PLoS One. 2025 Jul 28;20(7):e0326255. doi: 10.1371/journal.pone.0326255. eCollection 2025.

ABSTRACT

Intracranial haemorrhage (ICH) is a crucial medical emergency that entails prompt assessment and management. Compared to conventional clinical tests, the need for computerized medical assistance for properly recognizing brain haemorrhage from computer tomography (CT) scans is more mandatory. Various deep learning (DL) and artificial intelligence (AI) technologies have been successfully implemented for the analysis of medical images, namely grading of diabetic retinopathy (DR), breast cancer detection, skin cancer detection, and so on. Furthermore, the AI approach ensures accurate detection to facilitate early detection, drastically decreasing the mortality rate. Based on DL models, there are already various techniques for ICHdetection. This manuscript proposes the design of a Hyperparameter Tuned Deep Learning-Driven Medical Image Analysis for Intracranial Hemorrhage Detection (HPDL-MIAIHD) technique. The proposed HPDL-MIAIHD technique investigates the available CT images to classify and identify the ICH. In the presented HPDL-MIAIHD technique, the median filtering (MF) approach is utilized for the image preprocessing step. Next, the HPDL-MIAIHD approach uses an enhanced EfficientNet technique to extract feature vectors. To increase the efficiency of the EfficientNet method, the hyperparameter tuning process is performed by utilizing the chimp optimizer algorithm (COA) method. The ICH detection process is accomplished by the ensemble classification process, comprising long short-term memory (LSTM), stacked autoencoder (SAE), and bidirectional LSTM (Bi-LSTM) networks. Lastly, the Bayesian optimizer algorithm (BOA) is implemented for the hyperparameter selection of the DL technique. A comprehensive simulation was conducted on the benchmark CT image dataset to demonstrate the effectiveness of the HPDL-MIAIHD approach in detecting ICH. The performance validation of the HPDL-MIAIHD approach portrayed a superior accuracy value of 99.02% over other existing models.

PMID:40720400 | DOI:10.1371/journal.pone.0326255

PLoS One. 2025 Jul 28;20(7):e0328402. doi: 10.1371/journal.pone.0328402. eCollection 2025.

ABSTRACT

Skin cancer is a significant global public health issue, with millions of new cases identified each year. Recent breakthroughs in artificial intelligence, especially deep learning, possess considerable potential to enhance the accuracy and efficiency of screening. This study proposes an approach that employs smartphone images, which are preprocessed using adaptive learning and Black-Hat transformation. ViT is utilized for feature extraction, and a stacking model is constructed employing these features in conjunction with image-related variables, like patient age and sex, for final classification. The model's efficacy in identifying cancer-associated skin diseases was evaluated across six categories of skin lesions: actinic keratosis, basal cell carcinoma, melanoma, nevus, squamous cell carcinoma, and seborrheic keratosis. The suggested model attained an overall accuracy of 97.61%, with a PVV of 96.88%, a recall of 97.63%, and an F1 score of 97.19%, so illustrating its efficacy in detecting malignant skin lesions. This method could greatly aid dermatologists by enhancing diagnostic sensitivity and specificity, reducing delays in identifying the most suspicious lesions, and ultimately reaching more patients in need of timely screenings and patient care, thus saving lives.

PMID:40720382 | DOI:10.1371/journal.pone.0328402

IEEE Trans Med Imaging. 2025 Jul 28;PP. doi: 10.1109/TMI.2025.3589495. Online ahead of print.

ABSTRACT

Deep learning-based medical image segmentation techniques have shown promising results when evaluated based on conventional metrics such as the Dice score or Intersection-over-Union. However, these fully automatic methods often fail to meet clinically acceptable accuracy, especially when topological constraints should be observed, e.g., continuous boundaries or closed surfaces. In medical image segmentation, the correctness of a segmentation in terms of the required topological genus sometimes is even more important than the pixel-wise accuracy. Existing topology-aware approaches commonly estimate and constrain the topological structure via the concept of persistent homology (PH). However, these methods are difficult to implement for high dimensional data due to their polynomial computational complexity. To overcome this problem, we propose a novel and fast approach for topology-aware segmentation based on the Euler Characteristic (χ). First, we propose a fast formulation for χ computation in both 2D and 3D. The scalar χ error between the prediction and ground-truth serves as the topological evaluation metric. Then we estimate the spatial topology correctness of any segmentation network via a so-called topological violation map, i.e., a detailed map that highlights regions with χ errors. Finally, the segmentation results from the arbitrary network are refined based on the topological violation maps by a topology-aware correction network. Our experiments are conducted on both 2D and 3D datasets and show that our method can significantly improve topological correctness while preserving pixel-wise segmentation accuracy.

PMID:40720275 | DOI:10.1109/TMI.2025.3589495

IEEE J Biomed Health Inform. 2025 Jul 28;PP. doi: 10.1109/JBHI.2025.3593028. Online ahead of print.

ABSTRACT

Clinical coding translates medical information from Electronic Health Records (EHRs) into structured codes such as ICD-10, which are essential for healthcare applications. Advances in deep learning and natural language processing have enabled automatic ICD coding models to achieve notable accuracy metrics on in-domain datasets when adequately trained. However, the scarcity of clinical medical texts and the variability across different datasets pose significant challenges, making it difficult for current state-of-the-art models to ensure robust generalization performance across diverse data distributions. Recent advances in Large Language Models (LLMs), such as GPT-4o, have shown great generalization capabilities across general domains and potential in medical information processing tasks. However, their performance in generating clinical codes remains suboptimal. In this study, we propose a novel ICD coding paradigm based on code verification to leverage the capabilities of LLMs. Instead of directly generating accurate codes from a vast code space, we simplify the task by verifying the code assignment from a given candidate set. Through extensive experiments, we demonstrate that LLMs function more effectively as code verifiers rather than code generators, with GPT-4o achieving the best performance on the CodiEsp dataset under zero-shot settings. Furthermore, our results indicate that LLM-based systems can perform on par with state-of-the-art clinical coding systems while offering superior generalizability across institutions, languages, and ICD versions.

PMID:40720269 | DOI:10.1109/JBHI.2025.3593028

IEEE J Biomed Health Inform. 2025 Jul 28;PP. doi: 10.1109/JBHI.2025.3593388. Online ahead of print.

ABSTRACT

Heart failure (HF) is a complex and heterogeneous syndrome caused by diverse factors, such as atrial fibrillation, diabetes, and pulmonary hypertension. The intricate pathophysiology of HF, coupled with variability in patient demographics and presentations, poses significant challenges to the effectiveness of existing deep learning models in HF risk prediction. In this paper, we propose a novel deep neural network, PDSNet, which leverages a new dual patient-disease spatial similarity strategy to improve HF risk prediction using short electronic health records. First, we develop ontology graphs to capture hierarchical relationships between patients based on HF-related symptoms and causes; Then, a bipartite graph model is utilized to learn spatial similarities between patients with similar hospital visit histories; Finally, we design a transformer-based architecture to integrate both temporal and spatial dynamics for predicting future hospital visits associated with HF risk. We benchmarked the PDSNet on predicting HF risk for 7,346 patients from the MIMIC-III dataset. Compared to seven state-of-the-art deep learning methods, our PDSNet model achieved improvements of 2%-12% in the area under the curve (AUC) score and 3%-18% in the F1 score. These findings highlight the promising potential of our proposed PDSNet to provide accurate and robust HF risk predictions, paving the way for efficient clinical decision support and personalized HF management.

PMID:40720268 | DOI:10.1109/JBHI.2025.3593388

IEEE J Biomed Health Inform. 2025 Jul 28;PP. doi: 10.1109/JBHI.2025.3593443. Online ahead of print.

ABSTRACT

As Deep Learning (DL) algorithms become more widely adopted in healthcare applications, there is a greater emphasis on understanding and addressing potential privacy risks associated with these models. The purpose of this study is to investigate the privacy vulnerabilities of the Convolutional Neural Network (CNN) classifiers for Electroencephalogram (EEG) data in the Brain-Computer Interfaces (BCIs). Specifically, it focuses on the Membership Inference Attack (MIA), which seeks to determine if data from an individual were used in model training. The novelty of this work lies in its empirical analysis of MIA, specifically by addressing two key challenges that are less common in other domains: 1) datasets that are heterogeneous and 2) spatial-temporal design choices. Motivated by these challenges, we investigate the susceptibility to MIA based on: 1) specifics of the training dataset (number of participants, demographics), and 2) specifics of the CNN (such as architecture, regularization). Our experiments revealed that an adversary with limited knowledge of the model and its training process can compromise the privacy of training participants, noting that the same attack is not effective against deep learning models trained on image and tabular datasets. Some of our findings are: 1) training on diverse participant datasets improves the privacy of the most participants but increases risks of memorization and vulnerabilities for underrepresented groups; 2) regularization is less effective in defending against the MIA for EEG data CNN classifiers when compared to other types of input data; 3) depth and width of model architecture has no impact on membership attack effectiveness. We hope that the presented insights will assist future researchers develop more privacy-aware deep learning based BCI systems.

PMID:40720264 | DOI:10.1109/JBHI.2025.3593443

Appl Neuropsychol Child. 2025 Jul 28:1-11. doi: 10.1080/21622965.2025.2535017. Online ahead of print.

ABSTRACT

Attention-deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental disorder in children, impacting academic performance, social interactions, and overall well-being. Early and accurate diagnosis is crucial, yet current methods rely heavily on subjective assessments. This study presents a novel Wavelet-Attention deep model for objective ADHD diagnosis using electroencephalography signals. The model integrates a wavelet transform for feature extraction with a deep residual network (ResNet) augmented by an attention mechanism to enhance focus on salient features. Rigorous preprocessing, including Independent Component Analysis for artifact removal, was applied to a publicly available dataset of 121 children. To ensure a robust and clinically relevant evaluation that avoids data leakage, a strict Leave-One-Subject-Out cross-validation protocol was employed. The proposed model demonstrated strong diagnostic performance, achieving an accuracy of 96.69%, a sensitivity of 95.08%, and a specificity of 98.33% in distinguishing between children with ADHD and healthy controls. Furthermore, model-agnostic interpretability analysis revealed that features derived from frontal lobe channels and low-frequency wavelet coefficients were most critical for the model's decisions, aligning with established neurophysiological markers of ADHD. The results suggest that this approach holds significant potential for developing a reliable and objective diagnostic tool for ADHD, facilitating earlier and more personalized interventions.

PMID:40720204 | DOI:10.1080/21622965.2025.2535017

Commun Eng. 2025 Jul 26;4(1):135. doi: 10.1038/s44172-025-00467-6.

ABSTRACT

Healthcare wearables are transforming health monitoring, generating vast and complex data in everyday free-living environments. While supervised deep learning has enabled tremendous advances in interpreting such data, it remains heavily dependent on large labeled datasets, which are often difficult and expensive to obtain in clinical practice. Self-supervised contrastive learning (SSCL) provides a promising alternative by learning from unlabeled data, but conventional SSCL frequently overlooks important physiological similarities by treating all non-identical instances as unrelated, which can result in suboptimal representations. In this study, we revisit the enduring value of domain knowledge "embedded" in traditional domain feature engineering pipelines and demonstrate how it can be used to guide SSCL. We introduce a framework that integrates clinically meaningful features-such as heart rate variability from electrocardiograms (ECGs)-into the contrastive learning process. These features guide the formation of more relevant positive pairs through nearest-neighbor matching and promote global structure through clustering-based prototype representations. Evaluated across diverse wearable technologies, our method achieves comparable performance with only 10% labeled data, compared to conventional SSCL approaches with full annotations for fine-tuning. This work highlights the indispensable and sustainable role of domain expertise in advancing machine learning for real-world healthcare, especially for healthcare wearables.

PMID:40715702 | DOI:10.1038/s44172-025-00467-6

Mol Divers. 2025 Jul 25. doi: 10.1007/s11030-025-11301-8. Online ahead of print.

ABSTRACT

P-glycoprotein (P-gp) is a crucial drug transporter in several drug-resistant cases that are serious challenges in drug delivery and cancer treatment. Existing computational approaches mostly depended on small datasets for predicting P-gp interactions. To overcome these limitations, this paper proposes a Novel Ligand-based Convolutional Neural Network (NLCNN) framework to classify and predict P-gp substrates with high accuracy. The model is trained on a curated dataset of 197 P-gp substrates, integrating molecular docking and ligand-based deep learning methods for further predictive improvement. Experimental evaluations show that the NLCNN, on average, achieves prediction accuracy of 80%, which is 19-24% higher in precision and recall metrics compared to conventional Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs). Here, CNN is considered as the major model, whereas the SVM is emphasized as a baseline classifier. The proposed FLCNN algorithm obtains a noticeable accuracy, thus outperforming conventional SVM with a Gaussian RBF kernel. Moreover, by using the X-ray structure of mouse P-gp as a template, a homology model of human P-gp permits accurate molecular docking analysis. The proposed model is implemented in drug discovery and personalized medicine for P-gp interaction prediction. This is a landmark achievement in computational pharmacology as it puts out a powerful, accurate, and simple tool for determining P-gp inhibitors and substrates.

PMID:40715638 | DOI:10.1007/s11030-025-11301-8

Sci Rep. 2025 Jul 26;15(1):27302. doi: 10.1038/s41598-025-13221-x.

ABSTRACT

Accurate river discharge forecasting is essential for effective water resource management, particularly in regions prone to monsoonal variability and extreme weather events. This study presents an interpretable deep learning framework for daily river discharge forecasting in the Subarnarekha river basin (SRB), integrating Kolmogorov Arnold networks (KAN) with Shapley additive exPlanations (SHAP). Leveraging hydroclimatic inputs from five coupled model intercomparison project phase 6 (CMIP6) general circulation models (GCM) under the high emissions shared socioeconomic pathway (SSP585) scenario, the model was trained and evaluated across four active gauging stations: Muri, Adityapur, Jamshedpur, and Ghatsila covering the period 1980 to 2022, with projections extending to 2100. The main findings of this study are (1) KAN demonstrated high predictive performance with root mean squared error (RMSE) values ranging from 42.7 to 58.3 m3/s, Nash-Sutcliffe efficiency (NSE) between 0.80 and 0.87, mean absolute error (MAE) between 28.9 to 52.7 and R2 values between 0.84 and 0.90 across stations. (2) SHAP based feature contribution analysis identified Relative humidity (hurs), specific humidity (huss), and temperature (tas) as key predictors, while (pr) showed limited contribution due to spatial inherent inconsistencies in GCM precipitation data. (3) The bootstrapped SHAP distributions highlighted substantial variability in feature importance, particularly for humidity variables, revealing station specific uncertainty patterns in model interpretation. (4) The KAN framework results indicate strong temporal alignment and physical realism, confirming KAN's robustness in capturing seasonal discharge dynamics and extreme flow events under monsoon influence environments. (5) In this study KAN with SHAP (SHapley additive exPlanations) is implemented for hydrological modeling under monsoon-influenced and data-limited regions such as SRB, offering improved accuracy, functional precision and efficiency compared to traditional models. The explainability offered by SHAP confirms informed water resource planning. This novel framework presents a reproducible and climate-resilient decision support tool, particularly suitable for monsoon-influenced, data-limited basins susceptible to extreme hydroclimatic events.

PMID:40715581 | DOI:10.1038/s41598-025-13221-x

Sci Rep. 2025 Jul 25;15(1):27105. doi: 10.1038/s41598-025-12595-2.

ABSTRACT

This study explores an artificial intelligence-based approach to perform CT-free quantitative SPECT for kidney imaging using Tc-99 m DTPA, aiming to estimate glomerular filtration rate (GFR) without relying on CT. A total of 1000 SPECT/CT scans were used to train and test a deep-learning model that segments kidneys automatically based on synthetic attenuation maps (µ-maps) derived from SPECT alone. The model employed a residual U-Net with edge attention and was optimized using windowing-maximum normalization and a generalized Dice similarity loss function. Performance evaluation showed strong agreement with manual CT-based segmentation, achieving a Dice score of 0.818 ± 0.056 and minimal volume differences of 17.9 ± 43.6 mL (mean ± standard deviation). An additional set of 50 scans confirmed that GFR calculated from the AI-based CT-free SPECT (109.3 ± 17.3 mL/min) was nearly identical to the conventional SPECT/CT method (109.2 ± 18.4 mL/min, p = 0.9396). This CT-free method reduced radiation exposure by up to 78.8% and shortened segmentation time from 40 min to under 1 min. The findings suggest that AI can effectively replace CT in kidney SPECT imaging, maintaining quantitative accuracy while improving safety and efficiency.

PMID:40715554 | DOI:10.1038/s41598-025-12595-2

J Imaging Inform Med. 2025 Jul 25. doi: 10.1007/s10278-025-01592-6. Online ahead of print.

ABSTRACT

Chest X-ray is one of the main diagnostic tools in emergency medicine, yet its limited ability to capture fine anatomical details can result in missed or delayed diagnoses. To address this, we introduce XVertNet, a novel deep-learning framework designed to enhance vertebral structure visualization in X-ray images significantly. Our framework introduces two key innovations: (1) an unsupervised learning architecture that eliminates reliance on manually labeled training data-a persistent bottleneck in medical imaging, and (2) a dynamic self-tuned internal guidance mechanism featuring an adaptive feedback loop for real-time image optimization. Extensive validation across four major public datasets revealed that XVertNet outperforms state-of-the-art enhancement methods, as demonstrated by improvements in evaluation measures such as entropy, the Tenengrad criterion, LPC-SI, TMQI, and PIQE. Furthermore, clinical validation conducted by two board-certified clinicians confirmed that the enhanced images enabled more sensitive examination of vertebral structural changes. The unsupervised nature of XVertNet facilitates immediate clinical deployment without requiring additional training overhead. This innovation represents a transformative advancement in emergency radiology, providing a scalable and time-efficient solution to enhance diagnostic accuracy in high-pressure clinical environments.

PMID:40715865 | DOI:10.1007/s10278-025-01592-6

J Imaging Inform Med. 2025 Jul 25. doi: 10.1007/s10278-025-01613-4. Online ahead of print.

ABSTRACT

Breast cancer remains a global health challenge, ranking as a leading cause of mortality among women worldwide, emphasizing the need for early detection to improve treatment outcomes. Architectural distortion (AD), characterized by subtle disruptions in breast tissue patterns, is one of the earliest and most challenging signs of breast cancer to detect, often manifesting up to 2 years before other signs and providing a critical window for intervention. To address the limitations of traditional AD segmentation methods, this study introduces an advanced deep learning approach to automate and improve AD segmentation and classification on mammograms, significantly reducing radiologists' workload and enhancing diagnostic accuracy. The process began with implementing U-Net++ as a semantic segmentation model for pixel-level classification. Next, a standalone Mask R-CNN was applied, which incorporated instance segmentation for more precise detection of AD regions. Finally, the study introduced a two-phase pipeline that combines Mask R-CNN for segmentation with a ResNet-18 classification model to refine predictions and reduce false positives. Key enhancements, such as smooth L1 for bounding box regression loss and binary cross entropy with Dice loss for mask prediction, significantly improved segmentation metrics. The integrated approach achieved remarkable results with a segmentation accuracy of 0.852, a classification accuracy of 0.915, and a mean average precision (mAP) of 0.894. Furthermore, the sensitivity of our integrated approach was 92.4%. This enhances breast cancer screening and diagnosis, with the results highlighting its potential to improve patient outcomes through timely diagnosis and effective treatment planning.

PMID:40715861 | DOI:10.1007/s10278-025-01613-4

J Mol Model. 2025 Jul 26;31(8):218. doi: 10.1007/s00894-025-06440-6.

ABSTRACT

CONTEXT: The unregulated use of anionic surfactants poses significant environmental risks, necessitating methods for their rapid and accurate identification. While fluorescence spectroscopy is a powerful tool, its application faces a critical challenge: existing analytical strategies either rely on complex and costly sensor arrays to acquire rich data, or they apply traditional machine learning to simpler, single-spectrum data, which often requires pre-processing steps like PCA that risk information loss. Furthermore, standard deep learning approaches are often unsuitable due to the high cost and effort required to acquire the large datasets they need for training. To address this gap, we propose an end-to-end, few-shot learning method (CNN-PN) for the classification of anionic surfactant fluorescence emission spectra. Our approach leverages a one-dimensional convolutional neural network (1D-CNN) to automatically extract features from the full, raw spectrum, thus avoiding lossy pre-processing. It then employs a prototypical network to perform robust, similarity-based classification, a strategy highly effective for limited sample sizes. We validated our method on our FESS dataset (53 surfactant categories) and a public metal oxides dataset. In our experiments, the CNN-PN method consistently outperformed traditional techniques like LDA, SVM, and KNN. It achieved 76.36% accuracy when trained with only a single sample per class, 95.90% in a multi-sample scenario on our FESS dataset, and 84.86% on the public dataset. This work provides a powerful and data-efficient framework for spectral analysis, facilitating the development of more accessible and rapid fluorescence sensing technologies, particularly for applications where data collection is expensive or constrained.

METHODS: A few-shot learning classification method based on prototypical networks was employed. A one-dimensional convolutional neural network (1D-CNN) was utilized to extract spectral features from the full fluorescence emission spectra. Classification was then performed within the prototypical network framework using Euclidean distance as the similarity metric between features in the learned latent space. The Python programming language and the PyTorch library were used for all model implementations and data analysis.

PMID:40715840 | DOI:10.1007/s00894-025-06440-6

Evid Based Dent. 2025 Jul 26. doi: 10.1038/s41432-025-01181-0. Online ahead of print.

ABSTRACT

A COMMENTARY ON: Rokhshad R, Banakar M, Shobeiri, P, Zhang P. Artificial intelligence in early childhood caries detection and prediction: a systematic review and meta-analysis. Pediatr Dent. 2024;46:385-394.

DATA SOURCES: A literature search was performed in May 2024 via PubMed, Scopus, Embase, Web of Science, Institute of Electrical and Electronics Engineer database sources, and across the grey literature. Further studies were identified after analysis of reference lists. The research question was defined using the population-intervention-comparison-outcome (PICO) framework.

STUDY SELECTION: Studies published between 2010 and 2024 were included, that used artificial intelligence (AI) algorithms including machine learning (ML), deep learning (DL) and neutral networks (NN) for detecting and predicting early childhood caries (ECC). Exclusion occurred where the full text was inaccessible and non-English papers. Two independent reviewers screened titles and abstracts, with the use of a third reviewer in the case of any disagreement. The process was then repeated with the full texts to assess eligibility, again with a third reviewer where necessary. A total of 21 studies were used in the final analysis following assessment, 7 of which described ECC detection, and 14 for ECC prediction.

DATA EXTRACTION AND SYNTHESIS: The extracted data included author, publication year, study objectives, data modalities, datasets, annotation procedures, follow ups, ML test, AI model architecture, outcome measures and evaluation metrics. The findings were summarised descriptively. Quantitative synthesis was performed on six studies that reported sensitivity and specificity. Summary receiver operator characteristic curves were used to assess discriminatory ability. Statistical analysis was completed.

RESULTS: A total of 21 studies were included in the final analysis. It revealed that AI based methods, especially DL algorithms showed promising results in detecting ECC, with accuracy range of 78-86%, sensitivity of 67-96%, and specificity from 81-99%. ECC prediction had accuracy range of 60-100%, sensitivity of 20-100%, and specificity of 54-94%. The pooled sensitivity and specificity of these studies was 80% and 81% respectively, with confidence intervals of 95%, indicating statistically significant effects.

CONCLUSIONS: AI has demonstrated substantial potential in the detection and prediction of ECC. Further research is required to refine the technology and establish its application in paediatric dentistry.

PMID:40715738 | DOI:10.1038/s41432-025-01181-0

Sci Rep. 2025 Jul 25;15(1):27043. doi: 10.1038/s41598-025-12468-8.

ABSTRACT

Ensuring proper Personal Protective Equipment (PPE) compliance is crucial for maintaining worker safety and reducing accident risks on construction sites. Previous research has explored various object detection methodologies for automated monitoring of non-PPE compliance; however, achieving higher accuracy and computational efficiency remains critical for practical real-time applications. Addressing this challenge, the current study presents an extensive evaluation of You Only Look Once version 10 (YOLOv10)-based object detection models designed specifically to detect essential PPE items such as helmets, masks, vests, gloves, and shoes. The analysis utilized an extensive dataset gathered from multiple sources, including surveillance cameras, body-worn camera footage, and publicly accessible benchmark datasets, ensuring thorough and realistic evaluation conditions. The analysis was conducted using an extensive dataset compiled from multiple sources, including surveillance cameras, body-worn camera footage, and publicly available benchmark datasets, to ensure a thorough evaluation under realistic conditions. Experimental outcomes revealed that the Swin Transformer-based YOLOv10 model delivered the best overall performance, achieving AP50 scores of 92.4% for non-helmet, 88.17% for non-mask, 87.17% for non-vest, 85.36% for non-glove, and 83.48% for non-shoes, with an overall average AP50 of 87.32%. Additionally, these findings underscored the superior performance of transformer-based architectures compared to traditional detection methods across multiple backbone configurations. The paper concludes by discussing the practical implications, potential limitations, and broader applicability of the YOLOv10-based approach, while also highlighting opportunities and directions for future advancements.

PMID:40715598 | DOI:10.1038/s41598-025-12468-8

Sci Rep. 2025 Jul 26;15(1):27296. doi: 10.1038/s41598-025-12848-0.

ABSTRACT

Estimation and fraud detection in the case of insurance claims play a cardinal role in the insurance sector. With accurate estimation of insurance claims, insurers can have good risk perceptions and disburse compensation within proper time, while fraud prevention helps deter massive monetary loss from fraudulent activities. Financial fraud has done significant damage to the global economy, thus threatening the stability and efficiency of capital markets. Deep learning techniques have proven highly effective in addressing these challenges to analyse complex patterns and relationships in extensive datasets. Unlike traditional statistical methods, which often struggle with the intricate nature of insurance claims data, deep learning models performs well in handling diverse variables and factors influencing claim outcomes. To this extent, it explores the deep learning models like VGG 16 & 19, ResNet 50, and a custom 12 & 15-layer Convolutional Neural Network for accurate estimation of insurance claims and detection of fraud. The proposed work enhanced with Enhanced Hippopotamus Optimization Algorithm (EHOA) combined with a custom 12-layer CNN to optimize the hyperparameters and enhance the performance of the model. Overcoming challenges such as local minima and slow convergence, dynamic population adjustment, momentum-based updates, and hybrid fine-tuning are used with the EHOA. The experimental results reveal that the newly proposed EHOA-CNN-12 attains excellent accuracy (92%) and efficiency in comparison to other state-of-the-art approaches in claims estimation and fraud detection tasks.

PMID:40715558 | DOI:10.1038/s41598-025-12848-0

Sci Rep. 2025 Jul 25;15(1):27079. doi: 10.1038/s41598-025-12796-9.

ABSTRACT

One key barrier to applying deep learning (DL) to omics and other biological datasets is data scarcity, particularly when each gene or protein is represented by a single sequence. This fundamental challenge is mainly relevant in research involving genetically constrained organisms, organelles, specialized cell types, and biological cycles and pathways. This study introduces a novel data augmentation strategy designed to facilitate the application of DL models to omics datasets. This approach generated a high number of overlapping subsequences with controlled overlaps and shared nucleotide features through a sliding window technique. A hybrid model of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) layers was applied across augmented datasets comprising genes and proteins from eight microalgae and higher plant chloroplasts. The data augmentation strategy enabled employing DL methods on these datasets and significantly improved the model performance by avoiding common issues such as overfitting and non-representative sequence variations. The current augmentation process is highly adaptable, providing flexibility across different types of biological data repositories. Furthermore, a complementary k-mer-based data augmentation strategy was introduced for unlabeled datasets, enhancing unsupervised analysis. Overall, these innovative strategies provide robust solutions for optimizing model training potential in the study of datasets with limited data availability.

PMID:40715495 | DOI:10.1038/s41598-025-12796-9

Sci Rep. 2025 Jul 25;15(1):27164. doi: 10.1038/s41598-025-12633-z.

ABSTRACT

Type 2 Diabetes Mellitus (T2DM) remains a significant global health challenge, underscoring the need for early and accurate risk prediction tools to enable timely interventions. This study introduces ECG-DiaNet, a multimodal deep learning model that integrates electrocardiogram (ECG) features with established clinical risk factors (CRFs) to improve the prediction of T2DM onset. Using data from the Qatar Biobank (QBB), we compared ECG-DiaNet against unimodal models based solely on ECG or CRFs. A development cohort (n = 2043) was utilized for model training and internal validation, while a separate longitudinal cohort (n = 395) with a median five-year follow-up served as the test set. ECG-DiaNet demonstrated superior predictive performance, achieving a higher area under the receiver operating characteristic curve (AUROC) compared to the CRF-only model (0.845vs.0.8217), which was statistically significant based on the DeLong test (p < 0.001), thus highlighting the added predictive value of incorporating ECG signals. Reclassification metrics reinforced these improvements, with a significant Net Reclassification Improvement (NRI = 0.0153,p < 0.001) and Integrated Discrimination Improvement (IDI = 0.0482,p = 0.0099), confirming the enhanced risk stratification. Furthermore, stratifying participants into Low-, Medium-, and High-risk categories revealed that ECG-DiaNet achieved a higher positive predictive value (PPV) in the high-risk group compared to CRF-only models. These findings, together with the non-invasive nature and wide accessibility of ECG technology, suggest the potential of ECG-DiaNet for clinical implementation. However, further validation using larger and more diverse datasets is needed to improve generalizability.

PMID:40715481 | DOI:10.1038/s41598-025-12633-z

Sci Rep. 2025 Jul 26;15(1):27207. doi: 10.1038/s41598-025-12602-6.

ABSTRACT

An effective diagnosis system and suitable treatment planning require the precise segmentation of thyroid nodules in ultrasound imaging. The advancement of imaging technologies has not resolved traditional imaging challenges, which include noise issues, limited contrast, and dependency on operator choices, thus highlighting the need for automated, reliable solutions. The researchers developed TATHA, an innovative deep learning architecture dedicated to improving thyroid ultrasound image segmentation accuracy. The model is evaluated using the digital database of thyroid ultrasound images, which includes 99 cases across three subsets containing 134 labelled images for training, validation, and testing. It incorporates data pre-treatment procedures that reduce speckle noise and enhance contrast, while edge detection provides high-quality input for segmentation. TATHA outperforms U-Net, PSPNet, and Vision Transformers across various datasets and cross-validation folds, achieving superior Dice scores, accuracy, and AUC results. The distributed thyroid segmentation framework generates reliable predictions by combining results from multiple feature extraction units. The findings confirm that these advancements make TATHA an essential tool for clinicians and researchers in thyroid imaging and clinical applications.

PMID:40715468 | DOI:10.1038/s41598-025-12602-6