency in the actual target domain ETAM has consistently demonstrated good detail reproduction and perceptual consistency on transmission monitoring images which is especially valuable in weak signal high compression ratio transmission conditions The ETAM image compression method proposed in this paper adopts an end-to-end fully convolutional design in the model structure which does not rely on fixed-size inputs and thus has good input size adaptation capability Although a 256 × 256 image patch is mainly used in the training phase of the model to improve the training efficiency and enhance the feature diversity the model is able to seamlessly process higher resolution image data without any structural adjustments or additional modules in the testing phase This design makes the model not only suitable for image compression experiments at conventional resolutions but also has the potential to be extended to high-resolution images in real-world application scenarios we apply the trained model to image compression tasks at 1024 × 1024 and 1920 × 1080 resolutions without modifying the model parameters or structure and the test samples include industrial scene images The experimental results show that the ETAM model is suitable for image compression experiments at conventional resolutions The experimental results show that the ETAM model still maintains good compression and reconstruction performance under high-resolution inputs: the image structure sharpness (PSNR) and multi-scale structural similarity (MS-SSIM) are all consistent with those of the low-resolution test and the reconstructed image has clear edges since both ECA-Net and TAM adopt a lightweight local convolutional attention mechanism and do not introduce complex global self-attention or redundant fully-connected structures the growth in computational resource consumption remains linear when the image size is enlarged and no obvious inference bottleneck occurs the average inference time for the model to process a 1080p image is no more than 300ms showing good inference efficiency and practical deployability From the practical application point of view high-resolution image compression has a wide range of needs in the fields of telemedicine Traditional methods are often ineffective in high-resolution scenes due to high information density and limited compression ratio and the recovered images are prone to artifacts or blurring The method in this paper has strong expressive ability in structure and maintains high efficiency in design so it is more suitable to be deployed in this kind of edge equipment with high requirements on image quality and limited processing resources and provides a feasible solution for practical scenarios Although this study mainly focuses on the compression of still images the proposed ETAM module has been conceptually designed to extend to video compression and the Temporal Attention Module (TAM) in ETAM is essentially used to capture key information of the feature map in the “temporal dimension” or “feature dynamics” which is the “time dimension” or “feature dynamics” of the feature map The temporal attention module (TAM) in ETAM is essentially used to capture the key information of the feature map in the “time dimension” or “feature dynamics” In the image compression task of this study this temporal attention is more about modeling the context or internal trend of the features modeling the “dynamic” information by fusing the maximum pooling with the average pooling so as to enhance the model’s ability to perceive the feature changes This “pseudo-temporal modeling” mechanism does not involve explicit time series modeling between real video frames but it provides structural compatibility for subsequent extensions If ETAM is applied to a video compression task the TAM module can further utilize its ability to handle dynamic changes between frames TAM can capture temporal consistency and trends by aggregating the intermediate representations of the current frame with the preceding and following frames thus generating a more discriminative temporal attention map that enhances the critical frame regions which can help to reduce redundant coding and improve compression efficiency we propose a deep learning-based image compression method using the Efficient Channel-Temporal Attention Module (ETAM) designed to address the challenges of data transmission and storage efficiency in transmission line monitoring scenarios especially in regions with unstable or no signals The ETAM module jointly optimizes image spatial features and temporal dynamic features ECA-Net efficiently captures inter-channel correlations through adaptive one-dimensional convolution to improve spatial information representation while TAM captures key temporal dynamics through a combination of maximum and average pooling thereby enhancing the modeling capability of temporal features Coupled with the residual module and entropy coding sub-network this method further boosts the model’s ability to represent data and enhances compression efficiency significantly reducing information loss under high compression ratios The experimental results comprehensively validate the superiority of the ETAM method Tests on the STN PLAD and Kodak datasets show that our method outperforms traditional compression methods (e.g. JPEG2000) and other modern techniques (e.g. ETAM effectively preserves edge details and texture features under high compression ratios with the reconstructed images closely resembling the original ones the ablation experiments provide valuable insights into the contributions of the ETAM module and the entropy coding sub-network to the overall performance By jointly optimizing context modeling and super-network generation the entropy coding sub-network accurately models the distributional characteristics of latent variables further improving compression efficiency and reconstruction quality the significant improvement in the perceptual similarity (LPIPS) metric demonstrates that the reconstructed images generated by our method align closely with human visual perception The data involved in this study contains sensitive information and sharing it publicly mayresult in potential intellectual property or privacy issues The data has been used exclusively for this study and is not authorized for other uses This study does not involve resources that need to be submitted to public databases; therefore Expansion of high-voltage overhead transmission lines to remote areas[J] Impact of varying transmission bandwidth on image quality[J] Prantl, M. Image compression overview[EB/OL]. arXiv:1410.2259 Medical image compression based on region of interest using better portable graphics (BPG)[C] 2017 IEEE International Conference on Systems Context-based entropy coding of block transform coefficients for image compression[J] Image compression using deep learning: methods and techniques[J] Deep convolutional autoencoder-based lossy image compression[C] 2018 Picture Coding Symposium (PCS) IEEE Optimal deep learning based image compression technique for data transmission on industrial internet of things applications[J] Learned image compression with discretized gaussian mixture likelihoods and attention modules[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition High-fidelity generative image compression[C]//Advances Full resolution image compression with recurrent neural networks[C] In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Networking Scheme Of Non Signal Area In Transmission Line[C] In 2022 IEEE 6th Advanced Information Technology Electronic and Automation Control Conference (IAEAC) ECA-Net: Efficient channel attention for deep convolutional neural networks[C] In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Fcanet: Frequency channel attention networks[C] In Proceedings of the IEEE/CVF International Conference on Computer Vision Classification of motor imagery based on multi-scale feature extraction and the channel-temporal attention module[J] ECA-RetinaNet: A Novel Self-Attention RetinaNet for Environmental Microorganism Image Object Detection[C] In 2023 IEEE International Conference on Big Data (BigData) Learned video compression with efficient Temporal context learning[J] The importance of skip connections in biomedical image segmentation[C] In International Workshop on Deep Learning in Medical Image Analysis International Workshop on Large-Scale Annotation of Biomedical Data and Expert Label Synthesis Neural network pruning with residual-connections and limited-data[C] In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Kettunen, M., Härkönen, E. & Lehtinen, J. E-lpips: robust perceptual image similarity via random transformation ensembles[EB/OL]. arXiv:1906.03973(2019) Zhang, H., Li, L. & Liu, D. Generalized Gaussian model for learned image Compression[EB/OL]. arXiv:2411.19320 (2024) SSIM[C] In 2010 20th International Conference on Pattern Recognition JPEG 2000 encoder[C] 2009 Data Compression Conference Van Laarhoven, T. L2 regularization versus batch and weight normalization[EB/OL]. arXiv:1706.05350 (2017) Dropout: a simple way to prevent neural networks from overfitting[J] Objective assessment of the WebP image coding algorithm[J] Download references Thanks to State Grid Tonghua Power Supply Company for funding this project Present address: Changchun Institute of Technology Future Industry Innovation Research Institute Present address: School of Electrical and Electronic Engineering Present address: State Grid Tonghua Power Supply Company School of Electrical and Electronic Engineering conceived the overall research framework and designed the Efficient Channel-Temporal Attention Module (ETAM) along with the network structure They defined the primary research objectives and innovations and performed comparative evaluations with mainstream image compression methods Download citation DOI: https://doi.org/10.1038/s41598-025-00566-6 Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research Yahoo Finance senior columnist Rick Newman weighs in on Trump's comments and the constraint consumers will notice in available goods Chinese officials have stated they are open to enter into trade negotiations with the Trump administration US indices opened lower but were mixed by mid-morning Traders are looking to a busy week including key meetings for the Bank of England and Federal Reserve as well as new comments on trade from the US US president Donald Trump hinted late on Sunday that trade deals which could nullify extreme tariff decisions Trump said that agreements "could very well be" on their way I’ll set my own deals — because I set the deal “You keep asking the same question: ‘When will you agree?’ It’s up to me The Federal Reserve is set to begin its two-day rate setting meeting on Tuesday while the Bank of England monetary policy committee meets on Thursday Read more: Trending tickers: Berkshire Hathaway, Palantir, Shell, Netflix Germany's DAX (^GDAXI) ticked up 1.1% by the closing bell. Defence firm Rheinmetall (RHM.DE) was among the top gainers in the index The CAC 40 (^FCHI) fell 0.4% in Paris The pan-European STOXX 600 (^STOXX) was 0.2% higher having traded along a flatline for most of the day The US dollar (DX=F) also lost ground as Wall Street gauges whether recent confidence that Trump has moved to a dealmaking phase is misplaced Stocks: Create your watchlist and portfolio Head over to our US site for more market moving news Read more from our US team Following a trial, Uber (UBER) is now allowing customers based in the UK to pay for their rides in cash The rollout in London is currently under review Individual drivers can still opt out of receiving cash The moves come following a report by the Treasury Committee which suggested vendors might have to have cash as a payment option in the future to protect those that rely on it and was up 4% by the afternoon in European hours amid speculation that exporters are making moves to trade in the greenback for the island's own currency The Taiwan dollar is currently at its strongest position against the dollar in about three years with intraday gains the highest since the 1980s The central bank is set to hold a press conference later on to address the volatility in currency markets Streaming giant Netflix (NFLX) saw its stock price dip more than 4.5% in premarket trade on Monday following a plan laid out by US president Donald Trump to put a 100% tariff on movies "produced in foreign lands" The move is an attempt to draw filmmakers back to the US amid increasingly tempting offers by overseas nations AGAIN!” Trump wrote in a Truth Social post which also said he had authorised the commerce department and the US trade representative to immediately begin instituting such a tariff It's unclear whether the tariff will apply to US filmmakers making movies abroad or streaming giants with operations outside the US It's also unclear how the tariff will be calculated Bloomberg had a report on Sunday that oil giant Shell (SHEL.L) is weighing up the merits of buying rival oil major BP (BP.L) in a move that would create a London-listed energy behemoth The considerations are only in their early stages Any successful bid would capitalise on BP's ailing share price which has dropped more than 30% since this time last year Abel was named Buffett's successor back in 2021 As for where this leaves Buffett and Berkshire Buffett said he has no plans to sell Berkshire stock as a result of the change and won't be far away and could conceivably be useful in a few cases but the final word would be what Greg said On Friday, Berkshire Hathaway stock closed at a record high. Shares have gained over 17% this year against a 3% drop for the S&P 500 (^GSPC) How investors react to this weekend's news will not only shape the market discussion in the week ahead but given Berkshire is the seventh-largest company in the S&P 500 and sports a market cap north of $1.1tn how the stock trades could influence the broader market Read more on Yahoo Finance Chinese officials have signalled reopening trade talks with Washington President Trump on Sunday said he wants a "fair deal" with China but declined to specify what that would look like Wall Street’s attention turns to the Federal Reserve The central bank kicks off a two-day policy meeting on Tuesday with the Fed expected to hold steady with current rates even with Trump amping up pressure on the Fed and chair Jerome Powell in recent weeks ready to bring you the business and markets news of the day In the UK it's the May Day bank holiday, so the FTSE (^FTSE) and London indices are closed As are US futures — with markets across the pond opening at 2.30pm UK time Metrics details Complex pest and disease features appearing during the growth of wheat crops are difficult to capture and can seriously affect the normal growth of wheat crops The existing methods ignore the full pre-interaction of deep and shallow features which largely affects the accuracy of identification we rethink the feature representation and attention mechanism in intelligent recognition of wheat leaf diseases and pests and propose a representation and recognition network (RReNet) based on the feature attention mechanism RReNet captures key information more efficiently by focusing on complex pest and disease characteristics and fusing multi-semantic feature information RReNet further enhances the perception of complex disease and pest features by using four layers of detection units and fast IoU loss function which significantly improves the accuracy and robustness of wheat leaf disease and pest recognition Tests on a challenging wheat leaf pest and disease dataset with twelve pest and disease types show that RReNet achieves precision ablation experiments proved the effectiveness of all parts of the proposed method Crop pests and diseases limit crop yields and the sustainability of quality and efficient agriculture crop pests and diseases detection is mainly carried out by manual methods the method of transition relies on expert knowledge The rapid development and improvement of intelligent identification technology provides more advanced technical means for crop pest detection and crop pest detection has become a technical hotspot in the field of intelligent agriculture today the overlapping interference problem existing in wheat leaf disease and pest recognition scenarios is mainly solved by introducing attention module and pre-training knowledge migration to achieve the purpose of improving the recognition accuracy of wheat leaf disease and pest the existing methods ignore the full pre-interaction of deep and shallow features which greatly affects the accuracy of recognition we rethink the feature representation and attention mechanism in intelligent detection of wheat leaf pests and propose a representation and recognition network (RReNet) based on the feature attention mechanism the contributions of this paper are as follows: which can effectively improve the recognition ability of complex pests and diseases in wheat leaves and can be effectively monitored during wheat generation improving the problem of low detection accuracy of existing wheat pest and disease recognition algorithms We propose a complex pest and disease feature representation module This module effectively combines the convolutional downsampling module the convolutional feature extraction group and the channel attention which effectively improves the representation ability of complex pest and disease features We propose the pest and disease feature attention mechanism We fully optimize the deep and shallow feature representations in the complex pest and disease feature representation module while pre-interacting the deep and shallow features We add a Transformer4 detection unit to the recognition network and introduce a fast IoU This helped the recognition network to effectively improve the feature representation of complex contextual semantic information of pests and diseases A target detector in a natural scene usually is comprised of three components a backbone network composed of deep convolutional neural networks a neck structure incorporating a multi-scale feature fusion mechanism and a head designed for targeting the subject In order to be able to carry out sufficient feature interaction between deep semantic information and shallow semantics to improve the network’s representation of multiscale features BiFPN adds a bottom-up fusion path to the top-down feature fusion of FPN which introduces a flexible spatial fusion mechanism and an adaptive weight learning mechanism to fuse multi-scale feature information more effectively The primary focus of the multi-scale feature fusion mechanism discussed above is the integration of features at various levels to enhance the network’s feature representation performance these studies overlook the issue of information loss that may occur when features from different levels directly interact with each other and this class of methods has been widely used to simplify the model structure two-stage target detectors (Faster R-CNN series) and unanchored frame target detectors (CenterNet series) are the mainstream supervised learning-based target detectors and each of them has different application scenarios: the Yolo series is suitable for real-time response application scenarios and the Faster R-CNN series is suitable for scenarios with high accuracy and without real-time fast detection and the CenterNet series is suitable for scenarios that do not require high detection accuracy but require high detection speed All of the above methods are applications of supervised learning for target detection in natural scenes there are researchers who utilize unsupervised learning to achieve target detection in natural scenes The Yolo series may struggle with detecting tiny objects These small pests and diseases are initially difficult to detect and are often found in shallow features within deep networks This can result in lower accuracy when trying to recognize diseases and pests in wheat crops the Faster R-CNN family can improve detection accuracy and adaptability to complex scenes the complex network structure and slower inference speed of Faster R-CNN can hinder its development in the field of wheat crop pest and disease recognition while not as stable and accurate as traditional anchor frame-based detectors in complex scenarios may struggle due to its direct regression to the bounding box Unsupervised detection models also lack recognition accuracy and reliability making them unsuitable for high-accuracy detection in wheat crop pest and disease scenarios Segmentation models can impact inference speed and may not meet the real-time detection needs in wheat crop pest scenarios existing supervised detection methods often create overly complex models to enhance feature extraction leading to overfitting and weak generalization when dealing with small sample sizes In the field of crop pest and disease identification the detection accuracy and detection speed of the identification algorithms play a crucial role in the early screening and intervention of crop pests and diseases and in order to adapt to that challenge and demand the relevant researchers have developed a number of specifically optimized algorithms for crop pest and disease identification and proposed a novel identification method for winter wheat pests and diseases on which the attention is used for improvement and the migration learning technique is used to improve the training process thus effectively improving the classification accuracy of winter wheat pests and diseases In the above research in the field of crop pest and disease identification a variety of optimization algorithms for wheat and rice pest and disease detection techniques have emerged but most of the current algorithms still need to improve their detection accuracy and robustness in complex and variable real crop growth scenarios the above algorithms mainly focus on the replacement of the backbone network and there is a lack of enhancement of feature extraction and the introduction of contextual semantic information we will focus on the feature representation of complex pests and diseases the fusion network of pest attention features and the capture of contextual semantic information of pests and diseases focusing on the enhancement of the recognition ability of complex pests and microscopic pests and diseases Specifically, for the image F input to RReNet, which has been represented by a complex pest feature representation network to represent the image features, the output features of the last four layers are noted as: RReNet feeds the input images sequentially into a feature representation network a attention feature fusion network and a detection unit for complex disease and pest detection this four-layer feature outputs the feature notation after going through the pest and disease feature focus mechanism: the output features after going through the attention feature fusion network are noted as: Where Conv23 represents the convolutional module with kernel size 3 in step 2, ReLU represents the activation function, BN represents the normalization, Sigmoid represents the activation function, MaxPool represents the global maximum pooling layer, and Conv3 represents the kernel size 3 in step 1. Convolutional layer with kernel size 3 and step size 1. W0 is a one-dimensional convolution whose represents the convolutional layer with kernel size 1 in step 1. Individual feature extraction and representation layer in complex pest and disease feature representation network Simplified computational procedure for convolutional and pooling layers in network We consider pre-interacting the multi-scale features from different layers of the complex pest and disease feature representation network before sending them to the attention feature fusion network for feature fusion so as to obtain a better feature representation capability when they are fused in the attention feature fusion network This pre-interaction enables shallow features to incorporate global semantic guidance from deeper layers while preserving spatial details, thereby enhancing the network’s sensitivity to tiny pests and complex disease patterns. The detailed process is as follows. The complex pest and disease representation network outputs four hierarchical features {P2, P3, P4, P5}, where F2(shallowest layer) captures fine-grained spatial details (e.g., pest edges, lesion textures), and F5 (deepest layer) encodes high-level semantic information (e.g., disease categories, contextual relationships). Each feature Pi (i = 2,3,4,5) is first processed by the Efficient Channel Attention Network (ECANet) to generate channel-wise attention weights. This step adaptively enhances critical channels related to pest/disease characteristics while suppressing irrelevant ones. Deep-to-shallow interaction is performed to propagate semantic information from P5 to shallower layers. Specifically: P5 is upsampled to match the resolution of P4, then fused with P4 through element-wise addition: After pre-interaction, each layer’s feature contains both its original hierarchical information and cross-level contextual cues. For example, P2 (shallowest) now embeds semantic hints from P5. This pre-interaction ensures that features fed into the subsequent attention fusion network are already semantically enriched and spatially refined, significantly improving the detection of overlapping pests and subtle disease symptoms. Structure of pest and disease characterization concern mechanism Structure of the transformer block in the first detection unit the loss function is defined as formula (9) where the classification loss is defined as LC and the confidence loss is defined as LF The regression loss uses our proposed Fast IoU both LC and LF are calculated by the cross-entropy loss function and Focal Loss The CIoU loss does not accurately capture the relationship between the aspect ratios of the real frame and the predicted frame hindering improvements in detection accuracy we have developed a regression loss function called Fast IoU (fast regression intersection and merger ratio loss function) center point loss and aspect ratio loss follow CIoU the predicted frame and the minimum closure region of the two frames d2 and d3 into the Sigmoid function for mapping; the second step find the ratio of d3 to |d1−d2| to the value of d3/|d1−d2| introduce the parameter D that can be trained to update and multiply with W The third step is to introduce the parameter D that can be trained to update and multiply with W The first step is to calculate the area of the real frame and the minimum closure region of the two frames This study utilized a 64-bit Ubuntu 18.04 operating system with an NVIDIA A100 model GPU to expedite image processing The network training employed Mosaic as the data augmentation method with a BatchSize of 32 and an initial image size of 640 × 640 The model utilized SGD for gradient optimization with an initial learning rate of 0.01 and annealing cosine for learning rate updates The algorithm’s strength and detection accuracy were evaluated using Precision (P) Focus on feature converged networks vs Table 1 demonstrates the comparison results of our proposed RReNet method with other SOTA methods on the LWDCD2020Detection dataset using P, R, F1, and mAP. as can be seen from Table 1 our proposed RReNet achieves the optimal result of 98.3% on mAP and improves by 7.6% compared to the worse YOLOv3 model RReNet can be very effective in wheat pest and disease detection scenarios which is a small improvement over the second-stage Faster RCNN and a small gap with models such as YOLOv7 but it still has a large gap (5.8% mAP) with our proposed RReNet which is a greater illustration of the fact that our proposed RReNet is not only more effective than the one-stage YOLO series in the wheat leaf pest detection task but also achieves better recognition accuracy than two-stage detectors such as Faster RCNN and CenterNet without anchor frame RReNet performs well in complex pest and disease feature detection which is reflected through the following three aspects: RReNet realizes the pre-interaction between deep semantic information and shallow detailed features through the complex pest and disease feature representation module and feature attention mechanism This design compensates the problem of insufficient interaction between deep and shallow features in existing methods (e.g. especially in detecting tiny pests in wheat leaves the combination of shallow features (edges textures) and deep semantics (contextual information) significantly improves the feature representation ability The attention feature fusion network of RReNet contains a four-layer structure (compared with the three layers of FPN and PANet) which supports more complex multi-scale feature fusion and dynamically adjusts the feature weights through the mechanisms of channel attention (ECA-Net) and spatial attention to inhibit the redundant information and enhances the ability of sensing tiny targets YOLO series is prone to lose details of shallow features due to high downsampling rate; CenterNet is not stable enough in complex background due to direct regression to bounding box Through the pest and disease feature attention mechanism RReNet introduces the ECANet module in the feature interaction process to adaptively strengthen the key channel information and at the same time enhances the capture of complex contextual semantics by utilizing the long-distance dependency modeling capability of the Transformer detection unit although models such as YOLOv8 introduce attention mechanisms (e.g. they do not realize the dynamic guidance of cross-layer features the Transformer module is embedded in the detection unit to optimize the relationship between local and global features through the self-attention mechanism which solves the limitation of traditional convolutional networks in long-range dependency modeling and is especially suitable for the decentralized distribution scenario of leaf diseases In order to confirm the efficacy of our proposed modules, we systematically replaced the YOLOv5 baseline with our method. Ablation experiments were carried out on the intricate pest and disease feature representation network (P), pest and disease feature attention mechanism (A), attention feature fusion network (F), and detection unit (D) as shown in Table 2 we noticed a rise in mAP on the LWDCD2020Detection dataset We conducted incremental performance tests of our proposed modules, including complex pest and disease feature representation network (P), pest and disease feature focusing mechanism (A), focusing feature fusion network (F), and detection unit (D), based on the baseline. Table 2 illustrates the enhancement in performance achieved by incorporating individual components Our proposed method demonstrates a significant increase in accuracy compared to the baseline We first replaced the backbone network with the complex pest and disease feature representation network (P) on top of the baseline and the accuracy improved to some extent (2.1% mAP improvement for the LWDCD2020Detection dataset) On the basis of replacing the complex pest and disease feature representation network (P) we added the pest and disease feature attention mechanism (A) which achieved a total accuracy improvement of 3.0% mAP thus improving the recognition ability of RReNet in the face of wheat pests and diseases we proposed the attention feature fusion network (F) and RReNet achieved a total improvement of 3.9% mAP this also illustrates the effectiveness of the attention feature fusion network (F) in cross-layer information fusion which improves RReNet by a total of 4.1% mAP Table 3 presents the comparison results of our Fast-IoU method with other IoU methods on various metrics using the LWDCD2020Detection dataset These results were obtained by testing the YOLOv5L model and 1.2% higher compared to the worse Alpha-IoU It can be obtained from the experiment that Fast-IoU adds a new area loss term (based on the area relationship between the real frame the predicted frame and the minimum closure area) on the basis of CIoU and dynamically adjusts the weights through the trainable parameters so that the loss function optimizes the IoU and it is more comprehensive than the loss of only focusing on a single dimension The introduction of the area loss term is particularly suitable for scenarios with large differences in target dimensions in wheat disease detection the size difference between wheat aphids (tiny) and root rot (larger) is significant and Fast-IoU more accurately balances the regression weights of different targets through an area-sensitive mechanism by introducing a trainable scaling factor (parameter D) Fast-IoU is able to adapt itself to target size changes under different data distributions which improves the flexibility and generalization ability of bounding box regression This paper addresses the limitations of current methods in identifying wheat pests and diseases by incorporating contextual semantic information and key feature attention The study delves into feature representation and attention mechanisms in wheat pest and disease identification RReNet comprises four components: a complex pest and disease feature representation network a pest and disease feature attention mechanism The complex RReNet includes a complex pest and disease feature representation network a pest and disease feature focusing mechanism The complex pest and disease feature representation network extracts intricate disease and pest features while the pest and disease feature focusing mechanism interacts with features at different levels The focusing feature fusion network combines multi-level features and the detection unit incorporates a transformer module to focus on long-distance information enhancing the accuracy and robustness of wheat leaf pest and disease recognition Tests on a challenging wheat leaf pest dataset with twelve pest types demonstrate that RReNet achieves optimal recognition accuracy compared to the state-of-the-art method Ablation experiments further validate the effectiveness of the proposed method’s components The datasets used and analysed during the current study available from the corresponding author on reasonable request Wheat leaf disease identification based on deep learning algorithms Lightweighted wheat leaf diseases and pests detection model based on improved YOLOv8 Recognition of rice leaf diseases and wheat leaf diseases based on multi-task deep transfer learning ImageNet classification with deep convolutional neural networks Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. Preprint at http://arXiv.org/1409.1556 (2014) Deep residual learning for image recognition In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 770–778 (2016) Aggregated residual transformations for deep neural networks In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 1492–1500 (2017) In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 4700–4708 (2017) Fruit classification model based on improved Darknet53 convolutional neural network In 2020 International Conference on Intelligent Transportation Big Data & Smart City (ICITBS) 881–884 (IEEE Feature pyramid networks for object detection In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2117–2125 (2017) & Efficientdet Scalable and efficient object detection In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 10781–10790 (2020) Path aggregation network for instance segmentation In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 8759–8768 (2018) Liu, S., Huang, D. & Wang, Y. Learning spatial fusion for single-shot object detection. Preprint at http://arXiv.org/1911.09516 (2019) In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 7263–7271 (2017) In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 779–788 (2016) Wang, C. Y., Yeh, I. H. & Liao, H. Y. M. Yolov9: Learning what you want to learn using programmable gradient information. Preprint at http://arXiv.org/2402.13616 (2024) In Proceedings of the IEEE International Conference on Computer Vision 1440–1448 (2015) Faster r-cnn: towards real-time object detection with region proposal networks R-fcn: object detection via region-based fully convolutional networks Centernet: Keypoint triplets for object detection In Proceedings of the IEEE/CVF International Conference on Computer Vision 6569–6578 (2019) FCOS: Fully convolutional one-stage object detection In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV’20) (2020) & Cornernet Detecting objects as paired keypoints In Proceedings of the European Conference on Computer Vision (ECCV) 734–750 (2018) Bottom-up object detection by grouping extreme and center points In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 850–859 (2019) FCN-based approach for the automatic segmentation of bone surfaces in ultrasound images Segnet: A deep convolutional encoder-decoder architecture for image segmentation Swin-unet: Unet-like pure transformer for medical image segmentation In European Conference on Computer Vision 205–218 (2022) Applications of machine learning methods to genomic selection in breeding wheat for rust resistance Leaf and Spike wheat disease detection & classification using an improved deep convolutional architecture Novel method for identifying wheat leaf disease images based on differential amplification convolutional neural network Recognizing wheat aphid disease using a novel parallel real-time technique based on mask scoring RCNN 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE) 1372–1377 (IEEE A novel convolutional-recurrent hybrid network for Sunn pest-damaged wheat grain detection Identification of winter wheat pests and diseases based on improved convolutional neural network Deep learning-based object detection improvement for tomato disease Tomato diseases and pests detection based on improved Yolo V3 convolutional neural network An efficient tomato-detection method based on improved YOLOv4-tiny model in complex environment Rice pest detection based on MobileNetv2-YOLOv4 In 2022 China Automation Congress (CAC) 2499–2503 (IEEE Rice pest recognition method based on improved YOLOv8 2024 4th International Conference on Consumer Electronics and Computer Engineering (ICCECE) 418–422 (IEEE Maize-YOLO: a new high-precision and real-time method for maize pest detection Improved Pest-YOLO: Real-time pest detection based on efficient channel attention mechanism and transformer encoder A transformer-based model with feature compensation and local information enhancement for end-to-end pest detection A lightweight pest detection model for drones based on transformer and super-resolution sampling techniques ECA-Net: Efficient channel attention for deep convolutional neural networks In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 11534–11542 (IEEE Distance-IoU loss: Faster and better learning for bounding box regression In Proceedings of the AAAI Conference on Artificial Intelligence 12993–13000 (IEEE YOLOAPPLE: augment Yolov3 deep learning algorithm for Apple fruit quality detection A detection algorithm for Cherry fruits based on the improved YOLO-v4 model Nuts&bolts: YOLO-v5 and image processing based component identification system YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 7464–7475 (2023) Fruit ripeness identification using YOLOv8 model Focal and efficient IoU loss for accurate bounding box regression alpha-IoU: A family of power intersection over union losses for bounding box regression Download references School of Information Science and Technology School of Computer and Communication Engineering conceived and designed the study and made significant contributions during the revision stage conducted experiments and prepared the figures Download citation DOI: https://doi.org/10.1038/s41598-025-99027-3 It’s been a wild offseason for the UCLA Bruins football team. In this age of college sports, every offseason is a turbulent one. Bruin head coach DeShaun Foster reminded everyone of that on Saturday during his media availability I think every team in America probably had a lot of moving pieces because now we have a portal,” Foster said “Everybody can reload or get rid of people so it’s just basically what college football is and I think we just did a good job of finding a way to get these guys to gel.” “You can just tell he has presence,” Foster said of Nico Iamaleava he’s a huge guy and he’s not someone who’s just seeking out attention but just kind of has a poise to him Iamaleava completed 63.6% of his passes with Tennessee Iamaleava had 27 total touchdowns to just five interceptions The 20-year-old quarterback has all the traits you could look for now it’s just time for Iamaleava to prove he’s one of the top QBs in college football Metrics details We developed an experimental approach to compare how attentional orienting facilitates retrieval from spatial working memory (WM) and long-term memory (LTM) and how selective attention within these two memory types impacts incoming sensory information processing In three experiments with healthy young adults retrospective attention cues prioritize an item represented in WM or LTM Participants then retrieve a memory item or perform a perceptual task The retrocue is informative for the retrieval task but not for the perceptual task We show that attentional orienting benefits performance for both WM and LTM Eye-tracking reveals significant gaze shifts and microsaccades correlated with attention in WM but no statistically significant gaze biases were found for LTM Visual discrimination of unrelated visual stimuli is consistently improved for items matching attended WM locations Similar effects occur at LTM locations but less consistently The findings suggest at least partly dissociable attention-orienting processes for different memory types Although our conclusions are necessarily constrained to the type of WM and LTM representations relevant to our task attentional prioritization in LTM can operate independently from WM Future research should explore whether similar dissociations extend to non-spatial or more complex forms of LTM thus underscoring the strong role of WM in guiding external attention then the same internal attention mechanisms should select relevant mnemonic content for behaviour focusing on LTM content to guide retrieval may also occur directly and independently we introduce an experimental framework to test the ability to orient internal spatial attention selectively among competing LTM contents and to compare directly selective attention in WM and LTM We borrow from retrospective cue (referred to as “retrocue” hereafter) designs to direct attention to a specific item within a pre-learned array of items in LTM or to a specific item of an encoded WM array presented earlier in the trial we measured the consequences of orienting spatial attention in WM vs LTM for memory retrieval and additionally tested for spill-over effects on sensory processing participants either retrieved a memory item or performed a perceptual discrimination task The retrocue was informative for the retrieval task but not for the perceptual task This allowed us to investigate how orienting selective attention in WM and LTM enhanced the retrieval of relevant memoranda and biased perceptual processing in an unrelated task in similar ways The findings demonstrate the significant effects of internal attention in WM and LTM and show that dissociable mechanisms are at play Orienting attention in LTM consistently speeds selective retrieval and only sometimes improves retrieval accuracy Benefits of internal attention in WM are stronger with larger gains for response speed and consistent accuracy improvements Eye-tracking reveals notable differences in engagement of oculomotor mechanisms during shifts of attention within WM vs Perceptual discrimination of unrelated visual stimuli is consistently superior at locations associated with LTM items and is modulated by internal attention in both WM and LTM domains linear contrasts showed that learning performance improved significantly for both colour reproduction (F(1 partial η2 = 0.054) and shape reproduction (F(1 attentional orienting exerted stronger effects in WM than LTM featuring higher benefits in both retrieval speed and accuracy In experiment 1 (Fig. 2b) LTM) significantly impacted retrieval speed: RTs in the memory retrieval task were faster when retrocues were informative (F(1 partial η2 = 0.742) and when WM items were probed (F(1 retrocueing and memory conditions also interacted (F(1 Both WM and LTM retrocues conferred a significant benefit (WM: t(29) = 9.217 but retrocue benefits were stronger for the speed of retrieving WM items (t(29) = 4.316 Similar analyses on retrieval quality showed significant main effects of retrocueing (F(1 partial η2 = 0.130) and memory conditions (F(1 Post-hoc comparisons revealed a significant improvement in retrieval accuracy by retrocues for WM items (t(29) = 2.981 d = 0.465) but no significant effect of retrocues for LTM items (t(29) = 1.530 This could possibly be explained by the overall very high retrieval accuracy for LTM (0.977 ± 0.005) A shortcoming of experiment 1 was the spatial nature of the memory-retrieval report Upon presentation of an informative retrocue participants could immediately prepare their response thus confounding the quality and speed of item selection with response preparation The non-spatial nature of experiments 2 and 3 combined with using a randomly oriented response wheel In experiment 2 (Fig. 4b) retrocueing significantly shortened RTs (F(1 Although RTs were equivalent for both memory conditions (F(1 there was a significant interaction between the two factors (F(1 RT benefits were present for both WM (t(43) = 11.657 d = 0.273) but were stronger for WM (t(43) = 4.957 partial η2 = 0.173) and memory conditions (F(1 partial η2 = 0.434) exerted significant main effects Post-hoc comparisons revealed that WM retrocues significantly reduced shape reproduction errors (t(43) = 4.322 d = 0.429) but LTM retrocues did not (t(43) = 1.250 and errors were smaller when LTM shapes were reproduced (16.610° ± 1.519) as compared to WM shapes (27.603° ± 2.417) In experiment 3 (Fig. 5c) the one-day interval between training and testing brought LTM performance below ceiling allowing us to test for LTM accuracy benefits by internal attention the next-day LTM representations in experiment 3 were weaker than the same-day representations in experiment 2 (reproduction errors in experiment 3 vs Informative retrocues significantly improved RTs (F(1 Response times did not show statistically significant differences between memory condition (F(1 p = 0.603) but the two factors interacted (F(1 Post-hoc t tests showed that RT benefits were present for both WM (t(118) = 10.080 d = 0.253) but were stronger for WM (t(118) = 3.556 partial η2 = 0.320) and memory condition (F(1 partial η2 = 0.056) both exerted significant main effects and interacted (F(1 Post-hoc comparisons revealed that informative retrocues significantly reduced shape reproduction errors for both WM items (t(118) = 9.581 Error reduction was stronger for WM items (t(118) = 3.901 The results from experiment 3 thus showed that attentional orienting can improve LTM retrieval accuracy when overall LTM performance is below ceiling The current study also yielded insights on how prioritising items in WM or LTM impacts sensory information processing at the item location In experiment 1 (Fig. 2c) selective prioritisation of WM and LTM memoranda by retrocues impacted perceptual discrimination of items occurring at the matching location Perceptual discrimination accuracy showed a main effect of retrocue matching (F(1 with superior accuracy for discriminating stimuli at retrocue-matching locations than for stimuli on neutral-retrocue trials This accuracy benefit was significantly larger than zero for both memory conditions (WM: 0.094 ± 0.020 d = 0.797) and there was no statistically significant difference in size (t(29) = 0.962 We also observed a main effect of memory condition (F(1 Accuracy was higher when LTM locations were probed (0.655 ± 0.026) compared to WM locations (0.601 ± 0.032) A separate analysis found no statistically significant effect of congruence between item location and arrow-pointing direction (t(29) = 0.894 Improvement of perceptual discrimination in experiment 1 could have resulted from response preparation related to reporting item location rather than prioritisation of an item in WM or LTM. In experiment 2 (Fig. 4c) there was no direct link between spatial location and responses in the memory retrieval task so that any effect on perceptual discrimination could be linked to prioritisation of items in WM or LTM perceptual discrimination still benefitted at locations matching the retrocued items (F(1 Perceptual benefits for items matching retrocued WM and LTM locations were both significantly larger than zero (WM: 0.037 ± 0.003 d = 0.328) and there was no statistically significant difference in size (t(43) = 0.259 discrimination performance was better at LTM than WM locations overall (F(1 A separate analysis found no statistically significant effect of congruence between item location and arrow-pointing direction (t(43) = 0.354 In experiment 3 (Fig. 5d) the main effect of retrocue matching did not reach significance (F(1 Memory condition still exerted a main effect with perceptual discrimination better at LTM than WM locations (F(1 Retrocueing and memory condition interacted (F(1 Perceptual benefits of WM retrocues were significantly larger than zero (t(118) = 3.308 while benefits from LTM retrocues were not (t(118) = 0.896 a Time courses of horizontal gaze biases following neutral time courses are shown separately for conditions in which the retrocued item occupied the left side (top/bottom left location) or the right side (top/bottom right location) during encoding b Time courses of horizontal gaze bias towardness on trials with WM and LTM retrocues c Time courses of vertical gaze biases following neutral time courses were shown separately for conditions in which the retrocued item occupied the top side (top left/right location) or the bottom side (bottom left/right location) during encoding d Time courses of vertical gaze bias towardness on trials with WM and LTM retrocues away microsaccade rate in the horizontal channel Horizontal lines below the time courses in b d and e indicate significant temporal clusters (green lines: WM compared to zero; black lines in b and d: WM vs LTM difference; black line in e: toward vs Shading areas in the time courses in a–f represent ±1 SEM The differences between gaze biases following WM and LTM retrocues were quantified using the associated towardness time courses Cluster-based permutation testing identified significant differences following WM and LTM retrocues in both horizontal and vertical directions (horizontal: ~450–1000 ms after cue onset p = 0.002; vertical: ~420–870 ms after cue onset No statistically significant clusters occurred following LTM retrocues microsaccades occurred more often in the direction of the prioritised item than the other item (~380–490 ms after cue onset No statistically significant directional changes in microsaccade rates occurred after LTM retrocues We introduced an approach for investigating and comparing selective attention operating in LTM vs WM using equivalent stimulus parameters and response requirements within the same tasks orienting attention to contents in LTM significantly improved retrieval speed in a similar Accuracy benefits were more consistent for WM than LTM No statistically significant improvements in LTM accuracy occurred when retrieval performance was at very high levels overall (experiments 1 and 2) but improvements emerged when there was greater room for improvement in retrieval accuracy (experiment 3) The current pattern of results suggests that attention can consistently enhance the accessibility or “readiness” to act on LTM memory representations The modulations observed for WM compared to LTM retrieval were stronger shortening response times to a greater degree and consistently affecting retrieval accuracy The findings suggest differences in the mechanisms of internal attention within these different memory domains these behavioural results could result from variations in the strength of the memory representations themselves and the resulting retrieval demands we observed that gaze was biased toward retrocued items in WM no statistically significant gaze biases were detected following LTM retrocues Engagement of attention-related frontal areas has been less conspicuous established plasticity patterns related to the longer-term associations may confer alternative or additional mechanisms for prioritising feature values of LTM items Informative LTM retrocues may interact with latent functional states that have been intrinsically reinforced by associative plasticity The findings suggest that the spatial layout may also be a useful or intrinsic preserved property of encoded LTM arrays A combination of these accounts is also possible suggesting that focusing within LTM may rely on more flexibly activating the most relevant attributes of prior experiences to guide behaviour whereas WM representations are necessarily spatially tethered Greater variability of online performance may also have contributed It will be interesting to explore further the boundary conditions for perceptual spillover effects of internal attention in LTM Stronger activation of feature-specific sensory signals during attention to WM could interfere with processing other incoming visual stimuli this explanation would not account for better performance for discriminating stimuli at attended compared to unattended WM locations The mechanisms behind the intriguing findings warrant further study whereas retrieval in visual WM tasks often invokes spatial codes even when they are not strictly necessary for task performance the dependence on spatial frameworks for retrieving information from LTM may depend on the task requirements Our specific task conditions yielded a clear difference in the reliance on spatial codes when retrieving contents from WM vs LTM where spatial retrieval is an integral part of the LTM task the dissociation may not have been apparent spatial codes may be used in both WM and LTM when spatial retrocues are used or when participants are required to retrieve or reproduce the locations of stimuli within their original contexts in WM or LTM Further investigation will likely reveal a much richer array of interacting spatial frameworks in WM tasks we compared the effects of focusing attention within human WM vs LTM and discovered that both bring significant benefits to retrieval and subsequent sensory processing The distinct oculomotor signatures of covert spatial attention in WM and LTM corroborate a plurality of functional properties when memories of different durations guide adaptive behaviour and open novel opportunities for furthering our understanding of the relationship between WM and LTM SD = 4.15) with reported normal or corrected visual acuity volunteered and received monetary compensation for participation The experiment was approved by the Central University Research Ethics Committee of the University of Oxford and all participants provided informed consent before any experimental procedure began Stimuli appeared overlaid on a grey background four squares (2.5° in diameter) were always presented as placeholders at the four quadrants at 5° horizontally and vertically from the central fixation to the centre of each square The stimuli consisted of four equiluminant colours (brown [183.5 112.5]) drawn from a circle in CIELAB colour space The experiment included a learning session and a testing session, separated by a 5-minute break. During the learning session, participants were trained to encode two colours and their corresponding locations into LTM (Fig. 1) These two colours were randomly selected from the four colours defined above and they were always located along one of the two pairs of diagonal locations (i.e. We refer to this pair of locations as LTM locations Each learning trial began with a fixation display lasting randomly between 800 and 1000 ms after which the LTM display was presented for 150 ms participants were probed to reproduce either the colour at one location or the location of one colour a colour wheel (containing 360 colours) was presented at the centre and participants responded by rotating the dial and selecting a colour along the wheel The colour wheel was presented in a random orientation on every trial feedback was presented for 1000 ms in the form of an integer ranging from 0 to 100 with 100 indicating a perfect reproduction of the probed colour and 0 indicating the exact opposite on the wheel one of the two colours was presented at the centre and participants responded by pressing one of four keys mapped to the four locations (Q for top left Performance feedback was presented for 500 ms indicating whether the chosen location was correct or incorrect Each to-be-learned attribute (two colours and two locations) was probed on 20 trials resulting in a total of 80 learning trials presented in random order During the testing session, participants performed either a memory retrieval task or a perceptual discrimination task on each trial (Fig. 2a) Each testing trial began with a fixation display (800–1000 ms) where the two colours unused in the learning session were presented for 150 ms at the unused pair of diagonal locations We refer to this pair of locations as WM locations To make sure the contents in WM were not fixed across trials each WM colour was randomly assigned to one of the WM locations on every trial a retrocue that was either neutral or informative was presented for 200 ms The retrocue was neutral on one-third of the trials; the fixation display changed to white providing no information about the item to be probed The retrocue was informative on two-thirds of the trials; the fixation display changed to one of the four colours matching either a WM or LTM item with equal probability Informative retrocues indicated the item to be probed in the memory retrieval task with 100% validity Following a second delay of 800 ms after the retrocue the memory retrieval task and the perceptual discrimination tasks were equally likely to be presented participants were required to retrieve the location of the probed item On trials containing an informative retrocue participants reported the location of the WM or LTM item indicated by the retrocue the probed item was indicated by a centrally presented colour chosen randomly from the WM or LTM colours Responses were delivered by pressing one of four keys mapped to the four locations (the same keys as used in the location reproduction task during the learning session) indicating whether the response was correct or wrong 128]) were presented in the placeholders for 100 ms after which randomly generated Gaussian noise masks were applied to the four locations for 100 ms and participants pressed one of the arrow keys to report the arrow direction at that location Each of the four locations had an equal possibility of being probed The arrow direction at each location was independently drawn from four possible directions (↑ The choice of presenting arrow stimuli at all four locations and using a post-cue to elicit a response was intended to avoid the sensory capture by stimuli with different attributes Participants were informed that the memorised items and the retrocue bore no predictive relation concerning the location of the perceptual item to be discriminated participants always needed to maintain the retrocued item in mind for potential future use because of the randomisation of the memory retrieval and perceptual discrimination trial order which allowed us to examine any spill-over benefits elicited by the retrocue on the perceptual discrimination task The relationship between the perceptual discrimination task and the retrocue was totally incidental because all locations were equally likely to be probed no matter which location the retrocue would preferentially bias attention to only 25% of informative-retrocue trials in the perceptual discrimination task were “matching” trials on which the probed sensory location coincided with the location of the retrocued item The testing session consisted of 480 trials divided into 10 blocks (each including 48 trials) participants performed an additional 48 practice trials before testing we examined the average colour reproduction errors and location reproduction accuracy by sorting the learning trials of each type into 4 bins (each containing 10 trials) Colour reproduction errors (in units of degrees) were calculated by taking the absolute difference between the angle of the target colour and the reproduced colour on the colour wheel One-way repeated-measures ANOVAs with linear contrast weights ([−3 3]) across four bins tested for the efficacy of training data from the memory retrieval and perceptual discrimination tasks were analysed separately we excluded trials on which RTs were 3 SD above the individual mean across all conditions in either task an average of 98.35% (SD = 0.44%) trials were retained in the analyses To test for benefits of internal selective attention on WM and LTM retrieval we analysed the average RTs and accuracy for the memory retrieval task as a function of retrocueing (neutral vs informative) and the memory timescale of probed items (WM vs To examine whether orienting attention to a memory item benefited subsequent perceptual processing at matching locations we compared perceptual discrimination accuracy on retrocue matching trials vs neutral) when locations associated with WM or LTM items were probed (WM vs accuracy was the dependent variable of interest but RTs were also evaluated for completeness When evaluating potential perceptual benefits elicited by WM and LTM retrocues The effect sizes for these comparisons in experiment 1 were 0.877 and 0.797 We assumed a conservative approach and aimed to power for the detection of a medium effect size (0.5) because we expected that the manipulation in experiment 2 would lead to a smaller effect due to the incidental nature of the spatial attributes in the task Most of the experimental setup was identical to experiment 1 and then randomly assigned to each WM and LTM item adding a new feature dimension to the existing configurations in experiment 1 These same four shapes were used across all participants but the shapes assigned to WM and LTM items were randomised across participants During the learning session, participants were trained to memorise the colours and shapes of the two LTM items (Fig. 3) they were probed to reproduce either the colour or the shape of one item either a colour wheel or a shape wheel was presented at the centre indicating the feature dimension to be reproduced in this trial The spatial location of the item was used to probe the colour or shape reproduction but participants were never asked to report the item location Both the colour and shape wheels were presented in a random orientation each time The shape wheel consisted of 360 shapes from the VCS space eight shapes sampled from equidistant positions on the wheel were displayed along the cardinal axes (i.e. These eight shapes served as visual anchors which were also randomly chosen every time based on the orientation of the shape wheel Participants responded using a computer mouse that controlled the dial on the wheel Participants had unlimited time to retrieve the item from memory and to decide what to reproduce they had only 2500 ms to complete their reproduction This was intended to encourage participants to retrieve the exact colour or shape before moving the dial The position of the dial when participants clicked the left mouse button or when the time limit was reached was taken as the response Each colour and shape was probed on 20 trials During the testing session, each WM shape was randomly combined with one of the WM colours on every trial (Fig. 4a) participants performed either a memory retrieval task or a perceptual discrimination task on each trial equiprobably participants reproduced the shape of the item matching the retrocued colour or a randomly probed colour when the retrocue was neutral The shape wheel was identical to that used during the learning session and was randomly rotated across trials grey neutral retrocues appeared on one-fifth of the trials The remaining trials contained informative coloured retrocues matching each of the LTM or WM colours with equal probability the memory retrieval task in experiment 2 was non-spatial Retrieving the item shape based on the cued colour does not require using any spatial association Although spatial locations were used for training the colour-shape associations participants were never cued or asked to report the item location in the memory retrieval task The perceptual discrimination task was the same as that in experiment 1 The testing session consisted of 600 trials divided into 10 blocks (each including 60 trials) participants performed an additional 30 practice trials before testing The analyses of interest were basically the same as in experiment 1 with location reproduction in the learning and testing sessions replaced by shape reproduction Shape reproduction errors (in units of degrees) were calculated by taking the absolute difference between the angle of the target shape and the reproduced shape on the shape wheel The RTs in the memory retrieval task were calculated as the time from probe onset to when the response was recorded either when participants clicked the left mouse button or when the time limit was reached After excluding memory retrieval and perceptual discrimination trials on which RTs were 3 SD above the individual mean across all conditions an average of 98.66% (SD = 0.47%) trials were retained in the analyses with an average of 9.6% ± 1.7% (M ± SEM) trials excluded per participant six participants had to be removed due to a high number of excluded trials (>50%) an average of 14.6% ± 2.1% trials were excluded Gaze time courses were smoothed using a 25-ms average moving window We did this for both horizontal and vertical channels to obtain towardness in horizontal and vertical directions we focused exclusively on microsaccades along the horizontal axis Depending on the side of the retrocued item we labelled microsaccades as “toward” or “away” based on whether they were moving toward or away from the retrocued item The resulting time courses of “toward” and “away” microsaccade rates were smoothed using a moving average with a 50-ms sliding window Statistical evaluation of the towardness time courses used a cluster-based permutation approach79 implemented in the permuco package which is ideally suited to evaluate physiological effects across multiple time points while retaining high sensitivity Given the assumption that LTM would decay on the second day, we aimed to power for detecting an effect size of 0.3 for perceptual benefits, which led G*power to yield a sample size of 119 participants. Participants (86 males, 33 females, M = 28.69 years, SD = 6.04; 106 right-handed, 12 left-handed, 1 ambidextrous) were recruited on Prolific (https://www.prolific.co/) They were pre-screened on demographic criteria (age range 18 to 40 general health (normal or corrected-to-normal vision and participation history on Prolific Academic (participated in at least 10 studies All participants provided informed consent before participating and were paid $12 per hour Participants performed 80 learning trials on the first day which was identical to the learning session in experiment 2 Participants were invited to complete the testing session around 24 hours after training The differences in the testing session compared to experiment 2 are as follows the duration of the perceptual array was lengthened to 200 ms The testing session consisted of 480 trials divided into 12 blocks (each including 40 trials) Participants performed an additional block of 40 practice trials before the 480 testing trials we excluded memory retrieval and perceptual discrimination trials with RTs exceeding 3 SD above the individual mean across all conditions an average of 98.65% (SD = 0.61%) trials were retained in the analyses Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article Neural mechanisms of selective visual attention In Stevens’ Handbook of Experimental Psychology and Cognitive Neuroscience (ed Control of goal-directed and stimulus-driven attention in the brain A taxonomy of external and internal attention Turning attention inside out: how working memory serves behavior Orienting attention to locations in internal representations Large capacity storage of integrated objects before change blindness In search of the focus of attention in working memory: 13 years of the retro-cue effect Microsaccades uncover the orientation of covert attention Microsaccades as an overt measure of covert attention shifts Human gaze tracks attentional focusing in memorized visual space Goal-directed and stimulus-driven selection of internal representations Multiple spatial frames for immersive working memory Overlapping mechanisms of attention and spatial working memory Interactions between visual working memory and selective attention Dissociating the neural mechanisms of memory-based guidance of visual selection Contextual cueing in naturalistic scenes: global and local contexts Contextual cueing: Implicit learning and memory of visual context guides spatial attention Incidental biasing of attention from visual long-term memory Long-term memory prepares neural activity for perception Orienting attention based on long-term memory experience What to expect where and when: how statistical learning drives visual selection Reading scenes: how scene grammar guides attention and aids perception in real-world environments The parietal cortex and episodic memory: an attentional account Memory: enduring traces of perceptual and reflective attention Top-down and bottom-up attention to memory: a hypothesis (AtoM) on the role of the posterior parietal cortex in memory retrieval The contribution of the human posterior parietal cortex to episodic memory Posterior parietal cortex and episodic retrieval: convergent and divergent effects of attention and memory Alpha-band oscillations track the retrieval of precise spatial representations from long-term memory Pattern reinstatement and attentional control overlap during episodic long-term memory retrieval Human memory: a proposed system and its control processes In: Psychology of Learning and Motivation (eds What are the differences between long-term Visual working memory buffers information retrieved from visual long-term memory Long-term memory and working memory compete and cooperate to guide attention Functional but not obligatory link between microsaccades and neural modulation by covert spatial attention Premembering experience: a hierarchy of time-scales for proactive attention The cognitive neuroscience of working memory Orienting attention to locations in perceptual versus mental representations Frontoparietal and cingulo-opercular networks play dissociable roles in control of working memory Different states in visual working memory: when it guides attention and when it does not Automatic guidance of attention from working memory Searching for targets within the spatial layout of visual short-term memory The privileged role of location in visual working memory Neural architecture for feature binding in visual working memory Location and binding in visual working memory A review of visual memory capacity: Beyond individual items and toward structured representations Two types of representation in visual memory: evidence from the effects of stimulus contrast on image combination Feature-specific reaction times reveal a semanticisation of memories over time and with repeated remembering and hippocampal contributions to visual working memory maintenance and associative memory retrieval Perceptual difficulty regulates attentional gain modulations in human visual cortex Perceptual learning: toward a comprehensive theory How does the brain learn environmental structure Ten core principles for understanding the neurocognitive mechanisms of statistical learning Different features of real-world objects are represented in a dependent manner in long-term memory Feature binding in short-term memory and long-term learning The locus of recognition memory signals in human cortex depends on the complexity of the memory representations Eye movements and visual memory: detecting changes to saccade targets in scenes Shared representational formats for information maintained in working memory and information retrieved from long-term memory Cowan, N. Attention and Memory: An Integrated Framework. https://doi.org/10.1093/acprof:oso/9780195119107.001.0001 When natural behavior engages working memory Reference frames for spatial cognition: different brain areas are involved in viewer- and landmark-centered judgments about object location The use of egocentric and allocentric reference frames in static and dynamic conditions in humans Independent working memory resources for egocentric and allocentric spatial information Higher level visual cortex represents retinotopic Memory for retinotopic locations is more accurate than memory for spatiotopic locations Power 3: a flexible statistical power analysis program for the social The validated circular shape space: quantifying the visual similarity of shape No obligatory trade-off between the use of space and time for working memory Nonparametric statistical testing of EEG- and MEG-data The timing mega-study: comparing a range of experiment generators Gong, D., Draschkow, D. & Nobre, A. C. Focusing attention in working and long-term memory through dissociable mechanisms. Open Science Framework https://doi.org/10.17605/osf.io/n629s (2025) Gong, D., Draschkow, D. & Nobre, A. C. Focusing attention in working and long-term memory through dissociable mechanisms. Zenodo https://doi.org/10.5281/zenodo.14968822 (2025) Download references This research was funded by a Clarendon Scholarship a Medical Research Council Studentship and a New College-Yeotown Scholarship to D.G.; a Wellcome Trust Senior Investigator Award (104571/Z/14/Z) and a James S McDonnell Foundation Understanding Human Cognition Collaborative Award (220020448) to A.C.N.; and by the NIHR Oxford Health Biomedical Research Centre The Wellcome Centre for Integrative Neuroimaging is supported by core funding from the Wellcome Trust (203139/Z/16/Z) the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission Wellcome Centre for Integrative Neuroimaging Nature Communications thanks Edyta Sasin and the other reviewer(s) for their contribution to the peer review of this work Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations Download citation DOI: https://doi.org/10.1038/s41467-025-59359-0 This article was first published in the On the Hill newsletter. Sign up to receive the newsletter in your inbox on Friday mornings here Happy May and welcome to the very first edition of On The Hill I’ll still be bringing you all the same content and pulling back the curtain on the goings-on in Congress It will be based on my experiences walking the halls of Congress every day chasing lawmakers down hallways and revealing the authentic atmosphere of Capitol Hill as it happens — and now the name will reflect that It was a big week on Capitol Hill as House Republicans began their four-week sprint to draft and finalize President Donald Trump’s massive tax reconciliation package obstacles remain for GOP leaders to meet their self-imposed deadline on Memorial Day let’s check in on Democrats — and how they commemorated (read: mourned) Trump’s 100th day in office It’s no secret that Democrats have been struggling with an identity crisis over the last few months After a November election that left them with no control in Washington the party has scrambled to consolidate behind a clear message Democrats tried to put the focus on Medicaid and Social Security saying the programs could be trimmed as the GOP looks for ways to save money in the tax reconciliation bill But that message hasn’t seemed to resonate with voters who gave Democrats low approval ratings in the most recent polls Voters are looking for a stronger message as signs of restlessness within the base become more evident Some Democrats appear to be catching on — and are trying to up their game as they enter the next 100 days of Trump’s term It began with a 12-hour sit-in on the Capitol steps on Sunday hosted by House Minority Leader Hakeem Jeffries to protest the Trump administration’s actions so far particularly in relation to upcoming spending cuts a sing-along and activists who joined them This week the Democratic leader held three separate press conferences targeting Trump’s 100 days in addition to the Capitol steps sit-in House leaders typically only hold one media availability a week a group of Democratic senators kept the floor open all night on Tuesday giving speeches to highlight what they consider chaos caused by the Trump administration over the last three months President Trump has made Americans less safe Senate Democrats also forced a vote on a resolution disapproving of Trump’s tariffs, pressuring Republicans to go on the record on the issue. That resolution ultimately failed, but it was close — and three Republicans joined Democrats to oppose the tariffs Or is the initial excitement from Booker’s marathon speech wearing off It may be too early to tell. Recent approval ratings for Democrats have been at record lows, which gives them some work to do But: Republicans aren’t faring much better — which could give them some wiggle room Trump’s approval rating has also been falling in recent weeks with only 39% of voters saying they approve of job the president is doing according to a recent poll from ABC News/Washington Post/Ipsos Those numbers are especially low on how they view his handling of the economy it’s a long way until November 2026 when Democrats hope to win back at least one of the chambers in Congress — and as Republicans hope to defend their trifecta Whatever happens now may not have any influence by then Or it could be the start of a long-term strategy that leads to victory Among the many components causing snags in Republicans’ tax reconciliation package is the lingering question: What to do with former President Joe Biden’s Inflation Reduction Act Some Republicans want to repeal it in full — and they are threatening to tank the bill altogether if that’s not done A group of 38 House Republicans sent a letter to Ways and Means Committee Chairman Jason Smith and Speaker Mike Johnson all CC’d) demanding that the IRA is completely overturned have previously told me this is a red line for them “Republicans ran—and won—on a promise to completely dismantle the IRA and end the left’s green welfare agenda,” the lawmakers wrote “How do we retain some of these credits and not operate in hypocrisy …If every faction continues to defend their favored subsidies we risk preserving the entire IRA because no clearly defined principle will dictate what is kept and what is culled.” If you’ve been reading this newsletter for a while, you know this is an issue I’ve done quite a bit of reporting on. (Read last week’s edition where I walked through Utah Sens Mike Lee and John Curtis’ differing opinions on the matter.) But this letter shows that Republicans’ reconciliation efforts are far from over Some Republicans have made clear they want to preserve some green energy tax credits in Biden’s IRA it will increase utility costs for everyday Americans Remember: Republican lawmakers need to find $1.5 trillion in spending cuts to pay for the tax cut extensions in the reconciliation bill That could put the entirety of the IRA on the chopping block anyway just to foot the bill From the courts: Will the Supreme Court accept religious charter schools? … Disability-rights arguments grow heated at Supreme Court (AP) The House and Senate are out for the weekend and are set to return on Monday House Republicans will continue efforts to finalize their reconciliation bill with key committees such as Ways and Means and Energy and Commerce expected to hold their markups sometime in the next two weeks The Senate is patiently waiting for their chance to get their hands on those pieces of legislation As always, feel free to reach out to me by email with story ideas or questions you have for lawmakers. And follow me on X for breaking news and timely developments from the Hill Roshan Kenia presented a poster on how AI-CNet3D enhances glaucoma classification using cross-attention networks while improving interpretability and performance in OCT scan analysis At the ARVO 2025 meeting in Salt Lake City Roshan Kenia presented a poster on AI-CNet3D a cross-attention network for glaucoma classification Editor's note: The below transcript has been lightly edited for clarity It's a cross-attention network for glaucoma classification they're condensed into the reports that ophthalmologists typically use through their glaucoma diagnosis But there's a lot of rich structural information in the 3D data that isn't typically used and it's very time consuming to look at that's a convolutional [neural] network embedded with attention And the basic idea is glaucoma can be a symmetric disease where it affects both the superior and inferior nerve What we do is we split the volume in half based on the superior and inferior nerve and then we use cross attention to attend one of those halves to the other halves we embed that into a convolutional network And what we see is we get very good performance compared to other other types of networks And then the next step was we wanted our model to be interpretable which is a channel attention representation And what that allows us to do is visualize what the attention layers are using to make their diagnosis We can also use Grad CAM to use to visualize what the convolutional layers are using to make their diagnosis And we can see here it's very spread out and it's not very informative What we can do is enforce a consistency between the two so that the network is more interpretable and then the output of the heat map itself So we train our model further to enforce consistency between the two of these and then what we see is the performance increases even more we're regularizing the data or regularizing the model and so then when we evaluate it on more test data it performs better because it's just more regularized Don’t miss out—get Ophthalmology Times updates on the latest clinical advancements and expert interviews Subrata Batabyal, PhD, receives 2025 Carl Camras Translational Research Award from ARVO Foundation California dreaming: 2025 Glaucoma 360 brings celebration, innovation, and education to San Francisco Editorial Advisory Board members to present at Retina World Congress 2025 Connect, learn, and innovate in a family-friendly atmosphere: What to expect at EnVision Summit 2025 ASCRS 2025: Visual and patient-reported outcomes following bilateral Odyssey implantation iSTAR Medical announces positive 5-year data from STAR-GLOBAL trial for MINIject 609-716-7777 It’s the first week of no Minnesota Wild hockey so Judd and AJ look at some of the biggest points of attention for the Wild this … Minnesota #Vikings QB Sam Darnold has been arguably the best quarterback against the blitz this season Minnesota #Vikings QB Sam Darnold joins Brett Favre and Daunte Culpepper with ten games of a passer rating of 100+ in the same season... ESPN analyst and former NFL quarterback Alex Smith says the #Vikings are making a huge mistake if they don’t re-sign Sam Darnold The Vikings QB room is looking VERY different than it has in previous years 👀 JUST IN: The #Vikings are signing former Giants QB Daniel Jones Minnesota #Vikings star WR Justin Jefferson has no problems letting everyone else go off for big games When everyone is freaking out about the #Vikings going to overtime but you’re just a chill guy who’s happy to be watching a 9-2 fo.. Jordan Addison gets the #Vikings on the board Jonathan Greenard and Andrew Van Ginkel are not only monsters on the field for the Minnesota #Vikings; they’re making a combined $9.1.. Minnesota #Vikings All-Pro Jared Allen is a Semifinalist for the NFL Hall of Fame Is this the year he finally gets the call? Minnesota #Vikings QB Sam Darnold is one of the most accurate deep-ball passers in the NFL this season Sam Darnold is the reason the Vikings aren’t getting the respect they deserve Minnesota #Vikings QB Sam Darnold posted the highest PFF grade of his NFL career in Sunday’s win over the Titans Your Minnesota #Vikings are 8-2 through ten games Need help accessing the FCC Public File due to a disability Please contact Ross Brendell at publicfilemsp@hubbardradio.com or (651) 632-6675 This web site is not intended for users located within the European Economic Area Add Comment|1$TRUMPOfficial Trump$11.111.28%Stock Score Locked: Want to See it?Benzinga Rankings give you vital metrics on any stock – anytime President Donald Trump is leaning into crypto as a fundraising tool, with two events projected to generate millions in donations, despite his claim of earning no personal profit from the sector a $1.5 million-per-plate fundraiser on May 5 and will feature prominent tech investor and ‘crypto czar’ David Sacks targets holders of the Trump-themed meme coin OFFICIAL TRUMP TRUMP/USD The top 220 token holders will be granted dinner access The announcement has fueled speculation and price action with the TRUMP token soaring over 50% following the news CNBC reported on Monday that the token has generated over $324 million in trading fees with roughly 80% of the supply linked to Trump-affiliated wallets Critics have raised concerns that the setup allows wealthy and possibly foreign actors to buy access to the former president The official site notes that Trump's attendance isn't guaranteed and reserves the right to cancel the event for any reason in which case attendees will receive a Trump NFT instead Also Read: Bitcoin’s Key Level Is $96,500, Analyst Warns, But Here’s What Could Trigger A Breakdown Why It Matters: In a Sunday interview with ‘Meet the Press’ moderator Kristen Welker President Trump stated plainly that he does not benefit from crypto while stressing the strategic importance of the U.S He acknowledged crypto's growing popularity and resilience particularly in contrast to recent turbulence in traditional markets When asked about the meme coin bearing his name, Trump claimed he was unaware about any of the details. He declined to commit to donating any crypto-linked earnings, though he noted that he already donates his presidential salary, a sizable gesture he says not any other has made. Image created using artificial intelligence with Midjourney. Stock Score Locked: Want to See it?Benzinga Rankings give you vital metrics on any stock – anytime Momentum-Price TrendShortMediumLongOverviewThis content was partially produced with the help of AI tools and was reviewed and published by Benzinga editors Benzinga does not provide investment advice Posted In: CryptocurrencyNewsTop StoriesAI GeneratedDonald TrumpBenzinga simplifies the market for smarter investingTrade confidently with insights and alerts from analyst ratings free reports and breaking news that affects the stocks you care about and trade ideas delivered to your inbox every weekday before and after the market closes Cheeky Prince Louis made his father hear his complaints during the VE Day celebrations Louis out on another adorable cheeky display when he got bored and pulled his dad’s aiguillette (braid) of the RAF No.1 uniform He also brushed his father’s shoulder a few times the William continued to talk to someone seated on his other side Louis seemed to be complaining to William about how boring he found the procession to be He then explained the route of the procession of armed forces to Louis "It's not going to be long and I want you to be taking an interest," William asserted The event marked the first appearance of Prince George Princess Charlotte and Prince Louis together since Christmas Day George will also attend the tea party King Charles will host later today for the last surviving members of the World War II generation 2025 2:26 PM EDTCaitlin Clark left Iowa as one of the most decorated college basketball players in the history of the game Clark racked up just about every personal award possible all while setting the all-time scoring records - in men's or women's NCAA history Following her record-setting Hawkeyes career she became the No which she helped to the playoffs for the first time in eight years After winning Rookie of the Year in her first season in the league Clark is making a brief return to Iowa before her second WNBA season kicks off She and the Fever are hitting the court against Brazil at Carver-Hawkeye Arena in Iowa City whipped out a tribute for the Iowa superstar "Lexie Hull reppin' Caitlin Clark's 22 for today's game in Iowa," the Fever said showing a photo of Hull's mid-riff revealing outfit The comments second was flooded with love from fans over the look "Lexie looks great reppin 22," one person said "Our girls are soooo beautiful Rep that Lex!!!" offered another "YEA BUDDY REP IT LEXIE," added a third Clark missed the team's first preseason game against the Washington Mystics with a leg injury but head coach Stephanie White loved her communication from the bench "Her communication was outstanding," White said and I felt like at one point she wanted to go back and put her uniform on and come out and play she's had a lot of progress in the last 24 hours We'll see what happens and how she feels after the workouts that she went through today she'll be ready to go [Sunday]." an augmented reality (AR) game released on TikTok Mashreq’s recent SMASHREQ campaign has proven to be a masterclass in using innovative digital content to capture the attention of Gen Z the demographic that has proven to be notoriously elusive for brands Here's how this game not only grabbed attention but created a cultural phenomenon that brands can learn from Check out the Best Gaming & eSports Influencer Marketing Campaigns Mashreq’s SMASHREQ campaign was introduced as the first-ever padel game on TikTok, taking advantage of the platform’s immersive AR capabilities to engage users in a fun and interactive experience Designed to connect Mashreq with the growing padel community the game’s primary objective was to raise brand awareness and solidify Mashreq’s position as a prominent player in the sports and wellness space A post shared by Mashreq (@mashreq) Unlike typical social media ads this initiative integrated seamlessly into TikTok’s culture making it feel less like an advertisement and more like an entertainment experience that users actively wanted to engage with The game itself was a simple yet captivating concept: users could participate in a digital padel game directly within the app and the AR elements enhanced the immersion making the gameplay feel interactive and dynamic Mashreq was able to create something that was not only fun to play but also highly shareable While engagement metrics are often a crucial indicator of a campaign’s success SMASHREQ’s performance is especially noteworthy the campaign delivered a staggering 230 million impressions outperforming over 40 other games released on the platform in the past two years One of the key factors behind this success was the game’s exceptional dwell time users spent 200 seconds engaging with SMASHREQ a remarkable 143% higher than the typical dwell time seen with other TikTok content This level of engagement is not just impressive—it’s a testament to how well the game resonated with the audience keeping them hooked and coming back for more The Gen Z audience is highly sought after for its social media influence but capturing their attention is no easy feat Gen Z values interactive and personalized content over passive consumption SMASHREQ's gamified format hit the mark by offering a compelling interactive experience that users could engage with repeatedly a sport that aligns with Gen Z's love for dynamic The game allowed users to actively participate in challenges making them feel more connected to the content while also providing a sense of accomplishment through score tracking To fuel further engagement, Mashreq collaborated with top athletes from the UAE Padel Association and the Egyptian Padel Federation These collaborations were not only authentic but also played a pivotal role in driving the game’s popularity Influencers and athletes promoted the game through challenges and by encouraging their followers to beat their high scores This strategy created a sense of community around the game and led to organic promotion as users were motivated to engage with the game and share their progress with their own social circles These strategic partnerships were key in amplifying the game’s reach turning it into more than just a branded campaign but a viral sensation that resonated with the core values of TikTok’s community: entertainment Check out the Top 8 Virtual Reality (VR) & Augmented Reality (AR) Social Media Marketing Campaigns Timing is everything when it comes to digital campaigns and Mashreq smartly aligned the launch of SMASHREQ with the peak social traffic during Ramadan tapping into the increased online activity during this period Mashreq ensured maximum reach and engagement from users who were already active on TikTok during Ramadan which led to a significant surge in game participation Related Content:TikTok Money Calculator [Influencer Engagement & Earnings Est...12 Examples of Influencer Marketing on TikTok (Case Studies)44 Vital TikTok Stats to Inform Your Marketing Strategy What sets SMASHREQ apart from other branded games is its ability to transcend beyond a marketing campaign and become a cultural phenomenon. The high engagement levels, coupled with user-generated content and the competitive nature of the game helped the game become a social media sensation Users weren’t just playing the game; they were sharing their high scores Mashreq has shown that when you create something that resonates with the target audience—whether it's through gamification, influencer collaboration or strategic timing—the results can be phenomenal The campaign’s success isn’t just about numbers; it’s about building a community and making a lasting impact on your audience Check out the Top Virtual Reality (VR) & Augmented Reality (AR) Influencer Marketing Campaigns The SMASHREQ campaign is a shining example of how brands can harness the power of gamification and digital engagement to connect with younger Mashreq’s SMASHREQ game isn’t just a marketing success; it’s a case study in how to connect with Gen Z through gamification By tapping into the growing trend of interactive content Mashreq created a campaign that resonated with users and left a lasting impact on TikTok SMASHREQ provides a roadmap for how to effectively blend gaming Forget vague predictions—here’s what’s actually moving the needle in beauty.. YouTube celebrates two decades of transforming the social media and.. Content creators are constantly battling copyright claims over music usage on YouTube It’s the first week of no Minnesota Wild hockey so Judd and AJ look at some of the biggest points of attention for the Wild this summer What sort of reinforcements are out there to add a quality center to the mix What prospects would you part with if that was the case Learn more about your ad choices. Visit podcastchoices.com/adchoices Hield poured in 33 points in the 103-89 drubbing of Houston, probably the biggest factor in getting the Warriors through to the second round of the playoffs and a date with the Minnesota Timberwolves.  But Green's pouting behavior after the Warriors' Game 6 loss to Houston was one of the big storylines, notably because he owned up to it in a team meeting and subsequently changed his approach in time to lead his team in Game 7.  Golden State Warriors forward Draymond Green (23) greets Houston Rockets guard Fred VanVleet (5) after game seven of the first round for the 2025 NBA Playoffs at Toyota Center.  Yes, Green was doling out hugs to his foes.  "Draymond Green greeting Alperen Sengun, Fred VanVleet, Ime Udoka, Amen Thompson and Steven Adams after the series. Pretty friendly postgame scene in Houston," Slater wrote on Twitter/X.  Coach Steve Kerr said that Green helped reverse the momentum of the series, which saw Houston win Games 5 and 6 after the Warriors grabbed a 3-1 lead, by taking ownership of his inability to control his emotions in the losses. And he changed that in Game 7.  "He's our leader,” Kerr said. “When he’s right, like he was tonight, he is an incredible player to watch. The defense, just kind of owning the court on that side of things, and then just being patient and not turning it over and being in the right spot offensively.” The conference semifinals vs. the Minnesota Timberwolves will tip off at Target Center on Tuesday.  By Expertise: NBA horse racingEducation: Northwestern University The series includes research and commentary on actionable steps democracy actors can take to strengthen democratic institutions and protect freedoms in the U.S and we therefore separate the two here to highlight both complementary challenges we examine the 232 races for state legislature and supreme court in 2025 post-May 1 that have yet to take place across 15 states we specifically focus on the subset of those races that we have identified as higher risk for election subversion; in section 2 we look at the races that have the highest propensity for voter suppression Pennsylvania and Virginia elections fall under both categories and so increased attention should be paid to races in those states 189 elections across four states could change the balance of power in those states with all 80 seats of the General Assembly in New Jersey and all 100 seats of the House of Delegates in Virginia up for election this year The General Assembly in New Jersey and the House of Delegates in Virginia each represent one of the two legislative bodies in their respective states meaning their make-up following the elections could keep or break a unified state legislature Outcome-determinative races often garner more attention and thus are often more prone to election subversion where the gubernatorial and legislative branches are currently under Democratic control a change in party control could make or break a trifecta States that fall both within the election subversion category and which have also been identified as having a higher propensity for voter suppression (see subsequent section) are highlighted below in orange State assembly elections have not previously faced the same level of malfeasance they merit attention to ensure that remains the case in 2025 Losing candidates should concede their races and quash any claims of cheating from supporters. In addition, those involved in the ministerial and mandatory duty to certify election results should continue to do so show the highest potential for voter suppression states that have a higher propensity for both election subversion and voter suppression are highlighted in orange In South Carolina, S108 was signed into law in 2022 and made absentee voting more difficult by narrowing the qualification criteria for an absentee ballot increasing the information required on the application and banning the use of drop boxes in the state While two recent laws in Mississippi—S.B. 2576 and H.B. 1406—increased access to the polls by expanding voter ID options, removing affidavit requirements, and establishing a notice-and-cure process for mail-in ballots, the state remains one of the most restrictive (Elections in Mississippi have yet to be scheduled but are currently anticipated for November 2025.) Though it is often difficult to quickly determine the exact causal impact of a new piece of restrictive legislation on voter turnout these races and states present opportunities for increased vigilance and areas for future study Elections, including at the state level, are a fundamental component of American democracy, but they have come under threat in various ways. As Jacob Grumbach analyzed in his State Democracy Index 2.0 (SDI 2.0) electoral democracy in the states was strong in the 2000s but began to decline in the 2010s Recent data shows an uptick in electoral democracy scores in 2023 compared to 2018 but levels still remain well below those in the 2000s (data for 2024 and 2025 are not yet available in the SDI 2.0 analysis) Given the decline since the 2000s and the possibility of state electoral recovery in 2023 it is crucial to continue analyzing the health of state down-ballot elections The races identified in this piece should be monitored to ensure they remain or become safe pro-democracy actors must ensure that this standard is upheld The authors would like to thank Peter Beck and Gigi Liman for research support; Julianna Melendez and Eric Urby for fact-checking and copyediting assistance; and Jordan Muchnick and Eric Urby for editorial assistance Written By:Advertorial Team there’s one name floating around more than usual: Qubetics ($TICS) While everyone’s scrambling to catch the next moonshot this project is slowly (and quietly) setting the stage for something much bigger that juicy potential upside that makes even skeptics do a double-take Qubetics isn’t another clone project—it’s aiming to fix some serious mess-ups that old-school cryptos couldn’t handle From making cross-border payments simpler to offering tools for pros and businesses it’s packing features that go beyond meme magic and empty roadmaps So if you’re sitting there wondering which cryptocurrencies are worth watching today—take notes The Qubetics wallet isn’t just any wallet—it’s non-custodial and built to handle a future that’s anything but simple Whether you’re a casual user or a tech nerd stacking tools for business automation And some of the smarter heads in the game are calling a $10–$15 tag after mainnet launch That’s not fantasy—it’s backed by a growing use case and buzz spreading like wildfire in Telegram groups Qubetics isn’t just about holding a coin and waiting—it’s built to do work where cross-border payments can be slower than dial-up internet Simple: it’s not just a presale—it’s a movement And when a token’s laying down actual infrastructure before hitting the mainnet Avalanche has been on a lowkey grind AVAX has been building partnerships and scaling dApps like nobody’s business Ava Labs partnered up with Stripe to simplify fiat-to-crypto onboarding with projects like Shrapnel and Ascenders catching serious hype More developers are starting to choose Avalanche for its subnets—custom blockchains with tailor-made rules Now here’s where it gets really juicy: Avalanche is positioning itself as a decentralized alternative to AWS It’s already onboarding real-world assets and enterprise use cases that hint it’s not a pipe dream bouncing from under $30 to flirting with $40 it could retest the $50+ zone before mid-year Polkadot’s been quiet—but only if you’re not paying attention The ecosystem just rolled out asynchronous backing for parachains a technical upgrade that could supercharge transaction speeds and scalability DOT’s staking options have become way more flexible More people are jumping in for those tasty passive rewards especially as centralized savings options feel like a joke with today’s inflation And let’s not ignore Moonbeam and Astar—two parachains that are bringing real dev action into the Polkadot mix Cross-chain apps are finally getting some shine Especially with institutional players quietly stacking DOT on the low Here’s the deal with NEAR—it’s smooth as butter This chain’s been gunning for simplicity from day one It just rolled out its Chain Signatures tech letting users interact with other chains without leaving NEAR The foundation announced a massive grant pool for AI x Web3 startups and some early projects are showing serious promise Add that to the fact that transaction fees are dirt cheap and NEAR starts looking like a sleeper altcoin ready to snap out of its nap NEAR’s also integrating with major DeFi protocols to make it easier for folks in places like Tajikistan and Armenia to enter Web3 without needing a PhD in crypto Price-wise, it’s sitting just under $7—but there’s upside, especially if those AI plays take off. XRP doesn’t need much of an intro It’s been fighting legal wars for what feels like forever but 2025 is starting to look like its redemption arc That alone is enough to make XRP one of the Best cryptos to Buy Today Ripple just landed a fresh round of banking partnerships across Asia and the Middle East focusing on remittances and CBDC integrations XRP is also rolling out private ledger integrations that’ll help enterprises move value across borders without touching traditional systems For Central Asia’s remittance-heavy economies Its price is still chilling around $0.60–$0.70 these five tokens—especially Qubetics—are standing tall when it comes to potential returns $TICS at $0.2302 might not stick around for long Watch where the smart money’s moving—and get in while the door’s still open Because it solves real problems with cross-border payments especially for freelancers and small businesses across countries like Kazakhstan Analysts predict up to over six thousand percent ROI if $TICS hits $15 after mainnet launch Qubetics is building from the ground up for modern use cases like automation Stay ahead of the curve with authentic news and exclusive in-sight reports only on The Crypto Times All NewsAI NewsExclusiveExplained About UsEditorial PolicyPress Release Contact UsCareerAdvertise With Us brilliant heroine who makes the mother of dragons look meek This new novel will conclude Pullman’s Book of Dust trilogy, a series that expands on a world first glimpsed in 1995’s The Golden Compass The Golden Compass introduced a frosty parallel universe wherein Oxford College is managed by a dark religious theocracy human souls exist as external animal companions and a mysterious subatomic particle governs all the magic in the universe we met Lyra as a plucky twelve year old in thrall to a prophecy Five books, several novellas, a marathon play having faced down destiny and survived several cosmic battles she starts her final quest in mortal (and spiritual) danger In The Guardian Pullman described his alleged final novel “as partly a thriller and partly a bildungsroman: a story of psychological The power over people’s lives once held by old institutions and governments is seeping away and reappearing in another form: that of money Image via Created by Grove Atlantic and Electric Literature Masthead About Sign Up For Our Newsletters How to Pitch Lit Hub Privacy Policy Support Lit Hub - Become A Member Lit Hub has always brought you the best of the book world for free—no paywall you'll keep independent book coverage alive and thriving Department of Health and Human Services is ignoring decades of research findings But the head of the U.S. Department of Health and Human Services, Robert F. Kennedy Jr., says he believes environmental factors—vaccines chief among them—have caused the increase in autism prevalence. As Axios and other news outlets reported, Kennedy said “One of the things I think we need to move away from today is this ideology that this diagnosis at which time researchers will be awarded grants to study autism causes Kennedy presumably will want researchers to test his theory that the increase in autism is caused by environmental factors such as ultrasound scans or potential hazards such as food additives and pesticides Kennedy’s environmental exposures theory is shockingly naive because it ignores decades of study findings This new research will also waste money better spent on evidence-based hypotheses But at least a test of his theory may prove him wrong Some environmental exposures will result in autism only for people with specific genetic vulnerabilities Improving maternal health can lower these chances. The MMR (measles, mumps, rubella) vaccine eliminated maternal rubella and, with it, fetal congenital rubella syndrome, the cause of at least  10 percent of autism in the 1960s and 1970s Kennedy believes there has been a massive increase in chemicals in the environment that has caused I think it’s unlikely that there could be 30 times more undetected autism than undetected ADHD and 46 times more undetected autism than undetected intellectual disability The “gold standard” to establish an actual increase in autism would be finding a longitudinal increase in the number of people diagnosed using a tightly constrained set of behaviors and a set of reliable and valid autism biomarkers But autism behavior patterns show immense heterogeneity and research has not yet found a reliable set of biomarkers in conjunction with a genetic predisposition An editorially independent publication supported by the Simons Foundation showing herself walking like a model at the beach “Ready to be what my fans need,” Maher captioned Set to the song “Father Figure (Remastered),” the clip gave followers a glimpse of the U.S rugby star enjoying her downtime after a busy season on the field.  the short clip quickly captured the internet's attention as social media users flocked to the comment section to express their awe Fans quickly filled the comments with praise for Maher’s look and message many celebrating her confidence and physique “That suit is iconic!!!!!” one follower wrote “Miss Ma’am… respectfully… you are effing stunning!” another added Comments continued pouring in with admiration not only for the bold swimsuit look but also for Maher’s unapologetic strength “Rugby body here—thanks for normalizing strong women!” one fan commented “You work hard to look like that… Proud of you for showing it off," a user agreed “People need to understand that this is a peak feminine athlete physique “Love that suit – and the silly tan lines it would leave.” The swimsuit video follows a string of recent milestones for the athlete. Maher, who was named one of USA Today's Women of the Year, recently returned from a three-month stint with England’s Bristol Bears she helped boost the team's online presence and brought fresh attention to women’s rugby in the U.K Olympic athlete Ilona Maher poses for a photo at the USOC Media Summit in preparation for the Paris 2024 Olympic Summer Games at Mariott Marquis Maher has become one of the most recognizable faces in women’s rugby she competed in the Pacific Four Series opener against Canada drawing a record-breaking crowd of over 10,000 fans to CPKC Stadium Though the USA Eagles fell short in the match Maher’s energy and connection with supporters stood out “I’ll give you what you give me,” she wrote on Instagram after the match thanking fans for their incredible support Maher’s influence isn’t just about sports She’s also known for using her platform to talk about body image often addressing the pressure female athletes face.  In a previous interview with The Telegraph she opened up about dealing with online criticism and body-shaming comments.  but I’ll continue to fight it off,” she said Maher’s latest post is another example of how she uses her spotlight not only to celebrate herself but to shift perspectives around beauty and strength in women’s sports Parked in a Churchtown car park and listening to the radio which was parked facing the side of my car the bonnet of the Audi was buried in my car door Her attention excursion cost me a mild brain injury Plus the loss of a year’s productive living – that I will never get back The driver just forgot she was driving a car One study by the University of California found that our attention span when looking at a screen has fallen from two and a half minutes in 2003 to 47 seconds in 2023 Today the printed word has been replaced by the screen What was once a serious and coherent discussion of public affairs at the heart of our culture has been trivialised into a constellation of snappy 35-second posts Like a mother feeding regurgitated food to her chicks these bite-size snippets offer little to digest or think about you’re left with a hollow sinking feeling; a sense of having been robbed Traditionally, oil, wheat, gold, coffee, etc are commodities that are traded on global markets. Today our attention is ranked as the most valuable commodity on the planet. Elon Musk and the gang pour billions into creating algorithms that track your likes and dislikes and plunder your precious attention – right under your nose these corporate bandits are on a mission to knock you off-guard decimate your ability to focus and steal your attention I overtook a wobbling cyclist on the busy road I saw he had a sandwich in the hand he was trying to grip his handlebar with Wiping his overstuffed mouth with his elbow the lad was simultaneously glued to his screen we meet the zombies with heads dipped; chin to chest often wearing headphones too (might this be you?) screen junkies strain to keep abreast of the event they’ve come out and paid to see The term nomophobia was coined to describe the irrational fear of being without a mobile phone Smartphone users can experience anxiety or panic when unable to access their device I heard one woman confess on radio that she even brings her (waterproof) device into the shower with her continuous scrolling and hours spent gaming our brains are being wired to restlessly skim and switch As with anyone addicted to something – gambling etc – the phone junkie craves the dopamine hit The mood-altering drug delivered by endless skimming and checking is triggered by the mindless Musk shoots off a minimum of 70 chest-thumping posts on X every day We have free will to think and process our own thoughts but we are in grave danger of being dumbed down into societies of passive nincompoops is just one of the reasons for cyberbullying You wouldn’t send your child out on the M50 to play so how does it seem appropriate to hand anyone under the age of 16 a smartphone [ Barriers to sleep: Research finds 83% of Irish teenagers have their phones in bedrooms at nightOpens in new window ] No one can claim that life is an easy ride and thoughts can be uncomfortable at times isolation and the fear visited on us by Covid we turned to our devices in our droves in search of some relief every click offered a fleeting moment of “relief” (dopamine hit) from our thoughts [ Parents, if you’re going to ban anything, ban devices from bedroomsOpens in new window ] advertising existed in the form of papyrus posters TV and radio adverts designed to win our attention You make a conscious decision to pick up your Irish Times You get to choose what you want to absorb or ignore It‘s every media editor’s job to catch yours when and where to deliver the loudest headline the most outrageous or grating comment or spectacular photo to reel you in – momentarily it‘s up to you to choose to continue to watch what‘s on offer in return The result can assist in the formation of opinions and decision-making and the algorithm has no truck with fluffy benefits to mankind it shamelessly seeks to pave the way for our free will to be plundered By saturating the world with catchy visuals and sounds masked as “entertainment” or “educational content” Even “innocent” retail platforms exist in part to harvest your attention What they actually sell is secondary to the big steal itself But there must be some accountability as to where this buck must be made to stop As I stood by a supermarket food cabinet last week the sight of an under-two-year-old child in its buggy utterly hypnotised by an irresistible colourful screen he shot that tight fist of his up in defiance; and scored the nation’s attention By spewing troughs of what is mostly unchallenged guff he won 49.8 per cent of a highly “entertained” US electorate‘s attention How many of these voters are laughing today Our attention lies at the very core of our humanity raises stress levels and increases mental fatigue how are children supposed to develop intellectually The potential impact of device/gaming addiction on their exam results can only be catastrophic warns that “it‘s highly important that children learn their mathematics well If they don’t learn basic problem-solving skills as children,” life will be very difficult for them “Because they have higher emotional intelligence girls are at a slightly lesser risk here than boys although girls face other serious social media challenges.” these apps are distracting and give instant gratification reels and memes don’t stimulate the important logical deductible way of thinking.” He emphasises: “by not exercising the deductive way of thinking this puts them at serious risk of never developing their automatic skills This in turn puts them at risk of long-term dementias.” Talking on the phone is now completely foreign to young people This limited form of communication automatically lacks the multilayered complexity within the intonation of the human voice Zoom deprives participants of the natural ability to read body language and other non-verbal communication cues worsening posture from supporting lowered heads – not to mentioned overactive thumbs – and upper-thoracic back pain has reached epidemic proportions In Brave New World (1932) Aldous Huxley depicted a future in which universal happiness is only achieved by thoroughly dehumanising humanity His fear was that we might one day be flooded with information Huxley worried that the undoing of people‘s capacity to think would deprive them of their autonomy and maturity and history; rendering them vulnerable to manipulation One modern book I’ve found useful is US TV host and journalist Chris Hayes’s recently published The Sirens’ Call: How Attention Became the World’s Most Endangered Resource Hayes shreds the veil that conceals the billionaires’ rampage to steal our compromised supply of attention he warns of the dire consequences that await if we don’t stop the onslaught and take back control of our free will – while we still can Hayes says: “Now our deepest neurological structures human evolutionary inheritances and social impulses are in a habitat designed to prey upon or destroy that which most fundamentally makes us human.” [ Laura Kennedy: https://www.irishtimes.com/abroad/2025/04/23/laura-kennedy-yes-smartphone-addiction-is-unhealthy-but-so-is-getting-a-dumb-phone-and-pretending-its-2003/Opens in new window ] Going straight for the jugular, it‘s Hayes’s mission to put a halt to this creeping evil that is attention-harvesting. Nuala Macklin is a journalist and documentary maker Facebook pageTwitter feed© 2025 The Irish Times DAC