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

Bioinformatics. 2025 May 6:btaf289. doi: 10.1093/bioinformatics/btaf289. Online ahead of print.

ABSTRACT

MOTIVATION: The intricate network of RNA-RNA interactions, crucial for orchestrating essential cellular processes like transcriptional and translational regulations, has been unveiling through high-throughput techniques and computational predictions. As experimental determination of RNA-RNA interactions at the base-pair resolution remains challenging, a timely update for assessing complementary computational tools is necessary, particularly given the recent emergence of deep learning-based methods.

RESULTS: Here, we employed base pairs derived from three-dimensional RNA complex structures as a gold standard benchmark to assess the performance of 23 different methods ranging from alignment-based methods, free-energy-based minimization to deep-learning techniques. The result indicates that a deep-learning-based method, SPOT-RNA, can be generalized to make accurate zero-shot predictions of RNA-RNA interactions not only between previously unseen RNA structures but also between RNAs without monomeric structures. The finding underscores the potential of deep learning as a robust tool for advancing our understanding of these complex molecular interactions.

AVAILABILITY: All data and codes are available at https://github.com/meilanglang/RNA-RNA-Interaction.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:40327448 | DOI:10.1093/bioinformatics/btaf289

JMIR Med Inform. 2025 May 6;13:e66556. doi: 10.2196/66556.

NO ABSTRACT

PMID:40327366 | DOI:10.2196/66556

Mol Biol Cell. 2025 May 6:mbcE25010009. doi: 10.1091/mbc.E25-01-0009. Online ahead of print.

ABSTRACT

The life cycle of eukaryotic microorganisms involves complex transitions between states such as dormancy, mating, meiosis, and cell division, which are often studied independently from each other. Therefore, most microbial life cycles are theoretical reconstructions from partial observations of cellular states. Here we show that complete microbial life cycles can be directly and continuously studied by combining microfluidic culturing, life cycle stage-specific segmentation of micrographs, and a novel cell tracking algorithm, FIEST, based on deep learning video frame interpolation. As proof of principle, we quantitatively imaged and compared cell growth and the activity state of the cell division kinase, Cdk1, across the life cycle of Saccharomyces cerevisiae for up to three sexually reproducing generations. Our analysis of S. cerevisiae's life cycle provided the following new insights: (1) the accumulation of cell cycle regulators, such as Whi5, is tailored to each life cycle stage; (2) cell growth always preceded exit from non-proliferative states in our conditions; (3) the temporal coordination of meiotic events is the same across sexually reproducing populations when each generation is exposed to same conditions; (4) information such as cell size and morphology resets after each sexual reproduction cycle. Image processing and tracking algorithms are available as the Python package Yeastvision, which could be used study pathogens such as Candida glabrata, Cryptococcus neoformans, Colletotrichum acutatum, and other unicellular systems. [Media: see text] [Media: see text] [Media: see text] [Media: see text] [Media: see text] [Media: see text] [Media: see text] [Media: see text] [Media: see text].

PMID:40327364 | DOI:10.1091/mbc.E25-01-0009

JMIR Mhealth Uhealth. 2025 May 6;13:e70473. doi: 10.2196/70473.

ABSTRACT

BACKGROUND: Encouraging physical activity improves mental health and is recommended in workplace mental health guidelines. Although mobile health (mHealth) interventions are promising for physical activity promotion, their impact on mental health outcomes is inconsistent. Furthermore, poor user retention rates of mHealth apps pose a major challenge.

OBJECTIVE: This study aimed to examine the effectiveness and implementation outcomes of the smartphone app ASHARE in Japanese workplace settings, leveraging a deep learning model to monitor depression and anxiety through physical activity.

METHODS: This hybrid effectiveness-implementation trial was a 3-month nonrandomized controlled trial conducted from October 2023 to September 2024. Work units and employees were recruited and allocated to the intervention or active control group based on preference. The intervention group installed the ASHARE app, whereas the control group participated in an existing multicomponent workplace program promoting physical activity. Changes in physical activity and psychological distress levels were compared between the groups. User retention rates, participation rates, acceptability, appropriateness, feasibility, satisfaction, and potential harm were also assessed.

RESULTS: A total of 84 employees from 7 work units participated (67 from 5 units in the intervention group and 17 from 2 units in the control group). In total, 78 employees completed the 3-month follow-up survey (follow-up rate: 93%). Both groups showed increased physical activity, and the intervention group showed reduced psychological distress; however, the differences between groups were not statistically significant (P=.20; P=.36). In a sensitivity analysis of protocol-compliant employees (n=21), psychological distress levels were significantly reduced in the intervention group compared with the control group (coefficient=-3.68, SE 1.65; P=.03). The app's 3-month user retention rate was 20% (12/61), which was lower than the participation rate in each component of the control programs. Implementation outcomes evaluated by employees were less favorable in the intervention group than in the control group, whereas health promotion managers found them to be similar.

CONCLUSIONS: The ASHARE app did not show superior effectiveness compared with an existing multicomponent workplace program for promoting physical activity. An implementation gap may exist between health promotion managers and employees, possibly contributing to the app's low user retention rate. Future research should focus on examining the effectiveness of strategies to get engagement from managers and from segments of employees with favorable responses in the workplace at an early stage.

PMID:40327360 | DOI:10.2196/70473

Med Biol Eng Comput. 2025 May 6. doi: 10.1007/s11517-025-03357-3. Online ahead of print.

ABSTRACT

The corticospinal tract (CST) is the primary neural pathway responsible for voluntary motor functions, and preoperative CST reconstruction is crucial for preserving nerve functions during neurosurgery. Diffusion magnetic resonance imaging-based tractography is the only noninvasive method to preoperatively reconstruct CST in clinical practice. However, for the largesize bundle CST with complex fiber geometry (fanning fibers), reconstructing its full extent remains challenging with local-derived methods without incorporating global information. Especially in the presence of tumors, the mass effect and partial volume effect cause abnormal diffusion signals. In this work, a CST reconstruction tractography method based on a novel direction filter was proposed, designed to ensure robust CST reconstruction in the clinical dataset with tumors. A direction filter based on a fourth-order differential equation was introduced for global direction estimation. By considering the spatial consistency and leveraging anatomical prior knowledge, the direction filter was computed by minimizing the energy between the target directions and initial fiber directions. On the basis of the new directions corresponding to CST obtained by the direction filter, the fiber tracking method was implemented to reconstruct the fiber trajectory. Additionally, a deep learning-based method along with tractography template prior information was employed to generate the regions of interest (ROIs) and initial fiber directions. Experimental results showed that the proposed method yields higher valid connections and lower no connections and exhibits the fewest broken fibers and short-connected fibers. The proposed method offers an effective tool to enhance CST-related surgical outcomes by optimizing tumor resection and preserving CST.

PMID:40327206 | DOI:10.1007/s11517-025-03357-3

Med Biol Eng Comput. 2025 May 6. doi: 10.1007/s11517-025-03372-4. Online ahead of print.

ABSTRACT

Rehabilitation exercises are critical for recovering from motor dysfunction caused by neurological conditions like stroke, back pain, Parkinson's disease, and spinal cord injuries. Traditionally, these exercises require constant monitoring by therapists, which is time-consuming and costly, often leading to therapist shortages. This paper introduces a deep learning model, convolutional neural network - squeeze excitation (CNN-SE), to automate rehabilitation exercise assessment. By optimizing its parameters with the grey wolf optimization algorithm, the model was fine-tuned for optimal performance. The model's effectiveness was tested on both healthy and unhealthy participants with motor dysfunction, providing a comprehensive evaluation of its capabilities. To interpret the model's decisions and understand its inner workings, we employed Shapley additive explanations (SHAP) to analyze feature importance at each time step. Our CNN-SE model achieved a state-of-the-art mean absolute deviation of 0.127 on the KIMORE dataset and a comparable MAD of 0.014 on the UI-PRMD dataset across various exercises, demonstrating its potential to provide a cost-effective, efficient alternative to traditional therapist-led evaluations.

PMID:40327204 | DOI:10.1007/s11517-025-03372-4

EJNMMI Phys. 2025 May 6;12(1):43. doi: 10.1186/s40658-025-00756-1.

ABSTRACT

PURPOSE: Dopamine transporter (DAT) SPECT is an effective tool for early Parkinson's disease (PD) detection and heavily hampered by attenuation. Attenuation correction (AC) is the most important correction among other corrections. Transfer learning (TL) with fine-tuning (FT) a pre-trained model has shown potential in enhancing deep learning (DL)-based AC methods. In this study, we investigate leveraging realistic Monte Carlo (MC) simulation data to create a pre-trained model for TL-based AC (TLAC) to improve AC performance for DAT SPECT.

METHODS: A total number of 200 digital brain phantoms with realistic 99mTc-TRODAT-1 distribution was used to generate realistic noisy SPECT projections using MC SIMIND program and an analytical projector. One hundred real clinical 99mTc-TRODAT-1 brain SPECT data were also retrospectively analyzed. All projections were reconstructed with and without CT-based attenuation correction (CTAC/NAC). A 3D conditional generative adversarial network (cGAN) was pre-trained using 200 pairs of simulated NAC and CTAC SPECT data. Subsequently, 8, 24, and 80 pairs of clinical NAC and CTAC DAT SPECT data were employed to fine-tune the pre-trained U-Net generator of cGAN (TLAC-MC). Comparisons were made against without FT (DLAC-MC), training on purely limited clinical data (DLAC-CLI), clinical data with data augmentation (DLAC-AUG), mixed MC and clinical data (DLAC-MIX), TL using analytical simulation data (TLAC-ANA), and Chang's AC (ChangAC). All datasets used for DL-based methods were split to 7/8 for training and 1/8 for validation, and a 1-/2-/5-fold cross-validation were applied to test all 100 clinical datasets, depending on the numbers of clinical data used in the training model.

RESULTS: With 8 available clinical datasets, TLAC-MC achieved the best result in Normalized Mean Squared Error (NMSE) and Structural Similarity Index Measure (SSIM) (TLAC-MC; NMSE = 0.0143 ± 0.0082/SSIM = 0.9355 ± 0.0203), followed by DLAC-AUG, DLAC-MIX, TLAC-ANA, DLAC-CLI, DLAC-MC, ChangAC and NAC. Similar trends exist when increasing the number of clinical datasets. For TL-based AC methods, the fewer clinical datasets available for FT, the greater the improvement as compared to DLAC-CLI using the same number of clinical datasets for training. Joint histograms analysis and Bland-Altman plots of SBR results also demonstrate consistent findings.

CONCLUSION: TLAC is feasible for DAT SPECT with a pre-trained model generated purely based on simulation data. TLAC-MC demonstrates superior performance over other DL-based AC methods, particularly when limited clinical datasets are available. The closer the pre-training data is to the target domain, the better the performance of the TLAC model.

PMID:40327202 | DOI:10.1186/s40658-025-00756-1

Eur Spine J. 2025 May 6. doi: 10.1007/s00586-025-08895-w. Online ahead of print.

ABSTRACT

PURPOSE: Scoliosis is a prevalent spine deformity that impacts millions of children globally. The Cobb angle, a crucial and widely-accepted metric, serves as the "gold standard" for assessing scoliosis in patients. However, the traditional manual measurement of spine curvature is time-consuming and labor-intensive. It also comes with issues like intra - and inter-observer variations. Moreover, accurately and robustly evaluating Cobb angles is extremely challenging. This is because it necessitates the correct identification of all the required vertebrae in both the anterior-posterior (AP) and lateral (LAT) views of full-spine digital radiography (DR).

METHODS: To solve these challenges, a deep learning-based framework is developed to fully automatically measure patient Cobb angels from full-spine DR of both AP and LAT views. First, a deep learning network was used to distinguish AP and LAT views. Then the region of interest (ROI) of the whole spine was located and extracted. Subsequently, a detection network was applied to detect and identify the boundaries and locations, the types, and the four corner points of each spinal vertebra. Finally, the Cobb angles was measured automatically. When taking into account the location, recognition, and key points detection of spinal vertebrae, YOLOv8 architecture with CBAM module was adopted as the backbone.

RESULTS: A total of 1,163 AP view and 1,378 LAT view DR images were used to train and evaluate the models. Experimental results in the evaluation testing showed a mean Cobb angle error of 2.56° for AP view and 2.498° for LAT view DR images. The intra-class correlation coefficient (ICC) with 95% confidence interval (CI) was 0.956 (0.932, 0.972) for AP view and 0.925 (0.888, 0.952) for LAT view. The Pearson correlation coefficient was 0.961 for AP view and 0.930 for LAT view. In the comprehensive reader study, for the major curve, a mean Cobb angle error of 3.918°, an ICC of 0.943 (0.912, 0.965), and a high correlation coefficient of 0.960 were obtained.

CONCLUSION: The results showed that the proposed framework had a significant accuracy and consistency advantage in measuring Cobb angle, which not only validated the effectiveness of the algorithm, but also provided strong support for the diagnosis of clinicians.

PMID:40327070 | DOI:10.1007/s00586-025-08895-w

Radiology. 2025 May;315(2):e240714. doi: 10.1148/radiol.240714.

ABSTRACT

Background Four-dimensional (4D) flow MRI provides assessment of thoracic aorta hemodynamic measures that are increasingly recognized as important biomarkers for risk assessment. However, long acquisition times and cumbersome data analysis limit widespread availability. Purpose To evaluate the feasibility and accuracy of a generative artificial intelligence (AI) approach (fluid physics-informed cycle generative adversarial network [FPI-CycleGAN]) in quantifying aorta hemodynamics directly from anatomic input as an alternative to 4D flow MRI. Materials and Methods Patients were retrospectively identified from a dataset of clinical cardiothoracic MRI examinations performed between November 2011 and July 2020. All patients underwent aortic 4D flow MRI, which served as a reference standard for training and testing of FPI-CycleGANs. A three-dimensional (3D) segmentation of the aortic geometry was used as the only input to predict systolic aortic hemodynamics, with separate networks for bicuspid aortic valve (BAV) (994 in the training set and 248 in the test set) and tricuspid aortic valve (TAV) (419 in the training set and 104 in the test set). Voxel-by-voxel and regional analyses were used to quantify and compare (AI vs the reference standard, 4D flow) systolic velocity vector fields, peak velocity, wall shear stress (WSS), and classification of aortic valve stenosis. Results In total, 1765 patients (median age, 53 years [IQR, 41-63 years]; 1242 patients had BAV and 523 had TAV) were included. Mean AI computation time was 0.15 second ± 0.11 (SD), and total training was 1500 and 3600 minutes for the TAV and BAV networks, respectively. The FPI-CycleGAN predicted systolic 3D velocity vector fields accurately, with low bias (<0.01 m/sec) and excellent limits of agreements (±0.06-0.08 m/sec). For peak velocities and WSS, there was strong agreement between FPI-CycleGAN and 4D flow (r2 = 0.930-0.957 [P < .001], with relative differences of 6.2%-9.8%). AI accurately classified aortic valve stenosis severity in 85.8% of patients (302 of 352) (κ = 0.80 [95% CI: 0.71, 0.89]). The FPI-CycleGAN was robust to one- and two-voxel dilation and erosion (bias, -0.05 to 0.1 m/sec) and ±5° rotation (bias, -0.02 to 0.03 m/sec) of the input data. The application of the trained FPI-CycleGAN in an external test set with contrast-enhanced MR angiography (n = 60 patients) as AI input data demonstrated strong to excellent performance for peak velocities and WSS (r2 = 0.944-0.965 [P < .001], with relative differences of 6.2%-9.2%). Conclusion Aorta 3D hemodynamics can be derived from anatomic input in less than 1 second using an FPI-CycleGAN and demonstrate strong agreement with in vivo 4D flow MRI systolic hemodynamics. © RSNA, 2025 Supplemental material is available for this article.

PMID:40326877 | DOI:10.1148/radiol.240714

J Acoust Soc Am. 2025 May 1;157(5):3508-3523. doi: 10.1121/10.0036456.

ABSTRACT

Underwater passive acoustic recognition, which focuses on classifying targets based on ship-radiated noise, is a key challenge in underwater acoustics. Deep learning-based methods have gained popularity in recent years because of their strong performance. However, these methods often fail to generalize well in real-world scenarios. This work reveals one underlying challenge: the characteristics of ship-radiated noise are influenced by factors such as vessel structures and propulsion systems. Although vessels of the same type may exhibit different patterns in these aspects, vessels of different categories share similarities. As a result, data-driven models often tend to overemphasize individual-specific features, leading to "overfitting" and poor generalization. The momentum-based adversarial training (MBAT) framework is proposed to mitigate this challenge. MBAT leverages a momentum adversarial strategy to use category information and individual vessel relationships, helping extract class-discriminative features. A homoscedastic uncertainty algorithm is employed to balance the optimization objectives of category-related and vessel-specific features. These strategies allow the model to capture category-discriminative patterns more effectively and generalize to unseen targets. Experiments on DeepShip and ShipsEar demonstrate that MBAT significantly improves generalization capability on unseen individual vessels, outperforming existing state-of-the-art methods. Visualizations further confirm the efficacy and necessity of the proposed approach.

PMID:40326792 | DOI:10.1121/10.0036456

Bioinformatics. 2025 May 6:btaf277. doi: 10.1093/bioinformatics/btaf277. Online ahead of print.

ABSTRACT

MOTIVATION: The availability of very large numbers of protein structures from accurate computational methods poses new challenges in storing, searching and detecting relationships between these structures. In particular, the new-found abundance of multi-domain structures in the AlphaFold structure database introduces challenges for traditional structure comparison methods.

RESULTS: We address these challenges using a fast, embedding-based structure comparison method called Foldclass which detects structural similarity between protein domains. We demonstrate the accuracy of Foldclass embeddings for homology detection. In combination with a recently developed deep learning-based automatic domain segmentation tool Merizo, we develop Merizo-search, which first segments multi-domain query structures into domains, and then searches a Foldclass embedding database to determine the top matches for each constituent domain. Combining the ability of Merizo to accurately segment complete chains into domains, and Foldclass to embed and detect similar domains, the Merizo-search tool can be used to rapidly detect per-domain similarities for complete chains, taking as little as 2 minutes to search all 365 million domains from the Encyclopedia of Domains. We anticipate that these tools will enable many analyses using the wealth of predicted structural data now available.

AVAILABILITY: Foldclass and Merizo-search are available at https://github.com/psipred/merizo_search. The version used in this publication is archived at https://doi.org/10.5281/zenodo.15120830. Merizo-search is also available on the PSIPRED web server at http://bioinf.cs.ucl.ac.uk/psipred.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:40326701 | DOI:10.1093/bioinformatics/btaf277

Curr Opin Pulm Med. 2025 May 6. doi: 10.1097/MCP.0000000000001179. Online ahead of print.

ABSTRACT

PURPOSE OF REVIEW: This review evaluates the role of artificial intelligence (AI) in diagnosing solitary pulmonary nodules (SPNs), focusing on clinical applications and limitations in pulmonary medicine. It explores AI's utility in imaging and blood/tissue-based diagnostics, emphasizing practical challenges over technical details of deep learning methods.

RECENT FINDINGS: AI enhances computed tomography (CT)-based computer-aided diagnosis (CAD) through steps like nodule detection, false positive reduction, segmentation, and classification, leveraging convolutional neural networks and machine learning. Segmentation achieves Dice similarity coefficients of 0.70-0.92, while malignancy classification yields areas under the curve of 0.86-0.97. AI-driven blood tests, incorporating RNA sequencing and clinical data, report AUCs up to 0.907 for distinguishing benign from malignant nodules. However, most models lack prospective, multiinstitutional validation, risking overfitting and limited generalizability. The "black box" nature of AI, coupled with overlapping inputs (e.g., nodule size, smoking history) with physician assessments, complicates integration into clinical workflows and precludes standard Bayesian analysis.

SUMMARY: AI shows promise for SPN diagnosis but requires rigorous validation in diverse populations and better clinician training for effective use. Rather than replacing judgment, AI should serve as a second opinion, with its reported performance metrics understood as study-specific, not directly applicable at the bedside due to double-counting issues.

PMID:40326426 | DOI:10.1097/MCP.0000000000001179

Circ Cardiovasc Imaging. 2025 May 6:e018291. doi: 10.1161/CIRCIMAGING.125.018291. Online ahead of print.

NO ABSTRACT

PMID:40326361 | DOI:10.1161/CIRCIMAGING.125.018291

Brain. 2025 May 6:awaf172. doi: 10.1093/brain/awaf172. Online ahead of print.

ABSTRACT

Migraine has an assumed polygenic basis, but the genetic risk variants identified in genome-wide association studies only explain a proportion of the heritability. We aimed to develop machine learning models, capturing non-additive and interactive effects, to address the missing heritability. This was a cross-sectional population-based study of participants in the second and third Trøndelag Health Study. Individuals underwent genome-wide genotyping and were phenotyped based on validated modified criteria of the International Classification of Headache Disorders. Four datasets of increasing number of genetic variants were created using different thresholds of linkage disequilibrium and univariate genome-wide associated p-values. A series of machine learning and deep learning methods were optimized and evaluated. The genotype tools PLINK and LDPred2 were used for polygenic risk scoring. Models were trained on a partition of the dataset and tested in a hold-out set. The area under the receiver operating characteristics curve was used as the primary scoring metric. Classification by machine learning was statistically compared to that of polygenic risk scoring. Finally, we explored the biological functions of the variants unique to the machine learning approach. 43,197 individuals (51% women), with a mean age of 54.6 years, were included in the modelling. A light gradient boosting machine performed best for the three smallest datasets (108, 7,771 and 7,840 variants), all with hold-out test set area under curve at 0.63. A multinomial naïve Bayes model performed best in the largest dataset (140,467 variants) with a hold-out test set area under curve of 0.62. The models were statistically significantly superior to polygenic risk scoring (area under curve 0.52 to 0.59) for all the datasets (p<0.001 to p=0.02). Machine learning identified many of the same genes and pathways identified in genome-wide association studies, but also several unique pathways, mainly related to signal transduction and neurological function. Interestingly, pathways related to botulinum toxins, and pathways related to the calcitonin gene-related peptide receptor also emerged. This study suggests that migraine may follow a non-additive and interactive genetic causal structure, potentially best captured by complex machine learning models. Such structure may be concealed where the data dimensionality (high number of genetic variants) is insufficiently supported by the scale of available data, leaving a misleading impression of purely additive effects. Future machine learning models using substantially larger sample sizes could harness both the additive and the interactive effects, enhancing precision and offering deeper understanding of genetic interactions underlying migraine.

PMID:40326299 | DOI:10.1093/brain/awaf172

Sci Rep. 2025 May 5;15(1):15719. doi: 10.1038/s41598-025-98873-5.

ABSTRACT

Polycystic Ovarian Disease (PCOD) is among the most prevalent endocrine disorders complicating the health of innumerable women worldwide due to lack of diagnosis and appropriate management. The diagnosis of PCOD, along with proper classification with the help of ultrasound imaging, would be of immense importance for early intervention and timely management of the condition. However, most of the existing approaches suffer from lots of problems, including low accuracy and capability in feature extraction, and may also be resilient to noise; it can further delay or lead to a wrong diagnosis. The main objective of this paper is to address these important issues by proposing a deep learning model, Holographic Convolutional Dense Network (Coco-HoloNet) that will be tailored for the precise detection and classification of PCOD in ultrasound images with high accuracy. These are multi-fold contributions which focus on improvement in diagnostic accuracy by overcoming the various limitations of conventional approaches. CoCo-HoloNet is using a layered architecture by integrating convolutional layers, dense blocks, and pooling strategies that leverage capturing and extraction of significant features from the input effectively. More importantly, the model is also embedded with the Tangent-Runner Adaptive Optimization (TRAdO) technique, which dynamically calculates the regularization parameters to overcome overfitting problems and improves the generalization capability of the model. The approach not only ensures the richest possible feature representation, but it also results in outstanding improvements within the performance measures of a model, such that the accuracy rate exceeds 99%. Further experimentation with CoCo-HoloNet on an extended Kaggle PCOD ultrasound image dataset proves its effectiveness by reporting higher precision, recall, and F1-scores than those obtained by state-of-the-art existing methods.

PMID:40325103 | DOI:10.1038/s41598-025-98873-5

Sci Rep. 2025 May 5;15(1):15713. doi: 10.1038/s41598-025-00401-y.

ABSTRACT

The forecasting of a patient's response to radiotherapy and the likelihood of experiencing harmful long-term health impacts would considerably enhance individual treatment plans. Due to the continuous exposure to radiation, cardiovascular disease and pulmonary fibrosis might occur. For forecasting the response of patients to chemotherapy, the Convolutional Neural Networks (CNN) technique is widely used. With the help of radiotherapy, cancer diseases are diagnosed, but some patients suffer from side effects. The toxicity of radiotherapy and chemotherapy should be estimated. For validating the patient's improvement in treatments, a patient response prediction system is essential. In this paper, a Deep Learning (DL) based patient response prediction system is developed to effectively predict the response of patients, predict prognosis and inform the treatment plans in the early stage. The necessary data for the response prediction are collected manually. The collected data are then processed through the feature selection segment. The Repeated Exploration and Exploitation-based Coati Optimization Algorithm (REE-COA) is employed to select the features. The selected weight features are input into the prediction process. Here, the prediction is performed by Multi-scale Dilated Ensemble Network (MDEN), where we integrated Long-Short term Memory (LSTM), Recurrent Neural Network (RNN) and One-dimensional Convolutional Neural Networks (1DCNN). The final prediction scores are averaged to develop an effective MDEN-based model to predict the patient's response. The proposed MDEN-based patient's response prediction scheme is 0.79%, 2.98%, 2.21% and 1.40% finer than RAN, RNN, LSTM and 1DCNN, respectively. Hence, the proposed system minimizes error rates and enhances accuracy using a weight optimization technique.

PMID:40325044 | DOI:10.1038/s41598-025-00401-y

J Biomed Inform. 2025 May 3:104826. doi: 10.1016/j.jbi.2025.104826. Online ahead of print.

ABSTRACT

OBJECTIVE: Deep learning approaches have demonstrated significant potential in predicting temporal health events in recent years. However, existing methods have not fully leveraged the complex interactions among comorbidities and have overlooked imbalances and temporal irregularities in admission records.

METHODS: This study proposes GLT-Net, a deep learning approach that combines Graph Learning with Transformer framework to tackle these challenges. GLT-Net begins by constructing a patient association graph to generate unique representations for each individual. At the same time, the hierarchical structure of diagnosis codes is utilized to pre-train the diagnosis code embeddings. Subsequently, a comorbidity association matrix is created to illustrate the relationships between comorbidities, and graph neural networks are employed to enhance the feature representations of diagnosis codes. Finally, a Transformer-Encoder framework captures the dependencies in historical admission records by incorporating time information.

RESULTS: We demonstrate our approach on two tasks in temporal health event predcition. Experimental results on real-world datasets show that GLT-Net outperforms baseline models in forecasting temporal health events. Additionally, a case study demonstrates the effectiveness of GLT-Net in predicting health events.

CONCLUSION: Understanding progression patterns over time, comorbidity associations, and patient characterization is essential for predicting temporal health events. Our study provides new insights and methods for a deeper understanding of patient health status and disease trends. Moreover, our model can be extended to other data sources, enhancing its versatility.

PMID:40324665 | DOI:10.1016/j.jbi.2025.104826

J Cheminform. 2025 May 5;17(1):70. doi: 10.1186/s13321-025-01023-2.

ABSTRACT

This study demonstrates the utility of a novel molecular representation, 3D APM and a deep learning model based on it for virtual screening, suggesting that many other prediction models would also benefit from adopting APM. An open-source script to generate 3D APM is available at https://github.com/rimeless/APM.

PMID:40325489 | DOI:10.1186/s13321-025-01023-2

BMC Med Imaging. 2025 May 5;25(1):149. doi: 10.1186/s12880-025-01694-1.

ABSTRACT

Diabetes is a widespread condition that can lead to serious vision problems over time. Timely identification and treatment of diabetic retinopathy (DR) depend on accurately segmenting retinal vessels, which can be achieved through the invasive technique of fundus imaging. This methodology facilitates the systematic monitoring and assessment of the progression of DR. In recent years, deep learning has made significant steps in various fields, including medical image processing. Numerous algorithms have been developed for segmenting retinal vessels in fundus images, demonstrating excellent performance. However, it is widely recognized that large datasets are essential for training deep learning models to ensure they can generalize well. A major challenge in retinal vessel segmentation is the lack of ground truth samples to train these models. To overcome this, we aim to generate synthetic data. This work draws inspiration from recent advancements in generative adversarial networks (GANs). Our goal is to generate multiple realistic retinal fundus images based on tubular structured annotations while simultaneously creating binary masks from the retinal fundus images. We have integrated a latent space auto-encoder to maintain the vessel morphology when generating RGB fundus images and mask images. This approach can synthesize diverse images from a single tubular structured annotation and generate various tubular structures from a single fundus image. To test our method, we utilized three primary datasets, DRIVE, STARE, and CHASE_DB, to generate synthetic data. We then trained and tested a simple UNet model for segmentation using this synthetic data and compared its performance against the standard dataset. The results indicated that the synthetic data offered excellent segmentation performance, a crucial aspect in medical image analysis, where smaller datasets are often common. This demonstrates the potential of synthetic data as a valuable resource for training segmentation and classification models for disease diagnosis. Overall, we used the DRIVE, STARE, and CHASE_DB datasets to synthesize and evaluate the proposed image-to-image translation approach and its segmentation effectiveness.

PMID:40325399 | DOI:10.1186/s12880-025-01694-1