YOLOv10-PDTFN framework combines ROI segmentation and transformer fusion for osteoporosis classification on knee radiographs

A deep learning model combining YOLOv10 region-of-interest detection and transformer fusion classified osteoporosis, osteopenia, and normal on knee radiographs with 98.87% accuracy and 98.23% AUC.

• The framework first detects regions of interest (tibiofemoral joints) with an improved YOLOv10 model incorporating context-aware and region-adaptive modules.
• The Parallel-Dynamic Transformer Fusion Network (PDTFN) uses parallel spatial transformer, dynamic sequential convolutional memory, and progressive attention fusion to capture fine trabecular textures and progressive bone density changes.
• Eigen-CAM visualization confirmed that the model focuses on clinically relevant bone regions, supporting its interpretability for radiologist assistance.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.