Diabetes Res Clin Pract. 2025 May 4:112221. doi: 10.1016/j.diabres.2025.112221. Online ahead of print.

ABSTRACT

Diabetes mellitus (DM) is a highly prevalent chronic condition with significant health and economic impacts; therefore, an accurate diagnosis is essential for the effective management and prevention of its complications. This review explores the latest advances in artificial intelligence (AI) focusing on machine learning (ML) and deep learning (DL) for the diagnosis of diabetes. Recent developments in AI-driven diagnostic tools were analyzed, with an emphasis on breakthrough methodologies and their real-world clinical applications. This review also discusses the role of various data sources, datasets, and preprocessing techniques in enhancing diagnostic accuracy. Key advancements in integrating AI into clinical workflows and improving early detection are highlighted along with challenges related to model interpretability, ethical considerations, and practical implementation. By offering a comprehensive overview of these advancements and their implications, this review contributes significantly to the understanding of how AI technologies can enhance the diagnosis of diabetes and support their integration into clinical practice, thereby aiming to improve patient outcomes and reduce the burden of diabetes.

PMID:40328407 | DOI:10.1016/j.diabres.2025.112221

Radiother Oncol. 2025 May 4:110914. doi: 10.1016/j.radonc.2025.110914. Online ahead of print.

ABSTRACT

PURPOSES: This study aimed to develop a computed tomography (CT)-based multi-organ segmentation model for delineating organs-at-risk (OARs) in pediatric upper abdominal tumors and evaluate its robustness across multiple datasets.

MATERIALS AND METHODS: In-house postoperative CTs from pediatric patients with renal tumors and neuroblastoma (n = 189) and a public dataset (n = 189) with CTs covering thoracoabdominal regions were used. Seventeen OARs were delineated: nine by clinicians (Type 1) and eight using TotalSegmentator (Type 2). Auto-segmentation models were trained using in-house (Model-PMC-UMCU) and a combined dataset of public data (Model-Combined). Performance was assessed with Dice Similarity Coefficient (DSC), 95 % Hausdorff Distance (HD95), and mean surface distance (MSD). Two clinicians rated clinical acceptability on a 5-point Likert scale across 15 patient contours. Model robustness was evaluated against sex, age, intravenous contrast, and tumor type.

RESULTS: Model-PMC-UMCU achieved mean DSC values above 0.95 for five of nine OARs, while the spleen and heart ranged between 0.90 and 0.95. The stomach-bowel and pancreas exhibited DSC values below 0.90. Model-Combined demonstrated improved robustness across both datasets. Clinical evaluation revealed good usability, with both clinicians rating six of nine Type 1 OARs above four and six of eight Type 2 OARs above three. Significant performance differences were only found across age groups in both datasets, specifically in the left lung and pancreas. The 0-2 age group showed the lowest performance.

CONCLUSION: A multi-organ segmentation model was developed, showcasing enhanced robustness when trained on combined datasets. This model is suitable for various OARs and can be applied to multiple datasets in clinical settings.

PMID:40328363 | DOI:10.1016/j.radonc.2025.110914

Proc Biol Sci. 2025 May;292(2046):20242233. doi: 10.1098/rspb.2024.2233. Epub 2025 May 7.

ABSTRACT

Traditional biomarkers, such as those obtained from blood tests, are essential for early disease detection, improving health outcomes and reducing healthcare costs. However, they often involve invasive procedures, specialized laboratory equipment or special handling of biospecimens. The retinal age gap (RAG) has emerged as a promising new biomarker that can overcome these limitations, making it particularly suitable for disease screening in low- and middle-income countries. This study aimed to evaluate the potential of the RAG as a biomarker for broad disease screening across a vast spectrum of diseases. Fundus images were collected from 86 522 UK Biobank participants aged 40-83 (mean age: 56.2 ± 8.3 years). A deep learning model was trained to predict retinal age using 17 791 images from healthy participants. The remaining images were categorized into disease/injury groups based on clinical codes. Additionally, 8524 participants from the Brazilian Multilabel Ophthalmological Dataset (BRSET) were used for external validation. Among the 159 disease/injury groups from the 2019 Global Burden of Disease Study, 56 groups (35.2%) exhibited RAG distributions significantly different from healthy controls. Notable examples included chronic kidney disease, cardiovascular disease, blindness, vision loss and diabetes. Overall, the RAG shows great promise as a cost-effective, non-invasive biomarker for early disease screening.

PMID:40328303 | DOI:10.1098/rspb.2024.2233

Biomed Eng Online. 2025 May 6;24(1):53. doi: 10.1186/s12938-025-01388-3.

ABSTRACT

BACKGROUND: Accurate measurement of anterior segment parameters is crucial for diagnosing and managing ophthalmic conditions, such as glaucoma, cataracts, and refractive errors. However, traditional clinical measurement methods are often time-consuming, labor-intensive, and susceptible to inaccuracies. With the growing potential of artificial intelligence in ophthalmic diagnostics, this study aims to develop and evaluate a deep learning model capable of automatically extracting key points and precisely measuring multiple clinically significant anterior segment parameters from ultrasound biomicroscopy (UBM) images. These parameters include central corneal thickness (CCT), anterior chamber depth (ACD), pupil diameter (PD), angle-to-angle distance (ATA), sulcus-to-sulcus distance (STS), lens thickness (LT), and crystalline lens rise (CLR).

METHODS: A data set of 716 UBM anterior segment images was collected from Tianjin Medical University Eye Hospital. YOLOv8 was utilized to segment four key anatomical structures: cornea-sclera, anterior chamber, pupil, and iris-ciliary body-thereby enhancing the accuracy of keypoint localization. Only images with intact posterior capsule lentis were selected to create an effective data set for parameter measurement. Ten keypoints were localized across the data set, allowing the calculation of seven essential parameters. Control experiments were conducted to evaluate the impact of segmentation on measurement accuracy, with model predictions compared against clinical gold standards.

RESULTS: The segmentation model achieved a mean IoU of 0.8836 and mPA of 0.9795. Following segmentation, the binary classification model attained an mAP of 0.9719, with a precision of 0.9260 and a recall of 0.9615. Keypoint localization exhibited a Euclidean distance error of 58.73 ± 63.04 μm, improving from the pre-segmentation error of 71.57 ± 67.36 μm. Localization mAP was 0.9826, with a precision of 0.9699, a recall of 0.9642 and an FPS of 32.64. In addition, parameter error analysis and Bland-Altman plots demonstrated improved agreement with clinical gold standards after segmentation.

CONCLUSIONS: This deep learning approach for UBM image segmentation, keypoint localization, and parameter measurement is feasible, enhancing clinical diagnostic efficiency for anterior segment parameters.

PMID:40329288 | DOI:10.1186/s12938-025-01388-3

BMC Cancer. 2025 May 6;25(1):830. doi: 10.1186/s12885-025-14113-z.

ABSTRACT

Non-coding RNAs (ncRNAs) play a crucial role in breast cancer progression, necessitating advanced computational approaches for precise disease classification. This study introduces a Deep Reinforcement Learning (DRL)-based framework for predicting ncRNA-disease associations in metaplastic breast cancer (MBC) using a multi-dimensional descriptor system (ncRNADS) integrating 550 sequence-based features and 1,150 target gene descriptors (miRDB score ≥ 90). The model achieved 96.20% accuracy, 96.48% precision, 96.10% recall, and a 96.29% F1-score, outperforming traditional classifiers such as support vector machines (SVM) and neural networks. Feature selection and optimization reduced dimensionality by 42.5% (4,430 to 2,545 features) while maintaining high accuracy, demonstrating computational efficiency. External validation confirmed model specificity to breast cancer subtypes (87-96.5% accuracy) and minimal cross-reactivity with unrelated diseases like Alzheimer's (8-9% accuracy), ensuring robustness. SHAP analysis identified key sequence motifs (e.g., "UUG") and structural free energy (ΔG = - 12.3 kcal/mol) as critical predictors, validated by PCA (82% variance) and t-SNE clustering. Survival analysis using TCGA data revealed prognostic significance for MALAT1, HOTAIR, and NEAT1 (associated with poor survival, HR = 1.76-2.71) and GAS5 (protective effect, HR = 0.60). The DRL model demonstrated rapid training (0.08 s/epoch) and cloud deployment compatibility, underscoring its scalability for large-scale applications. These findings establish ncRNA-driven classification as a cornerstone for precision oncology, enabling patient stratification, survival prediction, and therapeutic target identification in MBC.

PMID:40329245 | DOI:10.1186/s12885-025-14113-z

J Imaging Inform Med. 2025 May 6. doi: 10.1007/s10278-025-01528-0. Online ahead of print.

ABSTRACT

Quantitative cardiovascular PET/CT imaging is useful in the diagnosis of multiple cardiac perfusion and motion pathologies. The common approach for cardiac segmentation consists in using co-registered CT images, exploiting publicly available deep learning (DL)-based segmentation models. However, the mismatch between structural CT images and PET uptake limits the usefulness of these approaches. Besides, the performance of DL models is not consistent over low-dose or ultra-low-dose CT images commonly used in clinical PET/CT imaging. In this work, we developed a DL-based methodology to tackle this issue by segmenting directly cardiac PET images. This study included 406 cardiac PET images from 146 patients (43 18F-FDG, 329 13N-NH3, and 37 82Rb images). Using previously trained DL nnU-Net models in our group, we segmented the whole heart and the three main cardiac components, namely the left myocardium (LM), left ventricle cavity (LV), and right ventricle (RV) on co-registered CT images. The segmentation was resampled to PET resolution and edited through a combination of automated image processing and manual correction. The corrected segmentation masks and SUV PET images were fed to a nnU-Net V2 pipeline to be trained in fivefold data split strategy by defining two tasks: task #1 for whole cardiac segmentation and task #2 for segmentation of three cardiac components. Fifteen cardiac images were used as external validation set. The DL delineated masks were compared with standard of reference masks using Dice coefficient, Jaccard distance, mean surface distance, and segment volume relative error (%). Task #1 average Dice coefficient in internal validation fivefold was 0.932 ± 0.033. The average Dice on the 15 external cases were comparable with the fivefold Dice reaching an average of 0.941 ± 0.018. Task #2 average Dice in fivefold validation was 0.88 ± 0.063, 0.828 ± 0.091, and 0.876 ± 0.062 for LM, LV, and RV, respectively. There was no statistically significant difference among the Dice coefficients, neither between images acquired by three radiotracers nor between the different folds (P-values > > 0.05). The overall average volume prediction error in cardiac components segmentation was less than 2%. We developed an automated DL-based segmentation pipeline to segment the whole heart and cardiac components with acceptable accuracy and robust performance in the external test set and over three radiotracers used in nuclear cardiovascular imaging. The proposed methodology can overcome unreliable segmentations performed on CT images.

PMID:40329157 | DOI:10.1007/s10278-025-01528-0

J Imaging Inform Med. 2025 May 6. doi: 10.1007/s10278-025-01527-1. Online ahead of print.

ABSTRACT

Deep learning techniques have demonstrated potential in various fields, including segmentation, and have recently been applied to medical image processing. This study aims to develop and evaluate computer-based diagnostic software designed to assess the segmentation of the mandibular condyle in ultrasound images. A total of 668 retrospective ultrasound images of anonymous adult mandibular condyles were analyzed. The CranioCatch labeling program (CranioCatch, Eskişehir, Turkey) was utilized to annotate the mandibular condyle using a polygonal labeling method. These annotations were subsequently reviewed and validated by experts in oral and maxillofacial radiology. In this study, all test images were detected and segmented using the YOLOv8 deep learning artificial intelligence (AI) model. When evaluating the model's performance in image estimation, it achieved an F1 score of 0.93, a sensitivity of 0.90, and a precision of 0.96. The automatic segmentation of the mandibular condyle from ultrasound images presents a promising application of artificial intelligence. This approach can help surgeons, radiologists, and other specialists save time in the diagnostic process.

PMID:40329156 | DOI:10.1007/s10278-025-01527-1

J Imaging Inform Med. 2025 May 6. doi: 10.1007/s10278-025-01525-3. Online ahead of print.

ABSTRACT

The purpose of this study is to assess the ability of deep learning models to classify different subtypes of solid renal parenchymal tumors using contrast-enhanced ultrasound (CEUS) images and to compare their classification performance. A retrospective study was conducted using CEUS images of 237 kidney tumors, including 46 angiomyolipomas (AML), 118 clear cell renal cell carcinomas (ccRCC), 48 papillary RCCs (pRCC), and 25 chromophobe RCCs (chRCC), collected from January 2017 to December 2019. Two deep learning models, based on the ResNet-18 and RepVGG architectures, were trained and validated to distinguish between these subtypes. The models' performance was assessed using sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Matthews correlation coefficient, accuracy, area under the receiver operating characteristic curve (AUC), and confusion matrix analysis. Class activation mapping (CAM) was applied to visualize the specific regions that contributed to the models' predictions. The ResNet-18 and RepVGG-A0 models achieved an overall accuracy of 76.7% and 84.5% across all four subtypes. The AUCs for AML, ccRCC, pRCC, and chRCC were 0.832, 0.829, 0.806, and 0.795 for the ResNet-18 model, compared to 0.906, 0.911, 0.840, and 0.827 for the RepVGG-A0 model, respectively. The deep learning models could reliably differentiate between various histological subtypes of renal tumors using CEUS images in an objective and non-invasive manner.

PMID:40329155 | DOI:10.1007/s10278-025-01525-3

J Imaging Inform Med. 2025 May 6. doi: 10.1007/s10278-025-01465-y. Online ahead of print.

ABSTRACT

Breast cancer has remained one of the most frequent and life-threatening cancers in females globally, putting emphasis on better diagnostics in its early stages to solve the problem of therapy effectiveness and survival. This work enhances the assessment of breast cancer by employing progressive residual networks (PRN) and ResNet-50 within the framework of Progressive Residual Multi-Class Support Vector Machine-Net. Built on concepts of deep learning, this creative integration optimizes feature extraction and raises the bar for classification effectiveness, earning an almost perfect 99.63% on our tests. These findings indicate that PRMS-Net can serve as an efficient and reliable diagnostic tool for early breast cancer detection, aiding radiologists in improving diagnostic accuracy and reducing false positives. The separation of the data into different segments is possible to determine the architecture's reliability using the fivefold cross-validation approach. The total variability of precision, recall, and F1 scores clearly depicted in the box plot also endorse the competency of the model for marking proper sensitivity and specificity-highly required for combating false positive and false negative cases in real clinical practice. The evaluation of error distribution strengthens the model's rationale by giving validation of practical application in medical contexts of image processing. The high levels of feature extraction sensitivity together with highly sophisticated classification methods make PRMS-Net a powerful tool that can be used in improving the early detection of breast cancer and subsequent patient prognosis.

PMID:40329154 | DOI:10.1007/s10278-025-01465-y

Sci Rep. 2025 May 6;15(1):15773. doi: 10.1038/s41598-025-99694-2.

ABSTRACT

Passive localization is necessary for Internet of Things (IoT) applications to observe and follow people without requiring them to carry massive equipment. This is crucial in private settings like security and medical monitoring, where individuals are reluctant to wear tracking equipment. Localizing and tracking objects in these spaces are vital since wall loss causes GPS signals to perform poorly in indoor environments. Therefore, passive localization using Radio Tomography Images (RTI) has gained significant importance in present life. Because there are flaws in the RSSI data that models might exploit, previous problems with RTI sparked innovation and resulted in the development of more complex systems, such as a passive localization system that leverages deep learning. This paper employs a set of ESP32 nodes for a mesh network and utilizes a radio frequency sensor network with ESP32 modules to collect RSSI values. We have developed and thoroughly examined the working of radio tomography generation algorithms and present a deep learning approach using a convolutional neural network (CNN) to address the inverse problem. Two CNN models are developed to reconstruct static tomographic images, improve the quality of these images, and localize targeted objects. The targeted object localization accuracy is above 92% by using the proposed system. The results of the proposed system are also compared with previously developed approaches, and it is clearly shown that the proposed system outperforms the previously developed approaches.

PMID:40328896 | DOI:10.1038/s41598-025-99694-2

Sci Rep. 2025 May 6;15(1):15820. doi: 10.1038/s41598-025-96577-4.

ABSTRACT

Accurate high-resolution runoff predictions are essential for effective flood mitigation and water planning. In hydrology, conceptual models are preferred for their simplicity, despite their limited capacity for accurate predictions. Deep-learning applications have recently shown promise for runoff predictions; however, they usually require longer input data sequences, especially for high-temporal resolution simulations, thus leading to increased model complexity. To address these challenges, this study evaluates the robustness of two novel approaches using Long Short-Term Memory (LSTM) models. The first model integrates the outputs of a simple conceptual model with LSTM capabilities, while the second model is a stand-alone model that combines coarse and fine temporal inputs to capture both long and short dependencies. To ensure accuracy and reliability, we utilized a century-long meteorological dataset generated from a sophisticated physics-based model, eliminating any influence of measurement errors. The training phase employed multiple sub-periods ranging from 7- to 50-year, with a separate 50-year subset for validation. Our findings highlight the consistent improvement of both LSTM models with increasing training dataset lengths, while conceptual models show no notable enhancement beyond 15 years of training data. Both LSTM models demonstrate superior performance in capturing the reference flow duration curve, offering a promising pathway for more computationally efficient models for runoff predictions.

PMID:40328848 | DOI:10.1038/s41598-025-96577-4

Sci Rep. 2025 May 6;15(1):15845. doi: 10.1038/s41598-025-00592-4.

ABSTRACT

This study focuses on the application of virtual reality (VR) technology in English language teaching (ELT), and discusses the effect of combining VR with machine translation (MT) technology. VR technology is introduced into the study to provide innovative teaching methods for English teachers and create an immersive learning environment for students. Based on the latest development of deep learning (DL), a new MT model is proposed in this study, and it is successfully integrated with VR technology to optimize the quality of ELT. The experimental results show that the translation accuracy of the MT model designed in this study reaches 98.5%, the F1 score is stable at around 93%, and the semantic information recall rate is as high as 92%, all of which are better than the traditional model. In the preliminary test, the comparative experiment of 40 English majors further verified the effectiveness of the model in improving translation efficiency and quality. This study shows the great potential of the integration of VR and MT technology in ELT, and also proves its advantages by using experimental data, which provides technical support for ELT and provides reference for future practice.

PMID:40328816 | DOI:10.1038/s41598-025-00592-4

Nat Commun. 2025 May 6;16(1):4212. doi: 10.1038/s41467-025-59516-5.

ABSTRACT

Efficient and accurate road distress detection is crucial for infrastructure maintenance and transportation safety. Traditional manual inspections are labor-intensive and time-consuming, while increasingly popular automated systems often rely on computationally intensive devices, limiting widespread adoption. To address these challenges, this study introduces MobiLiteNet, a lightweight deep learning approach designed for mobile deployment on smartphones and mixed reality systems. Utilizing a diverse dataset collected from Europe and Asia, MobiLiteNet incorporates Efficient Channel Attention to boost model performance, followed by structural refinement, sparse knowledge distillation, structured pruning, and quantization to significantly increase the computational efficiency while preserving high detection accuracy. To validate its effectiveness, MobiLiteNet improves the existing MobileNet model. Test results show that the improved MobileNet outperforms baseline models on mobile devices. With significantly reduced computational costs, this approach enables real-time, scalable, and accurate road distress detection, contributing to more efficient road infrastructure management and intelligent transportation systems.

PMID:40328808 | DOI:10.1038/s41467-025-59516-5

J Med Internet Res. 2025 May 6;27:e65272. doi: 10.2196/65272.

ABSTRACT

BACKGROUND: Worldwide, 30%-45% of adults have sleep disorders, which are linked to major health issues such as diabetes and cardiovascular disease. Long-term monitoring with traditional in-lab testing is impractical due to high costs. Wearable artificial intelligence (AI)-powered solutions offer accessible, scalable, and continuous monitoring, improving the identification and treatment of sleep problems.

OBJECTIVE: This scoping review aims to provide an overview of AI-powered wearable devices used for sleep disorders, focusing on study characteristics, wearable technology features, and AI methodologies for detection and analysis.

METHODS: Seven electronic databases (MEDLINE, PsycINFO, Embase, IEEE Xplore, ACM Digital Library, Google Scholar, and Scopus) were searched for peer-reviewed literature published before March 2024. Keywords were selected based on 3 domains: sleep disorders, AI, and wearable devices. The primary selection criterion was the inclusion of studies that utilized AI algorithms to detect or predict various sleep disorders using data from wearable devices. Study selection was conducted in 2 steps: first, by reviewing titles and abstracts, followed by full-text screening. Two reviewers independently conducted study selection and data extraction, resolving discrepancies by consensus. The extracted data were synthesized using a narrative approach.

RESULTS: The initial search yielded 615 articles, of which 46 met the eligibility criteria and were included in the final analysis. The majority of studies focused on sleep apnea. Wearable AI was widely deployed for diagnosing and screening disorders; however, none of the studies used it for treatment. Commercial devices were the most commonly used type of wearable technology, appearing in 30 out of 46 (65%) studies. Among these, various brands were utilized rather than a single large, well-known brand; 19 (41%) studies used wrist-worn devices. Respiratory data were used by 25 of 46 (54%) studies as the primary data for model development, followed by heart rate (22/46, 48%) and body movement (17/46, 37%). The most popular algorithm was the convolutional neural network, adopted by 17 of 46 (37%) studies, followed by random forest (14/46, 30%) and support vector machines (12/46, 26%).

CONCLUSIONS: Wearable AI technology offers promising solutions for sleep disorders. These devices can be used for screening and diagnosis; however, research on wearable technology for sleep disorders other than sleep apnea remains limited. To statistically synthesize performance and efficacy results, more reviews are needed. Technology companies should prioritize advancements such as deep learning algorithms and invest in wearable AI for treating sleep disorders, given its potential. Further research is necessary to validate machine learning techniques using clinical data from wearable devices and to develop useful analytics for data collection, monitoring, prediction, classification, and recommendation in the context of sleep disorders.

PMID:40327852 | DOI:10.2196/65272

ACS Appl Mater Interfaces. 2025 May 6. doi: 10.1021/acsami.4c23103. Online ahead of print.

ABSTRACT

In this study, we present a hybrid Physics-Assisted Machine Learning (PAML) model that integrates Deep Learning (DL) techniques with the classical Discrete Element Method (DEM) to simulate slurry drying during a lithium-ion battery electrode manufacturing process. This model predicts the microstructure evolution leading to the formation of the electrode as a time-series along the drying process. The hybrid approach consists in performing a certain amount of DEM simulation steps, nDEM, after every DL prediction, mitigating the risk of unphysical predictions, like overlapping particles. Our PAML model was rigorously tested by evaluating different functional metrics of the predicted electrodes, including density, porosity, tortuosity factor, and radial distribution function. We conducted an in-depth analysis of performance versus accuracy, particularly focusing on the impact of the nDEM hyperparameter, which represents the number of DEM steps executed between two subsequent DL predictions. Despite the model being trained on a specific formulation (96% of Active Material, AM, and 4% of Carbon Binder Domain, CBD), it demonstrated exceptional generalization capability when used to extrapolate to a different formulation (94% AM and 6% CBD). This adaptability highlights the robustness of our PAML hybrid approach. Furthermore, the integration of DL significantly reduced the computational cost versus the original DEM model simulation, decreasing the calculation time from 615 to 36 min for the whole slurry drying simulation process. Our findings underscore the potential of combining ML with traditional simulation methods to enhance efficiency and accuracy in the field of electrode manufacturing.

PMID:40327815 | DOI:10.1021/acsami.4c23103

PLoS One. 2025 May 6;20(5):e0321854. doi: 10.1371/journal.pone.0321854. eCollection 2025.

ABSTRACT

with the intensification of market competition and the complexity of consumer behavior, enterprises are faced with the challenge of how to accurately identify potential customers and improve user conversion rate. This paper aims to study the application of machine learning in consumer behavior prediction and precision marketing. Four models, namely support vector machine (SVM), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and backpropagation artificial neural network (BPANN), are mainly used to predict consumers' purchase intention, and the performance of these models in different scenarios is verified through experiments. The results show that CatBoost and XGBoost have the best prediction results when dealing with complex features and large-scale data, F1 scores are 0.93 and 0.92 respectively, and CatBoost's ROC AUC reaches the highest value of 0.985. while SVM has an advantage in accuracy rate, but slightly underperformance when dealing with large-scale data. Through feature importance analysis, we identify the significant impact of page views, residence time and other features on purchasing behavior. Based on the model prediction results, this paper proposes the specific application of optimization marketing strategies such as recommendation system, dynamic pricing and personalized advertising. Future research could improve the predictive power of the model by introducing more kinds of unstructured data, such as consumer reviews, images, videos, and social media data. In addition, the use of deep learning models, such as Transformers or Self-Attention Mechanisms, can better capture complex patterns in long time series data.

PMID:40327711 | DOI:10.1371/journal.pone.0321854

Proc Natl Acad Sci U S A. 2025 May 13;122(19):e2426449122. doi: 10.1073/pnas.2426449122. Epub 2025 May 6.

ABSTRACT

Phase separation is an important mechanism to generate certain biomolecular condensates and organize the cell interior. Condensate formation and function remain incompletely understood due to difficulties in visualizing the condensate interior at high resolution. Here, we analyzed the structure of biochemically reconstituted chromatin condensates through cryoelectron tomography. We found that traditional blotting methods of sample preparation were inadequate, and high-pressure freezing plus focused ion beam milling was essential to maintain condensate integrity. To identify densely packed molecules within the condensate, we integrated deep learning-based segmentation with context-aware template matching. Our approaches were developed on chromatin condensates and were also effective on condensed regions of in situ native chromatin. Using these methods, we determined the average structure of nucleosomes to 6.1 and 12 Å resolution in reconstituted and native systems, respectively, found that nucleosomes form heterogeneous interaction networks in both cases, and gained insight into the molecular origins of surface tension in chromatin condensates. Our methods should be applicable to biomolecular condensates containing large and distinctive components in both biochemical reconstitutions and certain cellular systems.

PMID:40327693 | DOI:10.1073/pnas.2426449122

IEEE Trans Vis Comput Graph. 2025 May 6;PP. doi: 10.1109/TVCG.2025.3567117. Online ahead of print.

ABSTRACT

Urban segregation refers to the physical and social division of people, often driving inequalities within cities and exacerbating socioeconomic and racial tensions. While most studies focus on residential spaces, they often neglect segregation across "activity spaces" where people work, socialize, and engage in leisure. Human mobility data offers new opportunities to analyze broader segregation patterns, encompassing both residential and activity spaces, but challenges existing methods in capturing the complexity and local nuances of urban segregation. This work introduces InclusiViz, a novel visual analytics system for multi-level analysis of urban segregation, facilitating the development of targeted, data-driven interventions. Specifically, we developed a deep learning model to predict mobility patterns across social groups using environmental features, augmented with explainable AI to reveal how these features influence segregation. The system integrates innovative visualizations that allow users to explore segregation patterns from broad overviews to fine-grained detail and evaluate urban planning interventions with real-time feedback. We conducted a quantitative evaluation to validate the model's accuracy and efficiency. Two case studies and expert interviews with social scientists and urban analysts demonstrated the system's effectiveness, highlighting its potential to guide urban planning toward more inclusive cities.

PMID:40327496 | DOI:10.1109/TVCG.2025.3567117

IEEE Trans Neural Netw Learn Syst. 2025 May 6;PP. doi: 10.1109/TNNLS.2025.3562363. Online ahead of print.

ABSTRACT

Deep neural networks (DNNs) have shown significant progress in many application fields. However, overfitting remains a significant challenge in their development. While existing data-augmentation techniques such as MixUp have been successful in preventing overfitting, they often fail to generate hard mixed samples near the decision boundary, impeding model optimization. In this article, we present adversarial MixUp (AdvMixUp), a novel sample-dependent method for regularizing DNNs. AdvMixUp addresses this issue by incorporating adversarial training (AT) to create sample-dependent and feature-level interpolation masks, generating more challenging mixed samples. These virtual samples enable DNNs to learn more robust features, ultimately reducing overfitting. Empirical evaluations on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet demonstrate that AdvMixUp outperforms existing MixUp variants.

PMID:40327482 | DOI:10.1109/TNNLS.2025.3562363

IEEE Trans Neural Netw Learn Syst. 2025 May 6;PP. doi: 10.1109/TNNLS.2025.3561225. Online ahead of print.

ABSTRACT

Deep learning models achieve remarkable accuracy in computer vision tasks yet remain vulnerable to adversarial examples-carefully crafted perturbations to input images that can deceive these models into making confident but incorrect predictions. This vulnerability poses significant risks in high-stakes applications such as autonomous vehicles, security surveillance, and safety-critical inspection systems. While the existing literature extensively covers adversarial attacks in image classification, comprehensive analyses of such attacks on object detection systems remain limited. This article presents a novel taxonomic framework for categorizing adversarial attacks specific to object detection architectures, synthesizes existing robustness metrics, and provides a comprehensive empirical evaluation of state-of-the-art attack methodologies on popular object detection models, including both traditional detectors and modern detectors with vision-language pretraining. Through rigorous analysis of open-source attack implementations and their effectiveness across diverse detection architectures, we derive key insights into attack characteristics. Furthermore, we delineate critical research gaps and emerging challenges to guide future investigations in securing object detection systems against adversarial threats. Our findings establish a foundation for developing more robust detection models while highlighting the urgent need for standardized evaluation protocols in this rapidly evolving domain.

PMID:40327472 | DOI:10.1109/TNNLS.2025.3561225