Sci Rep. 2025 May 12;15(1):16409. doi: 10.1038/s41598-025-00116-0.

ABSTRACT

Accurate cortical surface parcellation is essential for elucidating brain organizational principles, functional mechanisms, and the neural substrates underlying higher cognitive and emotional processes. However, the cortical surface is a highly folded complex geometry, and large regional variations make the analysis of surface data challenging. Current methods rely on geometric simplification, such as spherical expansion, which takes hours for spherical mapping and registration, a popular but costly process that does not take full advantage of inherent structural information. In this study, we propose an Attention-guided Deep Graph Convolutional network (ADGCN) for end-to-end parcellation on primitive cortical surface manifolds. ADGCN consists of a deep graph convolutional layer with a symmetrical U-shaped structure, which enables it to effectively transmit detailed information of the original brain map and learn the complex graph structure, help the network enhance feature extraction capability. What's more, we introduce the Squeeze and Excitation (SE) module, which enables the network to better capture key features, suppress unimportant features, and significantly improve parcellation performance with a small amount of computation. We evaluated the model on a public dataset of 100 artificially labeled brain surfaces. Compared with other methods, the proposed network achieves Dice coefficient of 88.53% and an accuracy of 90.27%. The network can segment the cortex directly in the original domain, and has the advantages of high efficiency, simple operation and strong interpretability. This approach facilitates the investigation of cortical changes during development, aging, and disease progression, with the potential to enhance the accuracy of neurological disease diagnosis and the objectivity of treatment efficacy evaluation.

PMID:40355465 | DOI:10.1038/s41598-025-00116-0

Med Phys. 2025 May 12. doi: 10.1002/mp.17883. Online ahead of print.

ABSTRACT

BACKGROUND: Deep learning networks map input data to output predictions by fitting network parameters using training data. However, applying a trained network to new, unseen inference data resembles an interpolation process, which may lead to inaccurate predictions if the training and inference data distributions differ significantly.

PURPOSE: This study aims to generally improve the prediction accuracy of deep learning networks on the inference case by bridging the gap between training and inference data.

METHODS: We propose an inference-specific learning strategy to enhance the network learning process without modifying the network structure. By aligning training data to closely match the specific inference data, we generate an inference-specific training dataset, enhancing the network optimization around the inference data point for more accurate predictions. Taking medical image auto-segmentation as an example, we develop an inference-specific auto-segmentation framework consisting of initial segmentation learning, inference-specific training data deformation, and inference-specific segmentation refinement. The framework is evaluated on public abdominal, head-neck, and pancreas CT datasets comprising 30, 42, and 210 cases, respectively, for medical image segmentation.

RESULTS: Experimental results show that our method improves the organ-averaged mean Dice by 6.2% (p-value = 0.001), 1.5% (p-value = 0.003), and 3.7% (p-value < 0.001) on the three datasets, respectively, with a more notable increase for difficult-to-segment organs (such as a 21.7% increase for the gallbladder [p-value = 0.004]). By incorporating organ mask-based weak supervision into the training data alignment learning, the inference-specific auto-segmentation accuracy is generally improved compared with the image intensity-based alignment. Besides, a moving-averaged calculation of the inference organ mask during the learning process strengthens both the robustness and accuracy of the final inference segmentation.

CONCLUSIONS: By leveraging inference data during training, the proposed inference-specific learning strategy consistently improves auto-segmentation accuracy and holds the potential to be broadly applied for enhanced deep learning decision-making.

PMID:40356014 | DOI:10.1002/mp.17883

BMC Med Educ. 2025 May 12;25(1):691. doi: 10.1186/s12909-025-07259-8.

ABSTRACT

BACKGROUND: Dentistry is shifting from traditional to digital practices owing to the rapid development of "artificial intelligence" (AI) technology in healthcare systems. The dental curriculum lacks the integration of emerging technologies such as AI, which could prepare students for the evolving demands of modern dental practice. This study aimed to assess dental faculty members' knowledge, awareness, and attitudes toward AI and provide consensus-based recommendations for increasing the adoption of AI in dental education and dental practice.

METHOD: This mixed-method study was conducted via a modified version of the General Attitudes toward Artificial Intelligence Scale (GAAIS) and Focus Group Discussions (FGD). Four hundred faculty members from both public and private dental colleges in Pakistan participated. The quantitative data were analyzed using SPSS version 23. Otter.ai was used to transcribe the data, followed by thematic analysis to generate codes, themes, and subthemes.

RESULTS: The majority of the faculty members was aware of the application of AI in daily life and learned about AI mainly from their colleagues and social media. Fewer than 20% of faculty members were aware of terms such as machine learning and deep learning. 81% of the participants acknowledged the need for and limited opportunities to learn about AI. Overall, the dental faculty demonstrated a generally positive attitude toward AI, with a mean score of 3.5 (SD ± 0.61). The benefits of AI in dentistry, the role of AI in dental education and research, and barriers to AI adoption and recommendations for AI integration in dentistry were the main themes identified from the FGD.

CONCLUSIONS: The dental faculty members showed general awareness and positive attitudes toward AI; however, their knowledge regarding advanced AI concepts such as machine learning and deep learning was limited. The major barriers identified in AI adoption are financial constraints, a lack of AI training, and ethical concerns for data management and academics. There is a need for targeted education initiatives, interdisciplinary and multi-institutional collaborations, the promotion of local manufacturing of such technologies, and robust policy initiatives by the governing body.

PMID:40355937 | DOI:10.1186/s12909-025-07259-8

J Transl Med. 2025 May 12;23(1):532. doi: 10.1186/s12967-025-06507-1.

ABSTRACT

BACKGROUND: Accurate assessment of expected survival in melanoma patients is crucial for treatment decisions. This study explores deep learning-based body composition analysis to predict overall survival (OS) using baseline Computed Tomography (CT) scans and identify fully volumetric, prognostic body composition features.

METHODS: A deep learning network segmented baseline abdomen and thorax CTs from a cohort of 495 patients. The Sarcopenia Index (SI), Myosteatosis Fat Index (MFI), and Visceral Fat Index (VFI) were derived and statistically assessed for prognosticating OS. External validation was performed with 428 patients.

RESULTS: SI was significantly associated with OS on both CT regions: abdomen (P ≤ 0.0001, HR: 0.36) and thorax (P ≤ 0.0001, HR: 0.27), with lower SI associated with prolonged survival. MFI was also associated with OS on abdomen (P ≤ 0.0001, HR: 1.16) and thorax CTs (P ≤ 0.0001, HR: 1.08), where higher MFI was linked to worse outcomes. Lastly, VFI was associated with OS on abdomen CTs (P ≤ 0.001, HR: 1.90), with higher VFI linked to poor outcomes. External validation replicated these results.

CONCLUSIONS: SI, MFI, and VFI showed substantial potential as prognostic factors for OS in malignant melanoma patients. This approach leveraged existing CT scans without additional procedural or financial burdens, highlighting the seamless integration of DL-based body composition analysis into standard oncologic staging routines.

PMID:40355935 | DOI:10.1186/s12967-025-06507-1

BMC Cardiovasc Disord. 2025 May 12;25(1):362. doi: 10.1186/s12872-025-04818-1.

ABSTRACT

BACKGROUND: Heart failure (HF) after acute myocardial infarction (AMI) is a leading cause of mortality and morbidity worldwide. Accurate prediction and early identification of HF severity are crucial for initiating preventive measures and optimizing treatment strategies. This study aimed to develop an interpretable artificial intelligence (AI) model for HF severity prediction using multidimensional clinical data.

METHODS: This study included data from 1574 AMI patients, including medical history, clinical features, physiological parameters, laboratory test, coronary angiography and echocardiography results. Both deep learning (TabNet, Multi-Layer Perceptron) and machine learning (Random Forest, XGboost) models were employed in constructing model. Additionally, the Shapley Additive Explanation (SHAP) method was used to elucidate clinical factors importance and enhance model interpretability. A web platform ( https://prediction-killip-gby.streamlit.app/ ) was also developed to facilitate clinical application.

RESULTS: Among the models, TabNet demonstrated the best performance, achieving an AUROC of 0.827 for KILLIP four-class classification and 0.831 for KILLIP binary classification. Key clinical factors such as GRACE score, NT-pro BNP, and TIMI score were highly correlated with KILLIP classification, aligning with established clinical knowledge.

CONCLUSIONS: By leveraging easily accessible multidimensional data, this model enables accurate early prediction and personalized diagnosis of HF risk and severity following AMI. It supports early clinical intervention and improves patient outcomes, offering significant clinical application value.

CLINICAL TRIAL NUMBER: Not applicable.

PMID:40355836 | DOI:10.1186/s12872-025-04818-1

Aging Clin Exp Res. 2025 May 13;37(1):149. doi: 10.1007/s40520-025-03048-x.

ABSTRACT

BACKGROUND: Osteoporosis is often underdiagnosed due to limitations in traditional screening methods, leading to missed early intervention opportunities. AI-driven screening using chest radiographs could improve early detection, reduce fracture risk, and improve public health outcomes.

AIMS: To assess the cost-effectiveness of deep learning models (hereafter referred to as AI-driven) applied to chest radiographs for opportunistic osteoporosis screening in German women aged 50 and older.

METHODS: A decision tree and microsimulation Markov model were used to calculate the cost per quality-adjusted life year (QALY) gained (€2024) for screening with AI-driven chest radiographs followed by treatment, compared to no screening and treatment. Patient pathways were based on AI model accuracy and German osteoporosis guidelines. Women with a fracture risk below 5% received no treatment, those with 5-10% risk received alendronate, and women 65 + with a risk above 10% received sequential treatment starting with romosozumab. Data was validated by a German clinical expert, incorporating real-world treatment persistence, DXA follow-up rates, and treatment initiation. Sensitivity analyses assessed parameter uncertainty.

RESULTS: The cost per QALY gained from screening was €13,340, far below the typical cost-effectiveness threshold of €60,000. Optimizing follow-up, treatment initiation, and medication adherence further improved cost-effectiveness, with dominance achievable by halving medication non-persistence, and in women aged 50-64.

CONCLUSION: AI-driven chest radiographs for opportunistic osteoporosis screening is a cost-effective strategy for German women aged 50+, with the potential to significantly improve public health outcomes, reduce fracture burdens and address healthcare disparities. Policymakers and clinicians should consider implementing this scalable and cost-effective screening strategy.

PMID:40355760 | DOI:10.1007/s40520-025-03048-x

Nat Methods. 2025 May 12. doi: 10.1038/s41592-025-02698-z. Online ahead of print.

ABSTRACT

Light-field microscopy (LFM) and its variants have significantly advanced intravital high-speed 3D imaging. However, their practical applications remain limited due to trade-offs among processing speed, fidelity, and generalization in existing reconstruction methods. Here we propose a physics-driven self-supervised reconstruction network (SeReNet) for unscanned LFM and scanning LFM (sLFM) to achieve near-diffraction-limited resolution at millisecond-level processing speed. SeReNet leverages 4D information priors to not only achieve better generalization than existing deep-learning methods, especially under challenging conditions such as strong noise, optical aberration, and sample motion, but also improve processing speed by 700 times over iterative tomography. Axial performance can be further enhanced via fine-tuning as an optional add-on with compromised generalization. We demonstrate these advantages by imaging living cells, zebrafish embryos and larvae, Caenorhabditis elegans, and mice. Equipped with SeReNet, sLFM now enables continuous day-long high-speed 3D subcellular imaging with over 300,000 volumes of large-scale intercellular dynamics, such as immune responses and neural activities, leading to widespread practical biological applications.

PMID:40355725 | DOI:10.1038/s41592-025-02698-z

Sci Rep. 2025 May 12;15(1):16481. doi: 10.1038/s41598-025-01810-9.

ABSTRACT

This work proposes a personalized music learning platform model based on deep learning, aiming to provide efficient and customized learning recommendations by integrating audio, video, and user behavior data. This work uses Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks to extract audio and video features, while using multi-layer perceptrons to encode user behavior data. To further improve the recommendation accuracy, this work constructs a knowledge graph that integrates key entities and their relationships in the music field, and fuses them with the extracted feature vectors. The knowledge graph provides the platform with rich semantic information and relational data, helping the model better understand the correlation between user needs and music content, thereby improving the accuracy and personalization of recommendation results. Experimental analysis based on different datasets shows that the proposed music recommendation platform performs well in multiple key performance indicators. Especially under different TOP-K conditions, the accuracy reaches 0.90, significantly exceeding collaborative filtering and content-based recommendation methods. In addition, the platform can maintain high accuracy when processing sparse data, demonstrating stronger robustness and adaptability. The platform has significant advantages in overall performance, providing users with more reliable and efficient recommendation services.

PMID:40355699 | DOI:10.1038/s41598-025-01810-9

J Imaging Inform Med. 2025 May 12. doi: 10.1007/s10278-025-01534-2. Online ahead of print.

ABSTRACT

Pain assessment is a critical aspect of medical care, yet automated systems for clinical pain estimation remain rare. Tools such as the visual analog scale (VAS) are commonly used in emergency departments (EDs) but rely on subjective self-reporting, with pain intensity often fluctuating during triage. An effective automated system should utilize objective labels from healthcare professionals and identify key frames from video sequences for accurate inference. In this study, short video clips were treated as instance segments for the model, with ground truth (physician-rated VAS) provided at the video level. To address the weak label problem, we proposed flexible multiple instance learning approaches. Using a specialized loss function and sampling strategy, our instance-appraisable model, EDi Pain, was trained to estimate pain intensity while evaluating the significance of each instance segment. During inference, the VAS pain score for the entire video is derived from instance-level predictions. In retrospective analysis using the public UNBC-McMaster dataset, the EDi Pain model demonstrated competitive performance relative to prior studies, achieving strong performance in video-level pain intensity estimation, with a mean absolute error (MAE) of 1.85 and a Pearson correlation coefficient (PCC) of 0.63. Additionally, our model was validated on a prospectively collected dataset of 931 patients from National Taiwan University Hospital, yielding an MAE of 1.48 and a PCC of 0.22. In summary, we developed and validated a novel deep learning-based, instance-appraisable model for pain intensity estimation using facial videos. The EDi Pain model shows promise for real-time application in clinical settings, offering a more objective and dynamic approach to pain assessment.

PMID:40355693 | DOI:10.1007/s10278-025-01534-2

J Imaging Inform Med. 2025 May 12. doi: 10.1007/s10278-025-01529-z. Online ahead of print.

ABSTRACT

The purpose of this study was to investigate the utility of deep learning image reconstruction at medium and high intensity levels (DLIR-M and DLIR-H, respectively) for better delineation of liver metastases in pre-contrast and post-contrast CT, compared to conventional hybrid iterative reconstruction (IR) methods. Forty-one patients with liver metastases who underwent abdominal CT were studied. The raw data were reconstructed with three different algorithms: hybrid IR (ASiR-V 50%), DLIR-M (TrueFildelity-M), and DLIR-H (TrueFildelity-H). Three experienced radiologists independently rated the lesion conspicuity of liver metastases on a qualitative 5-point scale (score 1 = very poor; score 5 = excellent). The observers also selected each image series for pre- and post-contrast CT per patient that was considered most preferable for liver metastases assessment. For pre-contrast CT, lesion conspicuity scores for DLIR-H and DLIR-M were significantly higher than those for hybrid IR for two of the three observers, while there was no significant difference for one observer. For post-contrast CT, the lesion conspicuity scores for DLIR-H images were significantly higher than those for DLIR-M images for two of the three observers on post-contrast CT (Observer 1: DLIR-H, 4.3 ± 0.8 vs. DLIR-M, 3.9 ± 0.9, p = 0.0006; Observer 3: DLIR-H, 4.6 ± 0.6 vs. DLIR-M, 4.3 ± 0.6, p = 0.0013). For post-contrast CT, all observers most often selected DLIR-H as the best reconstruction method for the diagnosis of liver metastases. However, in the pre-contrast CT, there was variation among the three observers in determining the most preferred image reconstruction method, and DLIR was not necessarily preferred over hybrid IR for the diagnosis of liver metastases.

PMID:40355690 | DOI:10.1007/s10278-025-01529-z

J Imaging Inform Med. 2025 May 12. doi: 10.1007/s10278-025-01480-z. Online ahead of print.

ABSTRACT

Alveolar echinococcosis (AE) is a parasitic disease caused by Echinococcus multilocularis, where early detection is crucial for effective treatment. This study introduces a novel method for the early diagnosis of liver diseases by differentiating between tumor, AE, and healthy cases using non-contrast CT images, which are widely accessible and eliminate the risks associated with contrast agents. The proposed approach integrates an automatic liver region detection method based on RCNN followed by a CNN-based classification framework. A dataset comprising over 27,000 thorax-abdominal images from 233 patients, including 8206 images with liver tissue, was constructed and used to evaluate the proposed method. The experimental results demonstrate the importance of the two-stage classification approach. In a 2-class classification problem for healthy and non-healthy classes, an accuracy rate of 0.936 (95% CI: 0.925 - 0.947) was obtained, and that for 3-class classification problem with AE, tumor, and healthy classes was obtained as 0.863 (95% CI: 0.847 - 0.879). These results highlight the potential use of the proposed framework as a fully automatic approach for liver classification without the use of contrast agents. Furthermore, the proposed framework demonstrates competitive performance compared to other state-of-the-art techniques, suggesting its applicability in clinical practice.

PMID:40355689 | DOI:10.1007/s10278-025-01480-z

Int Dent J. 2025 May 11;75(4):100827. doi: 10.1016/j.identj.2025.04.007. Online ahead of print.

ABSTRACT

Dental caries is one of the most prevalent chronic diseases among both children and adults, despite being largely preventable. This condition has significant negative impacts on human health and imposes a substantial economic burden. In recent years, scientists and dentists have increasingly started to utilize artificial intelligence (AI), particularly machine learning, to improve the efficiency of dental caries management. This study aims to provide an overview of the current knowledge about the AI-enabled approaches for dental caries management within the framework of personalized patient care. Generally, AI works as a promising tool that can be used by both dental professionals and patients. For dental professionals, it predicts the risk of dental caries by analyzing dental caries risk and protective factors, enabling to formulate personalized preventive measures. AI, especially those based on machine learning and deep learning, can also analyze images to detect signs of dental caries, assist in developing treatment plans, and help to make a risk assessment for pulp exposure during treatment. AI-powered tools can also be used to train dental students through simulations and virtual case studies, allowing them to practice and refine their clinical skills in a risk-free environment. Additionally, AI tracks brushing patterns and provides feedback to improve oral hygiene practices of the patients and the general population, thereby improving their understanding and compliance. This capability of AI can inform future research and the development of new strategies for dental caries management and control.

PMID:40354695 | DOI:10.1016/j.identj.2025.04.007

Bioinformatics. 2025 May 12:btaf298. doi: 10.1093/bioinformatics/btaf298. Online ahead of print.

ABSTRACT

MOTIVATION: Predicting protein-ligand binding affinity accurately and quickly is a major challenge in drug discovery. Recent advancements suggest that deep learning-based computational methods can effectively quantify binding affinity, making them a promising alternative. Environmental factors significantly influence the interactions between protein pockets and ligands, affecting the binding strength. However, many existing deep learning approaches tend to overlook these environmental effects, focusing instead on extracting features from proteins and ligands based solely on their sequences or structures.

RESULTS: We propose a deep learning method, EM-PLA, which is based on an environment-aware heterogeneous graph neural network and utilizes multimodal data. This method improves protein-ligand binding affinity prediction by incorporating environmental information derived from the biochemical properties of proteins and ligands. Specifically, EM-PLA employs a heterogeneous graph neural network(HGT) with environmental information to improve the calculation of non-covalent interactions, while also considering the interaction calculations between protein sequences and ligand sequences. We evaluate the performance of the proposed EM-PLA through comprehensive benchmark experiments for binding affinity prediction, demonstrating its superior performance and generalization capability compared to state-of-the-art baseline methods. Furthermore, by analyzing the results of the ablation experiments and integrating visual analyses and case studies, we validate the rationale of the proposed method. These results indicate that EM-PLA is an effective method for binding affinity prediction and may provide valuable insights for future applications.

AVAILABILITY AND IMPLEMENTATION: The source code is available at https://github.com/littlemou22/EM-PLA.

CONTACT: pzhang@tju.edu.com.

SUPPLEMENTARY INFORMATION: Supplementary data are available in the submitted files.

PMID:40354612 | DOI:10.1093/bioinformatics/btaf298

Anal Chem. 2025 May 12. doi: 10.1021/acs.analchem.5c01082. Online ahead of print.

ABSTRACT

Precise discrimination of the crucial substances, e.g., synthetic cannabinoids (SCs) that are composed of low-active chemical groups and structurally similar to each other with tiny differences, is a pressing need and of great significance for safeguarding public security and human health. The structure-relevant vibrational spectroscopic techniques, e.g., Raman spectroscopy, could reflect structural fingerprint information on the target; however, the algorithm-assisted phrasing is inevitable. This work achieved the accurate identification of CA series SCs by proposing an attention mechanism involving a CNN algorithm to phrase the Raman data. Specifically, these SCs have only one different chemical group compared to each other, the attention mechanism was introduced to intensify the computation on their structural difference from the massive data, realizing the accurate discrimination. Furthermore, how the spectral peaks corresponded to the specific structure was revealed, which plays a decisive role for the algorithm to distinguish these substances, and provides an instructive reference for differentiating other SCs based on Raman spectra. Hence, this work provides a research paradigm for applying the advanced CNN algorithm-aided Raman spectral analysis to sub-differentiate the substances, strengthening the understanding of spectral information from the sub-molecular level and propelling the integration of interdisciplinary areas.

PMID:40354573 | DOI:10.1021/acs.analchem.5c01082

PLoS One. 2025 May 12;20(5):e0319562. doi: 10.1371/journal.pone.0319562. eCollection 2025.

ABSTRACT

The increasing importance of deep learning in software development has greatly improved software quality by enabling the efficient identification of defects, a persistent challenge throughout the software development lifecycle. This study seeks to determine the most effective model for detecting defects in software projects. It introduces an intelligent approach that combines Temporal Convolutional Networks (TCN) with Antlion Optimization (ALO). TCN is employed for defect detection, while ALO optimizes the network's weights. Two models are proposed to address the research problem: (a) a basic TCN without parameter optimization and (b) a hybrid model integrating TCN with ALO. The findings demonstrate that the hybrid model significantly outperforms the basic TCN in multiple performance metrics, including area under the curve, sensitivity, specificity, accuracy, and error rate. Moreover, the hybrid model surpasses state-of-the-art methods, such as Convolutional Neural Networks, Gated Recurrent Units, and Bidirectional Long Short-Term Memory, with accuracy improvements of 21.8%, 19.6%, and 31.3%, respectively. Additionally, the proposed model achieves a 13.6% higher area under the curve across all datasets compared to the Deep Forest method. These results confirm the effectiveness of the proposed hybrid model in accurately detecting defects across diverse software projects.

PMID:40354496 | DOI:10.1371/journal.pone.0319562

PLoS One. 2025 May 12;20(5):e0317758. doi: 10.1371/journal.pone.0317758. eCollection 2025.

ABSTRACT

Eliminating flickering from digital images captured by cameras equipped with a rolling shutter is of paramount importance in computer vision applications. The ripple effect observed in an individual image is a consequence of the non-synchronized exposure of rolling shutters utilized in CMOS sensor-based cameras. To date, there have been only a limited number of studies focusing on the mitigation of flickering in single images. Furthermore, it is more feasible to eliminate these flickers with prior knowledge, such as camera specifications or matching images. To solve these problems, we present an unsupervised framework Super-Resolution Generative Adversarial Networks and Partition-Based Adaptive Filtering Technique (SRGAN-PBAFT) trained on unpaired images from end to end Deflickering of a single image. Flicker artifacts, which are commonly caused by dynamic lighting circumstances and sensor noise, can severely reduce an image's visual quality and authenticity. To enhance image resolution SRGAN is used, while Partition based Adaptive Filtering technique detects and mitigates flicker distortions successfully. Combining the strengths of deep learning and adaptive filtering results in a potent approach for restoring image integrity. Experimental results shows that the Proposed SRGAN-PBAFT method is effective, with major improvements in visual quality and flicker aberration reduction compared to existing methods.

PMID:40354494 | DOI:10.1371/journal.pone.0317758

PLOS Digit Health. 2025 May 12;4(5):e0000831. doi: 10.1371/journal.pdig.0000831. eCollection 2025 May.

ABSTRACT

Diabetic retinopathy (DR) is a frequent complication of diabetes, affecting millions worldwide. Screening for this disease based on fundus images has been one of the first successful use cases for modern artificial intelligence in medicine. However, current state-of-the-art systems typically use black-box models to make referral decisions, requiring post-hoc methods for AI-human interaction and clinical decision support. We developed and evaluated an inherently interpretable deep learning model, which explicitly models the local evidence of DR as part of its network architecture, for clinical decision support in early DR screening. We trained the network on 34,350 high-quality fundus images from a publicly available dataset and validated its performance on a large range of ten external datasets. The inherently interpretable model was compared to post-hoc explainability techniques applied to a standard DNN architecture. For comparison, we obtained detailed lesion annotations from ophthalmologists on 65 images to study if the class evidence maps highlight clinically relevant information. We tested the clinical usefulness of our model in a retrospective reader study, where we compared screening for DR without AI support to screening with AI support with and without AI explanations. The inherently interpretable deep learning model obtained an accuracy of .906 [.900-.913] (95%-confidence interval) and an AUC of .904 [.894-.913] on the internal test set and similar performance on external datasets, comparable to the standard DNN. High evidence regions directly extracted from the model contained clinically relevant lesions such as microaneurysms or hemorrhages with a high precision of .960 [.941-.976], surpassing post-hoc techniques applied to a standard DNN. Decision support by the model highlighting high-evidence regions in the image improved screening accuracy for difficult decisions and improved screening speed. This shows that inherently interpretable deep learning models can provide clinical decision support while obtaining state-of-the-art performance improving human-AI collaboration.

PMID:40354306 | DOI:10.1371/journal.pdig.0000831

J Exp Psychol Gen. 2025 May 12. doi: 10.1037/xge0001740. Online ahead of print.

ABSTRACT

Pertaining to goal orientation and achievement, agency is a fundamental aspect of human cognition and behavior. Accordingly, detecting and quantifying linguistic encoding of agency are critical for the analysis of human actions, interactions, and social dynamics. Available agency-quantifying computational tools rely on word-counting methods, which typically are insensitive to the semantic context in which the words are used and consequently prone to miscoding, for example, in case of polysemy. Additionally, some currently available tools do not take into account differences in the intensity and directionality of agency. In order to overcome these shortcomings, we present BERTAgent, a novel tool to quantify semantic agency in text. BERTAgent is a computational language model that utilizes the transformers architecture, a popular deep learning approach to natural language processing. BERTAgent was fine-tuned using textual data that were evaluated by human coders with respect to the level of conveyed agency. In four validation studies, BERTAgent exhibits improved convergent and discriminant validity compared to previous solutions. Additionally, the detailed description of BERTAgent's development procedure serves as a tutorial for the advancement of similar tools, providing a blueprint for leveraging the existing lexicographical data sets in conjunction with the deep learning techniques in order to detect and quantify other psychological constructs in textual data. (PsycInfo Database Record (c) 2025 APA, all rights reserved).

PMID:40354292 | DOI:10.1037/xge0001740

IEEE Trans Image Process. 2025 May 12;PP. doi: 10.1109/TIP.2025.3567205. Online ahead of print.

ABSTRACT

All-in-one image restoration, which seeks to handle multiple types of degradation within a unified model, has become a prominent research topic in computer vision. While existing deep learning models have achieved remarkable success in specific restoration tasks, extending these models to heterogenous degradations presents significant challenges. Current all-in-one methods predominantly concentrate on extracting degradation priors, often employing learned and fixed task prompts to guide the restoration process. However, these static prompts are inclined to generate an average distribution characteristics of degradations, unable to accurately depict the unique attribute of the given input, consequently providing suboptimal restoration results. To tackle these challenges, we propose a novel dynamic prompt approach called Degradation Prototype Assignment and Prompt Distribution Learning (DPPD). Our approach decouples the degradation prior extraction into two novel components: Degradation Prototype Assignment (DPA) and Prompt Distribution Learning (PDL). DPA anchors the degradation representations to predefined prototypes, providing discriminative and scalable representations. In addition, PDL models prompts as distributions rather than fixed parameters, facilitating dynamic and adaptive prompt sampling. Extensive experiments demonstrate that our DPPD framework can achieve significant performance improvement on different image restoration tasks. Codes are available at our project page https://github.com/Aitical/DPPD.

PMID:40354220 | DOI:10.1109/TIP.2025.3567205

IEEE J Biomed Health Inform. 2025 May 12;PP. doi: 10.1109/JBHI.2025.3569160. Online ahead of print.

ABSTRACT

Respiratory diseases present significant challenges to global health due to their high morbidity and mortality rates. Traditional diagnostic methods, such as chest radiographs and blood tests, often lead to unnecessary costs and resource strain, as well as potential risks of cross-contamination during these procedures. In recent years, contactless sensing and intelligent technologies, particularly multi-modal sound-based deep learning methods, have emerged as promising solutions for the early detection of respiratory diseases. While these methods have shown encouraging results, the integration of multi-modal features has not been sufficiently explored, which limits the enhancement of diagnostic accuracy. To address this issue, we introduce GAMMNet, a novel multi-modal neural network designed to enhance the detection of respiratory diseases by leveraging multi-modal sound data collected from contactless recording devices. GAMMNet utilizes a unique gating mechanism that adaptively regulates the influence of each modality on the classification results. Additionally, our model incorporates multi-head attention and linear transformation modules to further enhance classification performance. Our GAMMNet achieves state-of-the-art classification results, compared to existing deep learning based methods, on real-world multi-modal respiratory sound datasets. These findings demonstrate the robustness and effectiveness of GAMMNet in the contactless monitoring and early detection of respiratory diseases.

PMID:40354202 | DOI:10.1109/JBHI.2025.3569160