Longitudinal analysis of sequential radiological images is hampered by a fundamental data challenge: how to effectively model a sequence of high-resolution images captured at irregular time intervals. This data structure contains indispensable spatial and temporal cues that current methods fail to fully exploit. Models often compromise by either collapsing spatial information into vectors or applying spatio-temporal models that are computationally inefficient and incompatible with non-uniform time steps. We address this challenge with Time-Aware $\Delta$t-Mamba3D, a novel state-space architecture adapted for longitudinal medical imaging. Our model simultaneously encodes irregular inter-visit intervals and rich spatio-temporal context while remaining computationally efficient. Its core innovation is a continuous-time selective scanning mechanism that explicitly integrates the true time difference between exams into its state transitions. This is complemented by a multi-scale 3D neighborhood fusion module that robustly captures spatio-temporal relationships. In a comprehensive breast cancer risk prediction benchmark using sequential screening mammogram exams, our model shows superior performance, improving the validation c-index by 2-5 percentage points and achieving higher 1-5 year AUC scores compared to established variants of recurrent, transformer, and state-space models. Thanks to its linear complexity, the model can efficiently process long and complex patient screening histories of mammograms, forming a new framework for longitudinal image analysis.

Despite standard rehabilitation protocols, many patients still suffer from limited mobility, delayed union, or even non-union. This underscores the need for personalized rehabilitation protocols. Fracture healing is a dynamic process governed by the interplay of mechanical stimuli and biochemical signalling pathways. This review first summarizes current understanding of the biological and mechanobiological mechanisms that regulate bone repair. It also discusses different simulation models, including the finite element method (FEM), agent-based models (ABM), reaction-diffusion models (RDM), and machine learning (ML), and evaluates their respective strengths. Recent advances in patient-specific models are also reviewed, particularly those integrating CT-derived geometry, bone properties, and musculoskeletal (MSK) loading. These approaches enable individualized predictions of healing and can inform clinical rehabilitation strategies. Finally, the key challenges and future priorities for implementing these technologies in clinical practice are discussed, providing insights to support the development of more precise and patient-specific fracture care.

The diffusion model, a cutting-edge deep generative technique, is gaining traction in biomedical informatics, showcasing promising applications across various domains. This review presents an overview of the working principles, categories, and numerous applications of diffusion models in biomedical research. In medical imaging, these models, through frameworks like Denoising Diffusion Probabilistic Models (DDPMs) and Stochastic Differential Equations (SDE), offer advanced solutions for image generation, reconstruction, segmentation, and denoising. Notably, they’ve been employed in synthesizing 2D/3D medical images, MRI, and PET image reconstruction, and segmentation tasks such as labeled MRI generation. In the realm of structured Electronic Health Records (EHR) data, diffusion models excel in data synthesis, offering innovative approaches in the face of challenges like data privacy and data gaps. Furthermore, these models are proving pivotal in physiological signal domains, such as EEG and ECG, for signal generation and restoration amidst data loss and noise disruptions. Another significant application lies in the design and prediction of small molecules and protein structures. These models unveil profound insights into the vast molecular space, guiding endeavors in drug design, molecular docking, and antibody construction. Despite their potential, there are inherent limitations, emphasizing the need for further research, validation, interdisciplinary collaboration, and robust benchmarking to ensure practical reliability and efficiency. This review seeks to shed light on the profound capabilities and challenges of diffusion models in the rapidly evolving landscape of biomedical research.

Lymph node metastasis (LNM) represents the predominant metastatic pathway and a critical prognostic determinant in gastric adenocarcinoma. Accurate preoperative prediction of LNM status is imperative for optimizing tumor staging, therapeutic decision-making, and prognostic evaluation. This study aims to develop and validate a radiomics model utilizing contrast-enhanced computed tomography (CT) features from both tumor and stomach regions for preoperative assessment of LNM status. A retrospective analysis was performed on 279 patients who underwent radical gastrectomy, randomly divided into a training cohort (<i>n</i> = 195) and a validation cohort (<i>n</i> = 84). Preoperative contrast-enhanced abdominal CT images were collected, and radiomics features were extracted from both the tumor and stomach. After Z-score normalization, feature selection was conducted using inter- and intra-class correlation, univariate analysis, and LASSO regression. Six machine learning algorithms were used to construct radiomics models based on tumor, stomach, and their combination. Each model was trained using five-fold cross-validation, and their performance was assessed using the area under the receiver operating characteristic (ROC) curve (AUC). Univariate analysis and logistic regression were performed to identify significant clinical features, leading to the development of a combined model incorporating both clinical and radiomics features. A total of 12 radiomics features were selected to construct the tumor - stomach wall radiomics model. Among the six algorithms, the LightGBM-based model demonstrated superior performance, achieving an AUC of 0.899 (95% CI, 0.858–0.941) in the training set and 0.740 (95% CI, 0.633–0.847) in the validation set. Anemia and abnormal CA199 levels were identified as independent risk factors for LN metastasis in GC. After integrating clinical and radiomics features, the combined model achieved an AUC of 0.903 (95% CI, 0.863–0.944) in the training set and 0.767 (95% CI, 0.664–0.869) in the validation set. Decision curve analysis demonstrated that the combined model had favorable clinical utility. As a non-invasive preoperative predictive tool, the combined model incorporating tumor and stomach radiomics features along with clinical factors shows promising clinical value in assessing LN status in GC patients.

This study aimed to explore the potential of integrating magnetic resonance imaging (MRI) features with inflammatory markers to predict glioma grading using machine learning models. A total of 179 glioma patients were analyzed. The dataset was randomly split into a training set (75%) and an independent test set (25%) to test hyperparameter. To enhance reliability, five-fold cross-validation was also applied during the training phase. Key MRI features and inflammatory markers, including relative apparent diffusion coefficient (rADC), monocyte count, lymphocyte-to-CRP ratio (LCR) and hemoglobin-albumin-lymphocyte-platelet index (HALP), were extracted and used as inputs for multiple machine learning classifiers. Model performance was assessed using metrics such as area under the curve (AUC), accuracy, and F1 score. The support vector machine model exhibited superior predictive performance, achieving an AUC of 0.92 and an F1 score of 0.91, effectively distinguishing between high-grade and low-grade gliomas. The combination of MRI features and inflammatory markers, analyzed through machine learning models like SVM, provides some clues for refining glioma prognosis and guiding personalized treatment strategies. The online version contains supplementary material available at 10.1186/s12880-025-01946-0.

Despite the rapid development of artificial intelligence (AI)-powered automated segmentation tools for PET/CT imaging, their prognostic value in predicting survival outcomes remains inadequately assessed. Our objective was to explore the prognostic significance of tumor burden quantification derived from PSMA PET/CT using AI for metastatic castration-resistant prostate cancer (mCRPC) patients receiving Lutetium-177 (¹⁷⁷Lu) PSMA therapy. A retrospective cohort of 107 consecutive patients with mCRPC treated with ¹⁷⁷Lu-PSMA therapy were analyzed. Utilizing a deep learning algorithm, PSMA-positive lesions were automatically delineated on baseline 68Ga-PSMA-11 PET/CT scans. Key metrics were derived from the segmented lesions: total tumor volume (PSMA_TV), total tumor load (PSMA_TU = PSMA_TV × SUV_mean), and total tumor quotient (PSMA_TQ = PSMA_TV / SUV_mean). A prognostic nomogram was developed through Cox regression analysis, incorporating LASSO regularization for variable selection. Univariate analysis revealed that higher PSMA_TV (HR 1.26), PSMA_TU (HR 1.18), and PSMA_TQ (HR 1.29) were significantly associated with shorter overall survival (OS). A prognostic nomogram that integrated PSMA_TQ alongside chemotherapy history, hemoglobin levels, alkaline phosphatase, and prostate-specific antigen demonstrated a bootstrap-corrected C-index of 0.71 (95% CI 0.64-0.78). Risk stratification using the nomogram showed significantly prolonged OS in low-risk vs. high-risk groups (median OS 30.9 vs. 7.9 months; HR 0.25, 95% CI 0.13-0.45, P < 0.001). The retrospective design is a study limitation. AI-based volumetric analysis of tumor burden on PSMA PET has prognostic significance for survival in ¹⁷⁷Lu-PSMA-treated mCRPC patients. The nomogram integrating PSMA_TQ with clinical factors might help in personalized risk stratification, facilitating AI-aided therapeutic decision-making.

In the domain of medical image segmentation, models utilizing convolutional neural network (CNN) and Transformer have been t extensively studied and widely implemented. However, the self-attention mechanism in Transformer is incapable of adapting its focus to target structures at varying scales, resulting in discontinuities in segmentation. The objective of this study is to propose a multi-directional dynamic modeling network for medical image segmentation. We propose a Cross-axis Mamba attention (CMA) to capture global info and establish long-range dependencies. It integrates both global context and local details, enhancing segmentation performance. We also introduce an Edge Feature Enhancement Model (EFCN) to improve edge feature detection. We evaluated the method on the ISIC2018 dataset, as well as the CVC-300 and Kvasir-SEG datasets. The dice similarity coefficient and intersection-over-union (IoU) metrics achieved values of 91.12 and 85.07, 90.35 and 83.43, and 94.14 and 89.62, respectively. These results outperform those of advanced models such as VM-Unet and Swin-UMamba. The experimental results indicate that the proposed method has good generalization ability and robustness. It also provides important support for clinical diagnosis and treatment.

Liver fibrosis is a pathological outcome of chronic liver injury and a hallmark of multiple chronic liver diseases. Magnetic resonance elastography (MRE) provides a non-invasive modality for evaluating the severity of liver fibrosis. This study aimed to develop and evaluate deep learning-based segmentation models for the automated assessment of liver fibrosis using MRE images, with a focus on comparing the performance of a conventional U-Net model and a UNet-ResNet50-32 × 4d architecture model. A retrospective analysis was conducted on 319 patients enrolled between January 2018 and December 2020. MRE images were processed and segmented using two U-Net-based models. Model performance was assessed through correlation coefficients, intersection over union (IoU), and additional segmentation metrics. The UNet-ResNet50-32 × 4d model demonstrated strong agreement with ground truth annotations, achieving correlation coefficients of 0.952 in the training phase and 0.943 in the validation phase, along with an Dice score of 85.68%, confirming its high segmentation accuracy. The UNet-ResNet50-32 × 4d model exhibited robust performance and may serve as a reliable tool for the rapid and accurate assessment of liver fibrosis severity. The integration of automated segmentation into MRE analysis has the potential to improve clinical workflows and support timely decision-making in the management of chronic liver disease.

Accurate correspondence matching in coronary angiography images is crucial for reconstructing 3D coronary artery structures, which is essential for precise diagnosis and treatment planning of coronary artery disease (CAD). Traditional matching methods for natural images often fail to generalize to X-ray images due to inherent differences such as lack of texture, lower contrast, and overlapping structures, compounded by insufficient training data. To address these challenges, we propose a novel pipeline that generates realistic paired coronary angiography images using a diffusion model conditioned on 2D projections of 3D reconstructed meshes from Coronary Computed Tomography Angiography (CCTA), providing high-quality synthetic data for training. Additionally, we employ large-scale image foundation models to guide feature aggregation, enhancing correspondence matching accuracy by focusing on semantically relevant regions and keypoints. Our approach demonstrates superior matching performance on synthetic datasets and effectively generalizes to real-world datasets, offering a practical solution for this task. Furthermore, our work investigates the efficacy of different foundation models in correspondence matching, providing novel insights into leveraging advanced image foundation models for medical imaging applications.

Ultrasound localization microscopy (ULM) enables super-resolution imaging of microvascular structures by localizing microbubbles from clutter-filtered ultrafast ultrasound data. However, conventional clutter filtering methods, particularly those based on singular value decomposition, are computationally intensive and thus impractical for real-time applications. In this study, we introduce AF-UNet, a lightweight multi-angle deep learning framework designed to accelerate clutter filtering in ULM. The model processes spatiotemporal slices from rotated 3D in-phase/quadrature data and fuses them to suppress tissue signals and reconstruct microvascular volumes. AF-UNet demonstrates robust performance across diverse anatomical organs, including brain, eye, and kidney, achieving strong generalization with consistently high image fidelity. Systematic analysis reveals optimal angular acquisition settings that enhance fusion performance, with peak improvements observed at 2$^\circ$-3$^\circ$ separations for ocular datasets and slightly larger angles for rat kidney and brain datasets. AF-UNet achieves over 20-fold computational speedup compared to conventional SVD filtering while preserving microvascular details, offering a practical pathway toward real-time, clinically applicable ULM.