BMC Med Imaging. 2025 Jul 15;25(1):284. doi: 10.1186/s12880-025-01810-1.

ABSTRACT

OBJECTIVE: This prospective study aimed to assess the feasibility of a half-Fourier single-shot turbo spin echo sequence (HASTE) with deep learning (DL) reconstruction for ultrafast imaging of the bladder with reduced susceptibility to motion artifacts.

METHODS: 50 patients underwent pelvic T2w imaging at 3 Tesla using the following MR sequences in sagittal orientation without antiperistaltic premedication: T2-TSE (time of acquisition [TA]: 2.03-4.00 min), standard HASTE (TA: 0.65-1.10 min), and DL-HASTE (TA: 0.25-0.47 min), with a slice thickness of 3 mm and a varying number of slices (25-45). Three radiologists evaluated the image quality of the three sequences quantitatively and qualitatively.

RESULTS: Overall image quality of DL-HASTE (average score: 5) was superior to HASTE and T2-TSE (p < .001). DL-HASTE provided the clearest bladder wall delineation, especially in the apical part of the bladder (p < .001). SNR (36.3 ± 6.3) and CNR (50.3 ± 19.7) were the highest on DL-HASTE, followed by T2-TSE (33.1 ± 6.3 and 44.3 ± 21.0, respectively; p < .05) and HASTE (21.7 ± 5.4 and 35.8 ± 17.5, respectively; p < .01). A limitation of DL-HASTE and HASTE was the susceptibility to urine flow artifact within the bladder, which was absent or only minimal on T2-TSE. Diagnostic confidence in assessment of the bladder was highest with the combination of DL-HASTE and T2-TSE (p < .05).

CONCLUSION: DL-HASTE allows for ultrafast imaging of the bladder with high image quality and is a promising addition to T2-TSE.

PMID:40665218 | DOI:10.1186/s12880-025-01810-1

Phys Eng Sci Med. 2025 Jul 15. doi: 10.1007/s13246-025-01595-1. Online ahead of print.

ABSTRACT

Organ segmentation using 18F-FDG PET images alone has not been extensively explored. Segmentation based methods based on deep learning (DL) have traditionally relied on CT or MRI images, which are vulnerable to alignment issues and artifacts. This study aimed to develop a DL approach for segmenting the entire liver based solely on 18F-FDG PET images. We analyzed data from 120 patients who were assessed using 18F-FDG PET. A three-dimensional (3D) U-Net model from nnUNet and preprocessed PET images served as DL and input images, respectively, for the model. The model was trained with 5-fold cross-validation on data from 100 patients, and segmentation accuracy was evaluated on an independent test set of 20 patients. Accuracy was assessed using Intersection over Union (IoU), Dice coefficient, and liver volume. Image quality was evaluated using mean (SUVmean) and maximum (SUVmax) standardized uptake value and signal-to-noise ratio (SNR). The model achieved an average IoU of 0.89 and an average Dice coefficient of 0.94 based on test data from 20 patients, indicating high segmentation accuracy. No significant discrepancies in image quality metrics were identified compared with ground truth. Liver regions were accurately extracted from 18F-FDG PET images which allowed rapid and stable evaluation of liver uptake in individual patients without the need for CT or MRI assessments.

PMID:40665198 | DOI:10.1007/s13246-025-01595-1

Neurosci Bull. 2025 Jul 15. doi: 10.1007/s12264-025-01455-8. Online ahead of print.

ABSTRACT

In modern society, people are increasingly exposed to chronic stress, leading to various mental disorders. However, the activities of brain regions, especially neural firing patterns related to specific behaviors, remain unclear. In this study, we introduce a novel approach, NeuroSync, which integrates open-field behavioral testing with electrophysiological recordings from emotion-related brain regions, specifically the central amygdala and the paraventricular nucleus of the hypothalamus, to explore the mechanisms of negative emotions induced by chronic stress in mice. By applying machine vision techniques, we quantified behaviors in the open field, and signal processing algorithms elucidated the neural underpinnings of the observed behaviors. Synchronizing behavioral and electrophysiological data revealed significant correlations between neural firing patterns and stress-related behaviors, providing insights into real-time brain activity underlying behavioral responses. This research combines deep learning and machine learning to synchronize high-resolution video and electrophysiological data, offering new insights into neural-behavioral dynamics under chronic stress conditions.

PMID:40665179 | DOI:10.1007/s12264-025-01455-8

Sci Rep. 2025 Jul 15;15(1):25468. doi: 10.1038/s41598-025-08024-z.

ABSTRACT

White blood cell (WBC) classification plays a vital role in hematology for diagnosing various medical conditions. However, it faces significant challenges due to domain shifts caused by variations in sample sources (e.g., blood or bone marrow) and differing imaging conditions across hospitals. Traditional deep learning models often suffer from catastrophic forgetting in such dynamic environments, while foundation models, though generally robust, experience performance degradation when the distribution of inference data differs from that of the training data. To address these challenges, we propose a generative replay-based Continual Learning (CL) strategy designed to prevent forgetting in foundation models for WBC classification. Our method employs lightweight generators to mimic past data with a synthetic latent representation to enable privacy-preserving replay. To showcase the effectiveness, we carry out extensive experiments with a total of four datasets with different task ordering and four backbone models including ResNet50, RetCCL, CTransPath, and UNI. Experimental results demonstrate that conventional fine-tuning methods degrade performance on previously learned tasks and struggle with domain shifts. In contrast, our continual learning strategy effectively mitigates catastrophic forgetting, preserving model performance across varying domains. This work presents a practical solution for maintaining reliable WBC classification in real-world clinical settings, where data distributions frequently evolve.

PMID:40665174 | DOI:10.1038/s41598-025-08024-z

Sci Rep. 2025 Jul 15;15(1):25609. doi: 10.1038/s41598-025-02303-5.

ABSTRACT

Accurate travel time prediction (TTP) is essential to freeway users, including drivers, administrators, and freight-related companies, for enabling them to plan trips effectively and mitigate traffic congestion. However, TTP is a complex challenge even for researchers due to the difficulty of capturing numerous and diverse factors such as driver behaviors, rush hours, special events, and traffic incidents, etc. A multitude of studies have proposed methods to address this issue, yet these approaches often involve multiple stages and steps, including data preprocessing, feature selection, data imputation, prediction model. The intricacy of these processes makes it difficult to pinpoint which steps or factors most significantly influence prediction accuracy. In this paper, we investigate the impact of various steps on TTP accuracy by examining existing methods. Beginning with the data pre-processing phase, we evaluate the effect of deep learning, interpolation, and max value imputation techniques on dealing with missing values. We also examine the influence of temporal features and weather conditions on the prediction accuracy. Furthermore, we compare five distinct hybrid models by assessing their strengths and limitations. To ensure our experiments align with real-world situations well, we conduct experiments using datasets from Taiwan and California. The experimental results reveal that the data-preprocessing phase, including feature editing, plays a pivotal role in TTP accuracy. Additionally, base models such as Long Short-Term Memory (LSTM) and eXtreme Gradient Boosting (XGBoost) outperform all hybrid models on real-world datasets. Based on these insights, we propose a baseline that fuses the complementary strengths of XGBoost and LSTM via a gating network. This approach dynamically allocates weights, guided by key statistical features, to each model, enabling the model to robustly adapt to both stable and volatile traffic conditions and achieve superior prediction accuracy compared to existing methods. By breaking down the TTP process and analyzing each component, this study provides insights into the factors which affect prediction accuracy most significantly, thereby offering guidance and foundation for developing more effective TTP methods in the future.

PMID:40665164 | DOI:10.1038/s41598-025-02303-5

Sci Rep. 2025 Jul 15;15(1):25493. doi: 10.1038/s41598-025-08342-2.

ABSTRACT

Birth complications, particularly jaundice, are one of the leading causes of adolescent death and disease all over the globe. The main severity of these illnesses may diminish if scholars study more about their sources and progress toward effective treatment. Assured developments were prepared, but they are inadequate. Newborns repeatedly have jaundice as their primary medical concern. A raised level of bilirubin is a symbol of jaundice. Generally, in newborns, hyperbilirubinemia peaks in the initial post-delivery week. The inability to perceive issues early is sufficient for quick treatment, and the resemblance of indications might lead to misdiagnosis. Therefore, appropriate technologies are instantly required. Nowadays, researchers have begun to implement an image-processing model for analyzing jaundice. Paediatricians can detect and classify neonatal jaundice with machine learning (ML) and deep learning (DL) techniques. This study proposes an Early Diagnosis of Neonatal Jaundice Image Classification Using Kernel Extreme Learning Machine (EDNJIC-KELM) approach in the Healthcare Sector. The main intention of the EDNJIC-KELM approach is to build an effective system for diagnosing neonatal jaundice based on advanced methods. Initially, the image pre-processing stage applies the Wiener filtering (WF) method to improve the quality of an image and make it more suitable for analysis by removing the noise. In addition, the vision transformer (ViT) method is employed for the feature extraction process. Furthermore, the EDNJIC-KELM method employs the kernel extreme learning machine (KELM) method for the jaundice image classification. Finally, the enhanced coati optimization algorithm (ECOA) method is implemented for the hyperparameter tuning of the KELM method, which results in a higher classification process. The experimental analysis of the EDNJIC-KELM technique is examined using the Jaundice Image data. The performance validation of the EDNJIC-KELM technique portrayed a superior accuracy value of 96.97% over existing models.

PMID:40665120 | DOI:10.1038/s41598-025-08342-2

Comput Med Imaging Graph. 2025 Jul 10;124:102602. doi: 10.1016/j.compmedimag.2025.102602. Online ahead of print.

ABSTRACT

Gliomas are the most common malignant primary brain tumors, and grading their severity, particularly the diagnosis of low-grade gliomas, remains a challenging task for clinicians and radiologists. With advancements in deep learning and medical image processing technologies, the development of Clinical Decision Support Systems (CDSS) for glioma grading offers significant benefits for clinical treatment. This study proposes a CDSS for glioma grading, integrating a novel feature extraction framework. The method is based on combining ensemble learning and knowledge distillation: teacher models were constructed through ensemble learning, while uncertainty-weighted ensemble averaging is applied during student model training to refine knowledge transfer. This approach bridges the teacher-student performance gap, enhancing grading accuracy, reliability, and clinical applicability with lightweight deployment. Experimental results show 85.96 % Accuracy (5.2 % improvement over baseline), with Precision (83.90 %), Recall (87.40 %), and F1-score (83.90 %) increasing by 7.5 %, 5.1 %, and 5.1 % respectively. The teacher-student performance gap is reduced to 3.2 %, confirming effectiveness. Furthermore, the developed CDSS not only ensures rapid and accurate glioma grading but also includes critical features influencing the grading results, seamlessly integrating a methodology for generating comprehensive diagnostic reports. Consequently, the glioma grading CDSS represents a practical clinical decision support tool capable of delivering accurate and efficient auxiliary diagnostic decisions for physicians and patients.

PMID:40663835 | DOI:10.1016/j.compmedimag.2025.102602

IEEE Trans Med Imaging. 2025 Jul 15;PP. doi: 10.1109/TMI.2025.3589456. Online ahead of print.

ABSTRACT

Colorectal cancer (CRC) is a significant global health concern, and early detection through screening plays a critical role in reducing mortality. While deep learning models have shown promise in improving polyp detection, classification, and segmentation, their generalization across diverse clinical environments, particularly with out-of-distribution (OOD) data, remains a challenge. Multi-center datasets like PolypGen have been developed to address these issues, but their collection is costly and time-consuming. Traditional data augmentation techniques provide limited variability, failing to capture the complexity of medical images. Diffusion models have emerged as a promising solution for generating synthetic polyp images, but the image generation process in current models mainly relies on segmentation masks as the condition, limiting their ability to capture the full clinical context. To overcome these limitations, we propose a Progressive Spectrum Diffusion Model (PSDM) that integrates diverse clinical annotations-such as segmentation masks, bounding boxes, and colonoscopy reports-by transforming them into compositional prompts. These prompts are organized into coarse and fine components, allowing the model to capture both broad spatial structures and fine details, generating clinically accurate synthetic images. By augmenting training data with PSDM-generated samples, our model significantly improves polyp detection, classification, and segmentation. For instance, on the PolypGen dataset, PSDM increases the F1 score by 2.12% and the mean average precision by 3.09%, demonstrating superior performance in OOD scenarios and enhanced generalization.

PMID:40663685 | DOI:10.1109/TMI.2025.3589456

IEEE Trans Med Imaging. 2025 Jul 15;PP. doi: 10.1109/TMI.2025.3588859. Online ahead of print.

ABSTRACT

Positron Emission Tomography (PET) is a valuable imaging method for studying molecular-level processes in the body, such as hyperphosphorylated tau (p-tau) protein aggregates, a hallmark of several neurodegenerative diseases including Alzheimer's disease. P-tau density and cerebral perfusion can be quantified from dynamic PET images using tracer kinetic modeling techniques. However, noise in PET images leads to uncertainty in the estimated kinetic parameters, which can be quantified by estimating the posterior distribution of kinetic parameters using Bayesian inference (BI). Markov Chain Monte Carlo (MCMC) techniques are commonly used for posterior estimation but with significant computational needs. This work proposes an Improved Denoising Diffusion Probabilistic Model (iDDPM)-based method to estimate the posterior distribution of kinetic parameters in dynamic PET, leveraging the high computational efficiency of deep learning. The performance of the proposed method was evaluated on a [18F]MK6240 study and compared to a Conditional Variational Autoencoder with dual decoder (CVAE-DD)-based method and a Wasserstein GAN with gradient penalty (WGAN-GP)-based method. Posterior distributions inferred from Metropolis-Hasting MCMC were used as reference. Our approach consistently outperformed the CVAE-DD and WGAN-GP methods and offered significant reduction in computation time than the MCMC method (over 230 times faster), inferring accurate (< 0.67% mean error) and precise (< 7.23% standard deviation error) posterior distributions.

PMID:40663684 | DOI:10.1109/TMI.2025.3588859

IEEE Trans Pattern Anal Mach Intell. 2025 Jul 15;PP. doi: 10.1109/TPAMI.2025.3588186. Online ahead of print.

ABSTRACT

Self-attention and transformers have been widely used in deep learning. Recent efforts have been devoted to incorporating transformer blocks into different neural architectures, including those with convolutions, leading to various visual transformers for computer vision tasks. In this paper, we propose a novel and compact transformer block, Transformer with Learnable Token Merging (LTM), or LTM-Transformer. LTM-Transformer performs token merging in a learnable scheme. LTM-Transformer is compatible with many popular and compact transformer networks, and it reduces the FLOPs and the inference time of the visual transformers while maintaining or even improving the prediction accuracy. In the experiments, we replace all the transformer blocks in popular visual transformers, including MobileViT, EfficientViT, ViT, and Swin, with LTM-Transformer blocks, leading to LTM-Transformer networks with different backbones. The LTM-Transformer is motivated by reduction of Information Bottleneck, and a novel and separable variational upper bound for the IB loss is derived. The architecture of mask module in our LTM blocks which generate the token merging mask is designed to reduce the derived upper bound for the IB loss. Extensive results on computer vision tasks evidence that LTM-Transformer renders compact and efficient visual transformers with comparable or much better prediction accuracy than the original visual transformers. The code of the LTM-Transformer is available at https://github.com/Statistical-Deep-Learning/LTM.

PMID:40663671 | DOI:10.1109/TPAMI.2025.3588186

IEEE J Biomed Health Inform. 2025 Jul 15;PP. doi: 10.1109/JBHI.2025.3586049. Online ahead of print.

ABSTRACT

High-quality 3D fetal brain MRI reconstruction from motion-corrupted 2D slices is crucial for precise clinical diagnosis and advancing our understanding of fetal brain development. This necessitates reliable slice-to-volume registration (SVR) for motion correction and super-resolution reconstruction (SRR) techniques. Traditional approaches have their limitations, but deep learning (DL) offers the potential in enhancing SVR and SRR. However, most of DL methods require large-scale external 3D high-resolution (HR) training datasets, which is challenging in clinical fetal MRI. To address this issue, we propose an unsupervised iterative joint SVR and SRR DL framework for 3D isotropic HR volume reconstruction. Specifically, our method conceptualizes SVR as a function that maps a 2D slice and a 3D target volume to a rigid transformation matrix, aligning the slice to the underlying location within the target volume. This function is parameterized by a convolutional neural network, which is trained by minimizing the difference between the volume slicing at the predicted position and the actual input slice. For SRR, a decoding network embedded within a deep image prior framework, coupled with a comprehensive image degradation model, is used to produce the HR volume. The deep image prior framework offers a local consistency prior to guide the reconstruction of HR volumes. By performing a forward degradation model, the HR volume is optimized by minimizing the loss between the predicted slices and the acquired slices. Experiments on both large-magnitude motion-corrupted simulation data and clinical data have shown that our proposed method outperforms current state-of-the-art fetal brain reconstruction methods. The source code is available at https://github.com/DeepBMI/SUFFICIENT.

PMID:40663667 | DOI:10.1109/JBHI.2025.3586049

Brief Bioinform. 2025 Jul 2;26(4):bbaf340. doi: 10.1093/bib/bbaf340.

ABSTRACT

The emergence of deep learning, particularly AlphaFold, has revolutionized static protein structure prediction, marking a transformative milestone in structural biology. However, protein function is not solely determined by static three-dimensional structures but is fundamentally governed by dynamic transitions between multiple conformational states. This shift from static to multi-state representations is crucial for understanding the mechanistic basis of protein function and regulation. This review outlines the fundamental concepts of protein dynamic conformations, surveys recent computational advances in modeling these dynamics in the post-AlphaFold era, and highlights key challenges, including data limitations, methodological constraints, and evaluation metrics. We also discuss potential strategies to address these challenges and explore future research directions to deepen our understanding of protein dynamics and their functional implications. This work aims to provide insights and perspectives to facilitate the ongoing development of protein conformation studies in the era of artificial intelligence-driven structural biology.

PMID:40663654 | DOI:10.1093/bib/bbaf340

Brief Bioinform. 2025 Jul 2;26(4):bbaf338. doi: 10.1093/bib/bbaf338.

ABSTRACT

Accurately predicting inhibitor-kinase binding affinity is crucial in drug discovery and medical applications, especially in the treatment of diseases such as cancer. Existing methods for predicting inhibitor-kinase affinity still face challenges, including insufficient data expression, limited feature extraction, and low performance. Despite the progress made through artificial intelligence methods, particularly deep learning technology, many current methods fail to capture the intricate interactions between kinases and inhibitors. Therefore, it is necessary to develop more advanced methods to solve the existing problems in inhibitor-kinase binding prediction. This study proposed Kinhibit, a novel framework for inhibitor-kinase binding affinity prediction. Kinhibit integrates self-supervised graph contrastive learning with multiview molecular graph representation and structure-informed protein language model (ESM-S) to extract features effectively. Kinhibit also employed a feature fusion approach to optimize the fusion of inhibitor and kinase features. Experimental results demonstrate the superiority of this method, achieving an accuracy of 92.6% in inhibitor prediction tasks of three mitogen-activated protein kinase (MAPK) signalling pathway kinases: Raf protein kinase (RAF), Mitogen-activated protein kinase kinase (MEK), and Extracellular Signal-Regulated Kinase (ERK). Furthermore, the framework achieves an impressive accuracy of 92.9% on the MAPK-All dataset. This study provides promising and effective tools for drug screening and biological sciences.

PMID:40663653 | DOI:10.1093/bib/bbaf338

PLoS One. 2025 Jul 15;20(7):e0326080. doi: 10.1371/journal.pone.0326080. eCollection 2025.

ABSTRACT

In recent times, chronic kidney cancer has been considered a significant cause of cancer, and Renal Cell Carcinoma (RCC) has become a significant prevalent among various kidney cancer conditions. The analysis of kidney cancer, an important and often time-sensitive medical task, has seen a breakthrough alteration by incorporating deep learning (DL) methods, mainly in analyzing histopathological images (HIs). Given manual analysis's inherent complexity and time-consuming aspect, automatic systems leveraging DL methods provide a promising solution. Automated techniques powered by DL methods showcase a notable capability to analyze intricate details within HIs. These methods are adept at recognizing complex patterns and anomalies within HIs, accelerating the diagnostic method and increasing accuracy. The combination of advanced computational methods with the assessment of kidney cancer HIs not only overcomes the demanding requirements for timely identification but also paves the way for more effective and reliable diagnostic processes in renal oncology. This study presents the Kidney Cancer Detection and Classification employing a Snake Optimizer with Deep Learning on Pathological Images (KCDC-SODLPI) technique. The main aim of the KCDC-SODLPI method is to analyze the pathological images to determine the presence of kidney cancer. In the multifaceted process, the KCDC-SODLPI technique utilizes a Gaussian filtering (GF)-based image preprocessing approach to eliminate the noise content. Furthermore, the KCDC-SODLPI method employs the SE-DenseNet model for extracting intricate patterns from the input images. Moreover, the SO model is used to tune the hyperparameter of the SE-DenseNet method. Finally, the bidirectional long shortterm memory (BiLSTM) model is implemented to detect and classify kidney cancer. The performance of the KCDC-SODLPI technique is evaluated under the biomedical image dataset. The experimental validation of the KCDC-SODLPI method portrayed a superior accuracy value of 88.90% over existing models.

PMID:40663560 | DOI:10.1371/journal.pone.0326080

PLoS One. 2025 Jul 15;20(7):e0326993. doi: 10.1371/journal.pone.0326993. eCollection 2025.

ABSTRACT

Multiple C2 Domains and Transmembrane region Proteins (MCTPs) in plants have been identified as important functional and structural components of plasmodesmata cytoplasmic bridges, which are vital for cell-cell communication. MCTPs are endoplasmic reticulum (ER)-associated proteins which contain three to four C2 domains and two transmembrane regions. In this study, we created structural models of Arabidopsis MCTP4 ER-anchor transmembrane region (TMR) domain using several prediction methods based on deep learning (DL). This region, critical for driving ER association, presents a complex domain organization and remains largely unknown. Our study demonstrates that using a single deep-learning method to predict the structure of membrane proteins can be challenging. Our models presented three different conformations for the MCTP4 structure, provided by different deep learning methods, indicating the potential complexity of the protein's conformational landscape. We then used physics-based molecular dynamics simulations to explore the behaviour of the TMR of MCTPs within the lipid bilayer.We found that the TMR of MCTP4 is not rigid but can adopt multiple conformations. The membrane-embedded region contains two helical pairs: HP1 (TM1-TM2) and HP2 (TM3-TM4). Deep learning predictions revealed three distinct types of inter-helical contact interfaces: ESMFold, AlphaFold-Multimer, trRosetta, and RoseTTAFold consistently predicted a TM2-TM3 contact; AlphaFold2 did not predict any contact between these two helical pairs, while OmegaFold instead suggested a TM1-TM4 interface. Our physics-based coarse-grained simulations not only confirmed the contacts predicted by these models but also revealed a broader conformational landscape. In particular, structural clustering identified five distinct conformational clusters, with additional and more extensive inter-helical contacts not captured by the deep learning predictions. These findings underscore the complexity of predicting protein structures. We learned that combining different methods, such as deep learning and simulations, enhances our understanding of complex proteins.

PMID:40663537 | DOI:10.1371/journal.pone.0326993

PLoS One. 2025 Jul 15;20(7):e0327906. doi: 10.1371/journal.pone.0327906. eCollection 2025.

ABSTRACT

In complex geological environments, the morphology, orientation and distribution characteristics of cracks in the rock directly affect the stability assessment for rock masses and engineering safety decisions. However, the traditional manual interpretation method is inefficient and influenced by subjective factors, which makes it tough to fulfill the requirements for high-precision and automated detection. Especially in the rock specimen analysis of prefabricated multi-angle cracks, image quality and algorithm adaptability have emerged as the critical bottlenecks restricting the identification accuracy. For this reason, it is pressingly essential to realize high-precision and automatic identification in the crack tip of the rock. Firstly, in this study, SCB semi-circular disk specimens are exposed to three-point bending loading, which is sandstone with prefabricated cracks at 0°, 15°, 30°, 45° and 60°. The microsecond-level expansion process of multi-directional cracks is monitored by utilizing an ultrafast camera in the rock specimens. Secondly, three equalization methods are applied to the collected crack images of the rock specimens, including HE, AHE, and CLAHE, to enhance the accuracy of identifying cracks in the rock specimens. And the preprocessed crack images of the rock specimens are compared, which reveals the CLAHE method possesses the optimum preprocessing effect. Based on this, pixel-level annotations are performed on the pretreated crack images, and a dataset is established about cracks in the rock specimen at five different angles. The Deeplabv3 network and the U-Net network are adopted to build cracks recognition models of the rock specimen to predict and identify the crack tips on the rock. The final results demonstrate that the recognition accuracy of the U-net model is able to reach up to 99.4%, the precision is capable of amount to 97.3%, and the recall rate can attain to 95.6%, in the cracks identification of the rock sample with various angles. The recognition accuracy, the precision, and the recall rate of the U-net model have increased by 0.5%, 2.3%, and 4.3% respectively compared with the Deeplabv3 model. The research results provide new ideas for the intelligent detection of cracks in the rock mass, which offer high-confidence data support for engineering decisions in complex geological environments.

PMID:40663529 | DOI:10.1371/journal.pone.0327906

Ann Biomed Eng. 2025 Jul 15. doi: 10.1007/s10439-025-03798-9. Online ahead of print.

ABSTRACT

Meniscal extrusion (ME) has been identified as a key factor contributing to knee joint dysfunction and osteoarthritis progression. Traditional finite element analysis (FEA) methods, while accurate, are computationally expensive and time-consuming, limiting their application for real-time clinical assessments and large-scale studies. This study proposes a geometric deep learning (GDL) model to predict the biomechanical responses of knee joint soft tissues, specifically focusing on the effects of varying degrees of meniscal extrusion. The model, trained on finite element analysis (FEA)-derived data and leveraging advanced AI algorithms, significantly reduces computational time while maintaining high prediction accuracy. Validation against FEA results demonstrated that the GDL model reliably predicts stress and displacement distributions, with key performance metrics including Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Percent Error at Peak Location (PEatPEAK), and Percent Error in Peak Value (PEinPEAK). Compared to conventional FEA workflows, the GDL model eliminates time-consuming preprocessing steps, enabling real-time or near-real-time biomechanical assessments. This innovation provides rapid insights into knee joint mechanics, facilitating clinical decision-making, surgical planning, and personalized rehabilitation strategies. The findings underscore the potential of AI-driven approaches to revolutionize biomechanical research and clinical practice, offering scalable and personalized solutions for joint mechanics analysis.

PMID:40663282 | DOI:10.1007/s10439-025-03798-9

Vis Comput Ind Biomed Art. 2025 Jul 15;8(1):17. doi: 10.1186/s42492-025-00197-8.

ABSTRACT

Placental segmentation is critical for the quantitative analysis of prenatal imaging applications. However, segmenting the placenta using magnetic resonance imaging (MRI) and ultrasound is challenging because of variations in fetal position, dynamic placental development, and image quality. Most segmentation methods define regions of interest with different shapes and intensities, encompassing the entire placenta or specific structures. Recently, deep learning has emerged as a key approach that offer high segmentation performance across diverse datasets. This review focuses on the recent advances in deep learning techniques for placental segmentation in medical imaging, specifically MRI and ultrasound modalities, and cover studies from 2019 to 2024. This review synthesizes recent research, expand knowledge in this innovative area, and highlight the potential of deep learning approaches to significantly enhance prenatal diagnostics. These findings emphasize the importance of selecting appropriate imaging modalities and model architectures tailored to specific clinical scenarios. In addition, integrating both MRI and ultrasound can enhance segmentation performance by leveraging complementary information. This review also discusses the challenges associated with the high costs and limited availability of advanced imaging technologies. It provides insights into the current state of placental segmentation techniques and their implications for improving maternal and fetal health outcomes, underscoring the transformative impact of deep learning on prenatal diagnostics.

PMID:40663247 | DOI:10.1186/s42492-025-00197-8

Radiology. 2025 Jul;316(1):e243239. doi: 10.1148/radiol.243239.

ABSTRACT

Background Emphysema holes change longitudinally in various ways, but current CT measurements lack the ability to fully capture these changes beyond measuring the extent of emphysema. Purpose To track emphysema holes longitudinally, group them according to their dynamics, and investigate their relationship with change in forced expiratory volume in 1 second (FEV1), disease progression, and mortality. Materials and Methods In this secondary analysis, data from participants in the Korean Obstructive Lung Disease cohort study from June 2005 to October 2013 who completed baseline and 6-year follow-up CT with identical protocols were evaluated. Emphysema holes were identified and tracked using deep learning-based software and were grouped based on changes in diameter (in 2-mm increments) as increased in diameter (including both new and enlarged preexisting holes), stable, or decreased in diameter. The percentage of hole volume in each group and its relationship with FEV1 decline were analyzed using multiple linear regression, and comparisons were made among the subsets of participants on the basis of emphysema progression or severity. Overall survival according to the volume cutoff of the holes with increased diameter was compared using the log-rank test. Results Among 108 participants (mean age, 63.4 years ± 6.7 [SD]; 104 male), 39 had emphysema progression (based on whether the change in low-attenuation area less than -950 HU [LAA-950] exceeded 3.7%). Enlarged preexisting holes were marginally associated with a greater decline in FEV1 (β = -.25, P = .049). Compared with those without emphysema progression, those with emphysema progression had a significantly greater percentage of hole volume and percentage of holes with increased diameter (7.7% vs 1.9% and 18.3% vs 6.2%, respectively; both P < .001), with most of the volume attributed to new holes. Participants with severe disease or emphysema (FEV1 < 50% or LAA-950 ≥ 14%) had more holes with increased diameter (5.1% vs 2.4% [P = .02] and 6.7% vs 1.2% [P < .001], respectively) and new holes (3.8% vs 1.7% [P = .01] and 4.7% vs 1.1% [P < .001], respectively). Participants with 5% or greater volume of increased-diameter holes had worse overall survival (log-rank P < .001). Conclusion Emphysema hole-tracking results showed that a greater volume of holes that increased in diameter were related to change in FEV1, disease progression, and mortality. © RSNA, 2025 Supplemental material is available for this article. See also the editorial by van Beek in this issue.

PMID:40662972 | DOI:10.1148/radiol.243239

Bioinformatics. 2025 Jul 1;41(Supplement_1):i458-i465. doi: 10.1093/bioinformatics/btaf252.

ABSTRACT

MOTIVATION: Accurately predicting the impact of single-point mutations on protein stability is crucial for elucidating molecular mechanisms underlying diseases in life sciences and advancing protein engineering in biotechnology. With recent advances in deep learning and protein structure prediction, deep learning approaches are expected to surpass existing methods for predicting protein thermostability. However, structure-based deep learning models, specifically convolutional neural networks, are affected by orientation biases, leading to inconsistent predictions with respect to the input protein orientation.

RESULTS: In this study, we present OrgNet, a novel orientation-gnostic deep learning model using 3D convolutional neural networks to predict protein thermostability change upon point mutation. OrgNet encodes protein structures as voxel grids, enabling the model to capture fine-grained, spatially localized atomic features. OrgNet implements spatial transforms to standardize input protein orientations, thus eliminating orientation bias problem. When evaluated on established benchmarks, including Ssym and S669, OrgNet achieves state-of-the-art performance, demonstrating superior accuracy and robust performance compared to existing methods.

AVAILABILITY AND IMPLEMENTATION: OrgNet is available at https://github.com/i-Molecule/OrgNet.

PMID:40662839 | DOI:10.1093/bioinformatics/btaf252