Bone age assessment (BAA) and fracture detection are essential tasks in orthopedic radiography, since they aid diagnosis and clinical decision-making in trauma treatment. Recent technologies in artificial intelligence (AI), particularly deep learning (DL), have enhanced the ability to analyze radiological images with greater accuracy and consistency. However, current studies are incomplete, with inconsistent evaluation methodologies, limited external validation, and insufficient interpretability, limiting clinical implementation. This study provides a systematic PRISMA-based review of 116 studies (2017–2025) that analyze imaging modalities, datasets, model architectures, evaluation techniques, and clinical relevance. A uniform taxonomy is introduced to define techniques based on task, modality, and learning paradigm, along with a comparative methodology to address inconsistencies across studies. The findings indicate that CNN and transformer-based models achieve MAE as low as 0.4–0.7 years and accuracy exceeding 95
Medical image classification benchmarks inform model selection. We assessed how data splits, training seeds, and evaluation conditions affect selection stability and whether repeated evaluation budgets support reproducible rankings. The original study comprised 800 runs on OrganAMNIST and SIPaKMeD, using four ImageNet-initialized convolutional neural networks (CNNs), ten splits, five seeds, and two checkpoint selection and stopping procedures. The 300-run extension evaluated the International Skin Imaging Collaboration (ISIC) 2019 challenge dataset and MURA (musculoskeletal radiographs) using four CNNs, ten splits, and three seeds under procedure A; paired Swin Transformer-Tiny (Swin-T) analyses used five splits and three seeds. Balanced accuracy (BA) defined rankings. Paired resampling quantified selection frequencies and uncertainty; budget analyses compared discovery subsets with held-out splits. Single-context agreement with the full repeated evaluation reference was 44–62
Ophthalmic vascular segmentation supports quantitative analysis of the retinal vasculature, yet thin vessels, low contrast, and complex vascular topology remain challenging. We propose EDGE-Net (Enhanced Distribution Gated Eye-vessel Network), which combines the CLARITY preprocessing pipeline, a Dynamic Dual-Stream Attention Block (DDSAB), Multi-Stream Gated Feature Fusion (MSGFF), and topology-aware loss terms based on persistent homology and connectivity regularization. The framework is evaluated on three retinal vessel datasets (DRIVE, CHASE_DB1, and STARE) and two coronary angiography datasets (DCA1 and CHUAC); all hyperparameters are selected on a DRIVE validation split and then fixed for every dataset. EDGE-Net obtains the highest F1 score on all five datasets, ranging from 0.8394 to 0.8456 on the retinal datasets and from 0.7421 to 0.7518 on the coronary datasets, although it does not achieve the highest accuracy or specificity on every dataset. When the same dual-threshold iterative (DTI) post-processing is applied to every method on DRIVE, EDGE-Net obtains the lowest Vessel Connectivity Assessment score (3.31 versus 3.58 for the strongest competitor) and the highest Junction Preservation Rate (0.842 versus 0.812); both differences are significant after Holm correction (paired Wilcoxon signed-rank test, p<0.01 ). A controlled ablation attributes most of the pixel-level gain to DDSAB and MSGFF and most of the junction-preservation gain to the topology loss. Potential downstream uses include anatomically plausible vascular-network analysis; direct clinical benefit is not evaluated in this study.
Cardiovascular diseases continue to be the primary cause of mortality worldwide; however, precise diagnosis is often limited by dependence on individual data modalities and the unclear nature of traditional deep learning models. In this study, we introduce an explainable multimodal deep learning framework that simultaneously analyzes cardiac magnetic resonance imaging (MRI) and structured clinical variables through a bidirectional cross-modal attention mechanism that operates on token sequences. The architecture utilizes modality-specific encoders ResNet-50 for imaging and a multilayer perceptron for clinical data interconnected through attention that dynamically aligns complementary information, reducing feature dilution and modal dominance. A multitask objective that includes classification, contrastive alignment (InfoNCE), and regularization further reinforces semantic consistency across modalities. Integrated post hoc explainability modules, such as Gradient-weighted Class Activation Mapping for spatial localization and SHapley Additive exPlanations for clinical feature attribution, offer transparent decision support. Tested on a cohort of 12,847 patients using fivefold cross-validation, the framework achieved an accuracy of 94.2
Coronary angiography (CAG) is the clinical gold standard for assessing coronary artery stenosis. However, its two-dimensional projection nature often causes vessel overlap and limited contrast, making subtle or complex stenotic regions difficult to detect and localize, thereby increasing reliance on experienced cardiologists. Existing deep learning methods face a practical trade-off: CNN-based models struggle with low-contrast or small-scale lesions, while Transformer-based models incur high computational cost, limiting real-time applicability. To address these challenges, we propose Coronary Angiography Structure-aware Stenosis Transformer (CASSFormer), a lightweight CNN-Transformer hybrid network for real-time coronary artery stenosis detection and localization. Specifically, CASSFormer integrates three key components: (i) a multi-stage spatial feature downsampling module (MSFD) to preserve fine-grained vascular textures during early feature extraction; (ii) a multi-scale mixed-topology block (MMTB) to model elongated vessel structures and multi-scale stenotic patterns; (iii) a dual-attention cross-scale adaptive fusion module (DCAF) that jointly exploits local directional attention and global channel attention to enhance sensitivity to low-contrast and small-scale stenotic regions. Experiments on a large private clinical dataset and two public datasets (ARCADE, ADSD) show that CASSFormer achieves favorable detection and localization performance compared with mainstream detection models, while maintaining low computational overhead (27.1 GFLOPs, 11.9M parameters) and an average inference time of 8.0 ms per image. Visualization results further suggest its robustness under challenging angiographic conditions, including vessel overlap, distal small vessels, and severe stenosis or vessel interruption associated with myocardial infarction. These results support CASSFormer as an efficient assistive tool for stenosis detection and localization in complex CAG scenarios.
