Sci Rep. 2025 May 20;15(1):17531. doi: 10.1038/s41598-025-97718-5.

ABSTRACT

Breast cancer represents a significant global health challenge, which makes it essential to detect breast cancer early and accurately to improve patient prognosis and reduce mortality rates. However, traditional diagnostic processes relying on manual analysis of medical images are inherently complex and subject to variability between observers, highlighting the urgent need for robust automated breast cancer detection systems. While deep learning has demonstrated potential, many current models struggle with limited accuracy and lack of interpretability. This research introduces the Deep Neural Breast Cancer Detection (DNBCD) model, an explainable AI-based framework that utilizes deep learning methods for classifying breast cancer using histopathological and ultrasound images. The proposed model employs Densenet121 as a foundation, integrating customized Convolutional Neural Network (CNN) layers including GlobalAveragePooling2D, Dense, and Dropout layers along with transfer learning to achieve both high accuracy and interpretability for breast cancer diagnosis. The proposed DNBCD model integrates several preprocessing techniques, including image normalization and resizing, and augmentation techniques to enhance the model's robustness and address class imbalances using class weight. It employs Grad-CAM (Gradient-weighted Class Activation Mapping) to offer visual justifications for its predictions, increasing trust and transparency among healthcare providers. The model was assessed using two benchmark datasets: Breakhis-400x (B-400x) and Breast Ultrasound Images Dataset (BUSI) containing 1820 and 1578 images, respectively. We systematically divided the datasets into training (70%), testing (20%,) and validation (10%) sets, ensuring efficient model training and evaluation obtaining accuracies of 93.97% for B-400x dataset having benign and malignant classes and 89.87% for BUSI dataset having benign, malignant, and normal classes for breast cancer detection. Experimental results demonstrate that the proposed DNBCD model significantly outperforms existing state-of-the-art approaches with potential uses in clinical environments. We also made all the materials publicly accessible for the research community at: https://github.com/romzanalom/XAI-Based-Deep-Neural-Breast-Cancer-Detection .

PMID:40394112 | DOI:10.1038/s41598-025-97718-5

Biomol Biomed. 2025 May 16. doi: 10.17305/bb.2025.12475. Online ahead of print.

ABSTRACT

The preoperative human epidermal growth factor receptor type 2 (HER2) status of breast cancer is typically determined by pathological examination of a core needle biopsy, which influences the efficacy of neoadjuvant chemotherapy (NAC). However, the highly heterogeneous nature of breast cancer and the limitations of needle aspiration biopsy increase the instability of pathological evaluation. The aim of this study was to predict HER2 status in preoperative breast cancer using deep learning (DL) models based on ultrasound (US) and magnetic resonance imaging (MRI). The study included women with invasive breast cancer who underwent US and MRI at our institution between January 2021 and July 2024. US images and dynamic contrast-enhanced T1-weighted MRI images were used to construct DL models (DL-US: the DL model based on US; DL-MRI: the model based on MRI; and DL-MRI&US: the combined model based on both MRI and US). All classifications were based on postoperative pathological evaluation. Receiver operating characteristic analysis and the DeLong test were used to compare the diagnostic performance of the DL models. In the test cohort, DL-US differentiated the HER2 status of breast cancer with an AUC of 0.842 (95% CI: 0.708-0.931), and sensitivity and specificity of 89.5% and 79.3%, respectively. DL-MRI achieved an AUC of 0.800 (95% CI: 0.660-0.902), with sensitivity and specificity of 78.9% and 79.3%, respectively. DL-MRI&US yielded an AUC of 0.898 (95% CI: 0.777-0.967), with sensitivity and specificity of 63.2% and 100.0%, respectively.

PMID:40392960 | DOI:10.17305/bb.2025.12475

PLoS One. 2025 May 20;20(5):e0322711. doi: 10.1371/journal.pone.0322711. eCollection 2025.

ABSTRACT

The fish market is a crucial industry for both domestic economies and the global seafood trade. Accurate fish species classification (FSC) plays a significant role in ensuring sustainability, improving food safety, and optimizing market efficiency. This study introduces automatic FSC using Swin Transformer (ST) through transfer learning (SwinFishNet), which proposes an innovative approach to FSC by leveraging the ST model, a cutting-edge architecture known for its exceptional performance in computer vision tasks. The ST's unique ability to capture both local and global features through its hierarchical structure enhances its effectiveness in complex image classification tasks. The model utilizes three distinct datasets: the 12-class BD-Freshwater-Fish dataset, the 10-class SmallFishBD dataset, and the 20-class FishSpecies dataset, focusing on image processing-based classification. Images were preprocessed by resizing to 224 [Formula: see text] 224 pixels, normalizing, and converting to tensor format for compatibility with deep learning models. Transfer learning was applied using the ST, which was fine-tuned on these datasets and optimized with the AdamW algorithm. The model's performance was evaluated using classification accuracy (CA), F1-score, recall, precision, Matthews correlation coefficient, Cohen's kappa and confusion matrix metrics. The results yielded promising CAs: 0.9847 for BD-Freshwater-Fish, 0.9964 for SmallFishBD, and 0.9932 for the FishSpecies dataset. These results underscore the potential of the SwinFishNet in automating FSC and demonstrate its significant contributions to improving sustainability, market efficiency, and food safety in the seafood industry. This work offers a novel methodology with broad applications in both commercial and research settings, advancing the role of artificial intelligence in the fish market.

PMID:40392913 | DOI:10.1371/journal.pone.0322711

PLoS One. 2025 May 20;20(5):e0323776. doi: 10.1371/journal.pone.0323776. eCollection 2025.

ABSTRACT

To investigate the potential of artificial intelligence (AI) in Cervical Vertebral Maturation (CVM) staging, we developed and compared AI-based qualitative CVM and AI-based quantitative QCVM methods. A dataset of 3,600 lateral cephalometric images from 6 medical centers was divided into training, validation, and testing sets in an 8:1:1 ratio. The QCVM approach categorized images into six stages (QCVM I-IV) based on measurements from 13 cervical vertebral landmarks, while the qualitative method identified six stages (CS1-CS6) through morphological assessment of three cervical vertebrae. Statistical analyses evaluated the methods' performance, including the Pearson correlation coefficient, mean square error (MSE), success detection rate (SDR), precision-recall metrics, and the F1 score. For landmark prediction, our AI model demonstrated remarkable performance, achieving an SDR (error threshold of ≤ 1.0 mm) of 97.14% and with the mean prediction error across thirteen landmarks ranging narrowly from 0.17 to 0.55 mm. Based on the AI-predicted landmarks, the cervical vertebral measurements showed strong agreement with orthodontists, as indicated by a Pearson correlation coefficient of 0.98 and an MSE of 0.004. Besides, the CVM method attained an overall classification accuracy of 71.11%, while the QCVM method showed a higher accuracy of 78.33%. These findings suggest that the AI-based quantitative QCVM method offers superior performance, with higher agreement rates and classification accuracy compared to the AI-based qualitative CVM approach, indicating the fully automated QCVM model could give orthodontists a powerful tool to enhance cervical vertebral maturation staging.

PMID:40392884 | DOI:10.1371/journal.pone.0323776

PLoS One. 2025 May 20;20(5):e0323735. doi: 10.1371/journal.pone.0323735. eCollection 2025.

ABSTRACT

The role of resources and capabilities in shaping and implementing a firm's strategy is paramount. The COVID-19 pandemic underscored the necessity for managers to possess a decision-making model that facilitates the selection of resources and capabilities in a real-time, dynamic, adaptive, and iterative manner. However, the dynamic capabilities framework, which serves as a decision-making model, faces three significant issues when selecting resources and capabilities within complex contexts. These issues, identified as research gaps, include context mismatch, inappropriate treatment, and strategy alignment. These gaps serve as the foundation for decision making models. This study aims to develop a decision-making model for determining banking resources and capabilities. The novelty of this study is encapsulated in the proposed decision-making model for resource and capability determination in complex contexts. Furthermore, this study employed a methodology adapted from the International Society of Pharmacoeconomics and Outcomes Research-Society of Medical Decision Making (ISPOR-SMDM). The research methodology was conducted in ten stages to develop a decision-making model. This study used qualitative methods, a case study strategy, and an abductive approach. The research sample consists of Indonesian State-Owned Banks (SOB). This research culminated in a proposed decision-making model that includes seven managerial decisions: probe, sense, structuring, bundling, building, leverage, and reconfiguring. This model integrates fuzzy preference judgments as inputs, deep learning analytics (predictive analysis) as processes, and success rate predictions as outputs. Theoretically, this research contributes to the enhancement of dynamic capabilities through the complex domains of the cynefin framework. Practically, it offers a decision-making model for the board of directors (BOD) to determine resources and capabilities amid complex environmental changes.

PMID:40392866 | DOI:10.1371/journal.pone.0323735

J Comput Biol. 2025 May 20. doi: 10.1089/cmb.2025.0075. Online ahead of print.

ABSTRACT

It is estimated that approximately 15% of cancers worldwide can be linked to viral infections. The viruses that can cause or increase the risk of cancer include human papillomavirus, hepatitis B and C viruses, Epstein-Barr virus, and human immunodeficiency virus, to name a few. The computational analysis of the massive amounts of tumor DNA data, whose collection is enabled by the advancements in sequencing technologies, has allowed studies of the potential association between cancers and viral pathogens. However, the high diversity of oncoviral families makes reliable detection of viral DNA difficult, and the training of machine learning models that enable such analysis computationally challenging. We introduce XVir, a data pipeline that deploys a transformer-based deep learning architecture to reliably identify viral DNA present in human tumors. XVir is trained on a mix of sequencing reads coming from viral and human genomes, resulting in a model capable of robust detection of potentially mutated viral DNA across a range of experimental settings. Results on semi-experimental data demonstrate that XVir is able to achieve high classification accuracy, generally outperforming state-of-the-art competing methods. In particular, it retains high accuracy even when faced with diverse viral populations while being significantly faster to train than other large deep learning-based classifiers.

PMID:40392695 | DOI:10.1089/cmb.2025.0075

J Med Internet Res. 2025 May 20;27:e71560. doi: 10.2196/71560.

ABSTRACT

BACKGROUND: With the rapid development of digital biomarkers in Parkinson disease (PD) research, it has become increasingly important to explore the current research trends and key areas of focus.

OBJECTIVE: This study aimed to comprehensively evaluate the current status, hot spots, and future trends of global PD biomarker research, and provide a systematic review of deep learning models for freezing of gait (FOG) digital biomarkers.

METHODS: This study used bibliometric analysis based on the Web of Science Core Collection database to conduct a comprehensive analysis of the multidimensional landscape of Parkinson digital biomarkers. After identifying research hot spots, the study also followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines for a scoping review of deep learning models for FOG from 5 databases: Web of Science, PubMed, IEEE Xplore, Embase, and Google Scholar.

RESULTS: A total of 750 studies were included in the bibliometric analysis, and 40 studies were included in the scoping review. The analysis revealed a growing number of related publications, with 3700 researchers contributing. Neurology had the highest average annual participation rate (12.46/19, 66%). The United States contributed the most research (192/1171, 16.4%), with 210 participating institutions, which was the highest among all countries. In the study of deep learning models for FOG, the average accuracy of the models was 0.92, sensitivity was 0.88, specificity was 0.90, and area under the curve was 0.91. In addition, 31 (78%) studies indicated that the best models were primarily convolutional neural networks or convolutional neural networks-based architectures.

CONCLUSIONS: Research on digital biomarkers for PD is currently at a stable stage of development, with widespread global interest from countries, institutions, and researchers. However, challenges remain, including insufficient interdisciplinary and interinstitutional collaboration, as well as a lack of corporate funding for related projects. Current research trends primarily focus on motor-related studies, particularly FOG monitoring. However, deep learning models for FOG still lack external validation and standardized performance reporting. Future research will likely progress toward deeper applications of artificial intelligence, enhanced interinstitutional collaboration, comprehensive analysis of different data types, and the exploration of digital biomarkers for a broader range of Parkinson symptoms.

TRIAL REGISTRATION: Open Science Foundation (OSF Registries) OSF.IO/RG8Y3; https://doi.org/10.17605/OSF.IO/RG8Y3.

PMID:40392578 | DOI:10.2196/71560

Mol Divers. 2025 May 20. doi: 10.1007/s11030-025-11211-9. Online ahead of print.

ABSTRACT

MicroRNAs (miRNAs) are evolutionarily conserved small regulatory elements that are ubiquitous in cells and are found to be abnormally expressed during the onset and progression of several human diseases. miRNAs are increasingly recognized as potential diagnostic and therapeutic targets that could be inhibited by small molecules (SMs). The knowledge of SM-miRNA associations (SMAs) is sparse, mainly because of the dynamic and less predictable 3D structures of miRNAs that restrict the high-throughput screening of SMs. Toward augmenting the costly and laborious experiments determining the SM-miRNA interactions, machine learning (ML) has emerged as a cost-effective and efficient platform. In this article, various aspects associated with the ML-guided predictions of SMAs are thoroughly reviewed. Firstly, a detailed account of the SMA data resources useful for algorithms training is provided, followed by an elaboration of various feature extraction methods and similarity measures utilized on SMs and miRNAs. Subsequent to a summary of the ML algorithms basics and a brief description of the performance measures, an exhaustive census of all the 32 ML-based SMA prediction methods developed so far is outlined. Distinctive features of these methods have been described by classifying them into six broad categories, namely, classical ML, deep learning, matrix factorization, network propagation, graph learning, and ensemble learning methods. Trend analyses are performed to investigate the patterns in ML algorithms usage and performance achievement in SMA prediction. Outlining key principles behind the up-to-date methodologies and comparing their accomplishments, this review offers valuable insights into critical areas for future research in ML-based SMA prediction.

PMID:40392452 | DOI:10.1007/s11030-025-11211-9

J Imaging Inform Med. 2025 May 20. doi: 10.1007/s10278-025-01504-8. Online ahead of print.

ABSTRACT

Deep learning methods provide enormous promise in automating manually intense tasks such as medical image segmentation and provide workflow assistance to clinical experts. Deep neural networks (DNN) require a significant amount of training examples and a variety of expert opinions to capture the nuances and the context, a challenging proposition in oncological studies (H. Wang et al., Nature, vol. 620, no. 7972, pp. 47-60, Aug 2023). Inter-reader variability among clinical experts is a real-world problem that severely impacts the generalization of DNN reproducibility. This study proposes quantifying the variability in DNN performance using expert opinions and exploring strategies to train the network and adapt between expert opinions. We address the inter-reader variability problem in the context of prostate gland segmentation using a well-studied DNN, the 3D U-Net model. Reference data includes magnetic resonance imaging (MRI, T2-weighted) with prostate glandular anatomy annotations from two expert readers (R#1, n = 342 and R#2, n = 204). 3D U-Net was trained and tested with individual expert examples (R#1 and R#2) and had an average Dice coefficient of 0.825 (CI, [0.81 0.84]) and 0.85 (CI, [0.82 0.88]), respectively. Combined training with a representative cohort proportion (R#1, n = 100 and R#2, n = 150) yielded enhanced model reproducibility across readers, achieving an average test Dice coefficient of 0.863 (CI, [0.85 0.87]) for R#1 and 0.869 (CI, [0.87 0.88]) for R#2. We re-evaluated the model performance across the gland volumes (large, small) and found improved performance for large gland size with an average Dice coefficient to be at 0.846 [CI, 0.82 0.87] and 0.872 [CI, 0.86 0.89] for R#1 and R#2, respectively, estimated using fivefold cross-validation. Performance for small gland sizes diminished with average Dice of 0.8 [0.79, 0.82] and 0.8 [0.79, 0.83] for R#1 and R#2, respectively.

PMID:40392414 | DOI:10.1007/s10278-025-01504-8

Int J Surg. 2025 May 20. doi: 10.1097/JS9.0000000000002488. Online ahead of print.

ABSTRACT

BACKGROUND: Glioblastoma (GBM) is a highly aggressive brain tumor with poor prognosis. This study aimed to construct and validate a radiomics-based machine learning model for predicting overall survival (OS) in IDH-wildtype GBM after maximal safe surgical resection using magnetic resonance imaging.

METHODS: A total of 582 patients were retrospectively enrolled, comprising 301 in the training cohort, 128 in the internal validation cohort, and 153 in the external validation cohort. Volumes of interest (VOIs) from contrast-enhanced T1-weighted imaging (CE-T1WI) were segmented into three regions: contrast-enhancing tumor, necrotic non-enhancing core, and peritumoral edema using an ResNet-based segmentation network. A total of 4,227 radiomic features were extracted and filtered using LASSO-Cox regression to identify signatures. The prognostic model was constructed using the Mime prediction framework, categorizing patients into high- and low-risk groups based on the median OS. Model performance was assessed using the concordance index (CI) and Kaplan-Meier survival analysis. Independent prognostic factors were identified through multivariable Cox regression analysis, and a nomogram was developed for individualized risk assessment.

RESULTS: The Step Cox [backward] + RSF model achieved CIs of 0.89, 0.81, and 0.76 in the training, internal and external validation cohorts. Log-rank tests demonstrated significant survival differences between high- and low-risk groups across all cohorts (P < 0.05). Multivariate Cox analysis identified age (HR: 1.022; 95% CI: 0.979, 1.009, P < 0.05), KPS score (HR: 0.970, 95% CI: 0.960, 0.978, P < 0.05), rad-scores of the necrotic non-enhancing core (HR: 8.164; 95% CI: 2.439, 27.331, P < 0.05), and peritumoral edema (HR: 3.748; 95% CI: 1.212, 11.594, P < 0.05) as independent predictors of OS. A nomogram integrating these predictors provided individualized risk assessment.

CONCLUSION: This deep learning segmentation-based radiomics model demonstrated robust performance in predicting OS in GBM after maximal safe surgical resection. By incorporating radiomic signatures and advanced machine learning algorithms, it offers a non-invasive tool for personalized prognostic assessment and supports clinical decision-making.

PMID:40391963 | DOI:10.1097/JS9.0000000000002488

Adv Sci (Weinh). 2025 May 20:e2416161. doi: 10.1002/advs.202416161. Online ahead of print.

ABSTRACT

The clinical benefits of neoadjuvant chemoimmunotherapy (NACI) are demonstrated in patients with bladder cancer (BCa); however, more than half fail to achieve a pathological complete response (pCR). This study utilizes multi-center cohorts of 2322 patients with pathologically diagnosed BCa, collected between January 1, 2014, and December 31, 2023, to explore the correlation between tumor budding (TB) status and NACI response and disease prognosis. A deep learning model is developed to noninvasively evaluate TB status based on CT images. The deep learning model accurately predicts the TB status, with area under the curve values of 0.932 (95% confidence interval: 0.898-0.965) in the training cohort, 0.944 (0.897-0.991) in the internal validation cohort, 0.882 (0.832-0.933) in external validation cohort 1, 0.944 (0.908-0.981) in the external validation cohort 2, and 0.854 (0.739-0.970) in the NACI validation cohort. Patients predicted to have a high TB status exhibit a worse prognosis (p < 0.05) and a lower pCR rate of 25.9% (7/20) than those predicted to have a low TB status (pCR rate: 73.9% [17/23]; p < 0.001). Hence, this model may be a reliable, noninvasive tool for predicting TB status, aiding clinicians in prognosis assessment and NACI strategy formulation.

PMID:40391846 | DOI:10.1002/advs.202416161

Wiley Interdiscip Rev Nanomed Nanobiotechnol. 2025 May-Jun;17(3):e70015. doi: 10.1002/wnan.70015.

ABSTRACT

Surface-enhanced Raman scattering (SERS) is a transformative technique for molecular identification, offering exceptional sensitivity, signal specificity, and resistance to photobleaching, making it invaluable for disease diagnosis, monitoring, and spectroscopy-guided surgeries. Unlike traditional Raman spectroscopy, which relies on weak scattering signals, SERS amplifies Raman signals using plasmonic nanoparticles, enabling highly sensitive molecular detection. This technological advancement has led to the development of SERS nanotags with remarkable multiplexing capabilities for biosensing applications. Recent progress has expanded the use of SERS nanotags in bioimaging, theranostics, and more recently, liquid biopsy. The distinction between SERS and conventional Raman spectroscopy is highlighted, followed by an exploration of the molecular assembly of SERS nanotags. Significant progress in bioimaging is summarized, including in vitro studies on 2D/3D cell cultures, ex vivo tissue imaging, in vivo diagnostics, spectroscopic-guided surgery for tumor margin delineation, and liquid biopsy tools for detecting cancer and SARS-CoV-2. A particular focus is the integration of machine learning (ML) and deep learning algorithms to boost SERS nanotag efficacy in liquid biopsies. Finally, it addresses the challenges in the clinical translation of SERS nanotags and offers strategies to overcome these obstacles.

PMID:40391396 | DOI:10.1002/wnan.70015

Stat Comput. 2025;35(4):97. doi: 10.1007/s11222-025-10624-8. Epub 2025 May 17.

ABSTRACT

Individual treatment effect estimation has gained significant attention in recent data science literature. This work introduces the Double Neural Network (Double-NN) method to address this problem within the framework of extended fiducial inference (EFI). In the proposed method, deep neural networks are used to model the treatment and control effect functions, while an additional neural network is employed to estimate their parameters. The universal approximation capability of deep neural networks ensures the broad applicability of this method. Numerical results highlight the superior performance of the proposed Double-NN method compared to the conformal quantile regression (CQR) method in individual treatment effect estimation. From the perspective of statistical inference, this work advances the theory and methodology for statistical inference of large models. Specifically, it is theoretically proven that the proposed method permits the model size to increase with the sample size n at a rate of O ( n ζ ) for some 0 ≤ ζ < 1 , while still maintaining proper quantification of uncertainty in the model parameters. This result marks a significant improvement compared to the range 0 ≤ ζ < 1 2 required by the classical central limit theorem. Furthermore, this work provides a rigorous framework for quantifying the uncertainty of deep neural networks under the neural scaling law, representing a substantial contribution to the statistical understanding of large-scale neural network models.

SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s11222-025-10624-8.

PMID:40391382 | PMC:PMC12085359 | DOI:10.1007/s11222-025-10624-8

NAR Genom Bioinform. 2025 May 19;7(2):lqaf058. doi: 10.1093/nargab/lqaf058. eCollection 2025 Jun.

ABSTRACT

Regulation of DNA or RNA at the transcriptional, post-transcriptional, and translational levels are key steps in the central dogma of molecular biology. DNA-binding proteins (DBPs) and RNA-binding proteins (RBPs) play pivotal roles in the precise regulation of gene expression in these steps. Both of these two classes of proteins are nucleic acid-binding proteins (NABPs), so they exhibit significant similarity in both sequence and structure. However, traditional methods for identifying NABPs are typically time-consuming, costly, and challenging to scale up. Utilizing deep learning to classify proteins intelligently has emerged as a more efficient solution for these issues. In this study, we propose a phased classification method integrating ESM-2 with a dual-path neural network, called DRBP-EDP. Additionally, a refined approach to dataset construction is designed, resulting in the creation of high-quality protein classification datasets. The results demonstrated that the model achieved strong performance, with 90.03% accuracy in the first stage for classifying NABPs and non-nucleic acid-binding proteins, and 89.56% accuracy in the second stage for classifying DBPs and RBPs. To enhance accessibility and usability, DRBP-EDP has been developed in both executable and web-based versions, which are publicly available at https://doi.org/10.5281/zenodo.14092184 and https://github.com/MuQiang-MQ/DRBP-EDP.

PMID:40391089 | PMC:PMC12086546 | DOI:10.1093/nargab/lqaf058

Front Endocrinol (Lausanne). 2025 May 5;16:1506729. doi: 10.3389/fendo.2025.1506729. eCollection 2025.

ABSTRACT

INTRODUCTION: In recent years, artificial intelligence (AI) tools have become widely studied for thyroid ultrasonography (USG) classification. The real-world applicability of these developed tools as pre-operative diagnostic aids is limited due to model overfitting, clinician trust, and a lack of gold standard surgical histology as ground truth class label. The ongoing dilemma within clinical thyroidology is surgical decision making for indeterminate thyroid nodules (ITN). Genomic sequencing classifiers (GSC) have been utilised for this purpose; however, costs and availability preclude universal adoption creating an inequity gap. We conducted this review to analyse the current evidence of AI in ITN diagnosis without the use of GSC.

METHODS: English language articles evaluating the diagnostic accuracy of AI for ITNs were identified. A systematic search of PubMed, Google Scholar, and Scopus from inception to 18 February 2025 was performed using comprehensive search strategies incorporating MeSH headings and keywords relating to AI, indeterminate thyroid nodules, and pre-operative diagnosis. This systematic review and meta-analysis was conducted in accordance with methods recommended by the Cochrane Collaboration (PROSPERO ID CRD42023438011).

RESULTS: The search strategy yielded 134 records after the removal of duplicates. A total of 20 models were presented in the seven studies included, five of which were radiological driven, one utilised natural language processing, and one focused on cytology. The pooled meta-analysis incorporated 16 area under the curve (AUC) results derived from 15 models across three studies yielding a combined estimate of 0.82 (95% CI: 0.81-0.84) indicating moderate-to-good classification performance across machine learning (ML) and deep learning (DL) architectures. However, substantial heterogeneity was observed, particularly among DL models (I² = 99.7%, pooled AUC = 0.85, 95% CI: 0.85-0.86). Minimal heterogeneity was observed among ML models (I² = 0.7%), with a pooled AUC of 0.75 (95% CI: 0.70-0.81). Meta-regression analysis performed suggests potential publication bias or systematic differences in model architectures, dataset composition, and validation methodologies.

CONCLUSION: This review demonstrated the burgeoning potential of AI to be of clinical value in surgical decision making for ITNs; however, study-developed models were unsuitable for clinical implementation based on performance alone at their current states or lacked robust independent external validation. There is substantial capacity for further development in this field.

SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/, identifier CRD42023438011.

PMID:40391010 | PMC:PMC12086071 | DOI:10.3389/fendo.2025.1506729

Environ Sci Ecotechnol. 2025 Apr 24;25:100562. doi: 10.1016/j.ese.2025.100562. eCollection 2025 May.

ABSTRACT

Rainfall measurement at high quality and spatiotemporal resolution is essential for urban hydrological modeling and effective stormwater management. However, traditional rainfall measurement methods face limitations regarding spatial coverage, temporal resolution, and data accessibility, particularly in urban settings. Here, we show a novel rainfall estimation framework that leverages surveillance cameras to enhance estimation accuracy and spatiotemporal data coverage. Our hybrid approach consists of two complementary modules: the first employs an image-quality signature technique to detect rain streaks from video frames and selects optimal regions of interest (ROIs). The second module integrates depthwise separable convolution (DSC) layers with gated recurrent units (GRU) in a regression model to accurately estimate rainfall intensity using these ROIs. We evaluate the framework using video data from two locations with distinct rainfall patterns and environmental conditions. The DSC-GRU model achieves high predictive performance, with coefficient of determination (R2) values ranging from 0.89 to 0.93 when validated against rain gauge measurements. Remarkably, the model maintains strong performance during daytime and nighttime conditions, outperforming existing video-based rainfall estimation methods and demonstrating robust adaptability across variable environmental scenarios. The model's lightweight architecture facilitates efficient training and deployment, enabling practical real-time urban rainfall monitoring. This work represents a substantial advancement in rainfall estimation technology, significantly reducing estimation errors and expanding measurement coverage, and provides a practical, low-cost solution for urban hydrological monitoring.

PMID:40390707 | PMC:PMC12088784 | DOI:10.1016/j.ese.2025.100562

Zhen Ci Yan Jiu. 2025 May 25;50(5):531-537. doi: 10.13702/j.1000-0607.20250319.

ABSTRACT

The quantitative research on acupuncture manipulation techniques aims to transform traditional empirical operations into measurable parameters through interdisciplinary technologies. This paper comprehensively reviewed the biomechanical foundations and technological evolution of acupuncture manipulation quantification research. It delved into upgrades in biomechanical parameter collection, breakthroughs in analytical methods, and innovations and applications from computer vision, deep learning, and multimodal perception fusion. Looking ahead, the paper explored deepening multimodal fusion, construction of big data expert databases, and intelligent clinical applications. It proposed that through multidisciplinary technological integration, a digital acupuncture theory system with characteristics of traditional Chinese medicine can be established. These breakthroughs will promote the transformation of acupuncture from empirical practice to data-driven precision medicine, providing theoretical and technological foundations for modernization and international standardization of traditional Chinese medicine.

PMID:40390611 | DOI:10.13702/j.1000-0607.20250319

Adv Sci (Weinh). 2025 May 20:e2417635. doi: 10.1002/advs.202417635. Online ahead of print.

ABSTRACT

On-chip training in analog in-memory computing (AIMC) holds great promise for reducing data latency and enabling user-specific learning. However, analog synaptic devices face significant challenges, particularly during parallel weight updates in crossbar arrays, where non-uniform programming and disturbances often arise. Despite their importance, the disturbances that occur during training are difficult to quantify based on a clear mechanism, and as a result, their impact on training performance remains underexplored. This work precisely identifies and quantifies the disturbance effects in 6T1C synaptic devices based on oxide semiconductors and capacitors, whose endurance and variation have been validated but encounter worsening disturbance effects with device scaling. By clarifying the disturbance mechanism, three simple operational schemes are proposed to mitigate these effects, with their efficacy validated through device array measurements. Furthermore, to evaluate learning feasibility in large-scale arrays, real-time disturbance-aware training simulations are conducted by mapping synaptic arrays to convolutional neural networks for the CIFAR-10 dataset. A software-equivalent accuracy is achieved even under intensified disturbances, using a cell capacitor size of 50fF, comparable to dynamic random-access memory. Combined with the inherent advantages of endurance and variation, this approach offers a practical solution for hardware-based deep learning based on the 6T1C synaptic array.

PMID:40390534 | DOI:10.1002/advs.202417635

NMR Biomed. 2025 Jul;38(7):e70066. doi: 10.1002/nbm.70066.

ABSTRACT

Translating quantitative skeletal muscle MRI biomarkers into clinics requires efficient automatic segmentation methods. The purpose of this work is to investigate a simple yet effective iterative methodology for building a high-quality automatic segmentation model while minimizing the manual annotation effort. We used a retrospective database of quantitative MRI examinations (n = 70) of healthy and pathological thighs for training a nnU-Net segmentation model. Healthy volunteers and patients with various neuromuscular diseases, broadly categorized as dystrophic, inflammatory, neurogenic, and unlabeled NMDs. We designed an iterative procedure, progressively adding cases to the training set and using a simple visual five-level rating scale to judge the validity of generated segmentations for clinical use. On an independent test set (n = 20), we assessed the quality of the segmentation in 13 individual thigh muscles using standard segmentation metrics-dice coefficient (DICE) and 95% Hausdorff distance (HD95)-and quantitative biomarkers-cross-sectional area (CSA), fat fraction (FF), and water-T1/T2. We obtained high-quality segmentations (DICE = 0.88 ± 0.15/0.86 ± 0.14, HD95 = 6.35 ± 12.33/6.74 ± 11.57 mm), comparable to recent works, although with a smaller training set (n = 30). Inter-rater agreement on the five-level scale was fair to moderate but showed progressive improvement of the segmentation model along with the iterations. We observed limited differences from manually delineated segmentations on the quantitative outcomes (MAD: CSA = 65.2 mm2, FF = 1%, water-T1 = 8.4 ms, water-T2 = 0.35 ms), with variability comparable to manual delineations.

PMID:40390325 | DOI:10.1002/nbm.70066

Respir Res. 2025 May 19;26(1):192. doi: 10.1186/s12931-025-03270-1.

ABSTRACT

The standard approach to diagnosing idiopathic pulmonary fibrosis (IPF) includes identifying the usual interstitial pneumonia (UIP) pattern via high resolution computed tomography (HRCT) or lung biopsy and excluding known causes of interstitial lung disease (ILD). However, limitations of manual interpretation of lung imaging, along with other reasons such as lack of relevant knowledge and non-specific symptoms have hindered the timely diagnosis of IPF. This review proposes the definition of early IPF, emphasizes the diagnostic urgency of early IPF, and highlights current diagnostic strategies and future prospects for early IPF. The integration of artificial intelligence (AI), specifically machine learning (ML) and deep learning (DL), is revolutionizing the diagnostic procedure of early IPF by standardizing and accelerating the interpretation of thoracic images. Innovative bronchoscopic techniques such as transbronchial lung cryobiopsy (TBLC), genomic classifier, and endobronchial optical coherence tomography (EB-OCT) provide less invasive diagnostic alternatives. In addition, chest auscultation, serum biomarkers, and susceptibility genes are pivotal for the indication of early diagnosis. Ongoing research is essential for refining diagnostic methods and treatment strategies for early IPF.

PMID:40390073 | DOI:10.1186/s12931-025-03270-1