YOLOv10-PDTFN: An integrated ROI segmentation and transformer fusion framework for osteoporosis classification in knee radiographs.

INTRODUCTION: Reduced bone mineral density and degeneration of bone microarchitecture are hallmarks of Osteoporosis, a degenerative skeletal condition that raises the risk of fractures. Due to the subjective nature of conventional radiography assessment and the potential for inter-observer variability, early diagnosis and osteoporosis grading remain clinical challenges. Although recent developments in deep learning have demonstrated promise in automating diagnostic tasks, it is still challenging to identify contextual structural patterns and minor trabecular alterations in radiographs. OBJECTIVE: This research suggests a unique deep learning framework for the multi-class classification of Osteoporosis from knee X-ray images in order to overcome these constraints. A Parallel-Dynamic Transformer Fusion Network (PDTFN) is proposed that combines advanced ROI detection with hierarchical spatial-sequence feature fusion for multi-class classification of normal, Osteopenia, and osteoporosis cases. METHODOLOGY: ROI localization is performed using an Improved YOLOv10 integrated with context-aware and region-adaptive modules to capture fine tibiofemoral details. The extracted joint regions are then analyzed by PDTFN, where a Parallel Spatial Transformer Unit (PSTU) models captures fine-grained and global structures, a Dynamic Sequential Convolutional Memory Unit (DSCMU) captures progressive bone density variations, and a Progressive Spatial-Sequential Attention Fusion (PSSAF) module refines discriminative features. Training is optimized using the AdaBelief optimizer, while Eigen-CAM visualizations highlight clinically relevant bone regions. The proposed PDTFN has better sensitivity and specificity than conventional CNN-based classifiers for differentiating between Osteopenia, Osteoporosis, and normal. Both fine trabecular textures and larger structural features are effectively captured by the model when dynamic sequential modeling and adaptive attention mechanisms are used. RESULTS: The model's emphasis on clinically significant tibiofemoral areas is validated by Eigen-CAM representations. The proposed model achieved 98.87% accuracy, 98.52% F1-score, 97.67% kappa, and 98.23% AUC. CONCLUSIONS: With increased accuracy, robustness, and interpretability in osteoporosis evaluation, the suggested approach shows great promise as a computer-aided diagnostic tool to assist radiologists in clinical practice.