Deep Learning-Based Diagnosis and Classification of Femoral Head Necrosis

Osteonecrosis of the Femoral Head (ONFH) is a disabling orthopedic disease, and its early diagnosis and precise staging are crucial for clinical intervention. Although Deep Learning has advanced the medical image analysis of ONFH, existing studies lack a standardized automated diagnostic framework based on the China-Japan Friendship Hospital (CJFH) classification, which emphasizes the biomechanical features of the femoral head through three-column structural analysis. In this study, we propose a novel multi-stage deep learning framework, improved YOLOv12 combined with Mamba-like linear attention module for efficient detection of region of interest in hip MRI; Swin-Unet-based to accurately delineate the acetabulum, normal femoral head, and necrotic region; and DeiT-Small classifier based on knowledge distillation to realize the CJFH classification. The study used 3,937 labeled hip MRI images from 484 patients, and the model achieved 94.7% mAP in lesion detection, and the segmentation experiments showed a mIoU of 83.39% for SwinUnet. By refining the knowledge from the RegNetX-160 teacher network to the lightweight DeiT-Small student network, the system completes the five types of CJFH classification with an overall accuracy of 85.22%, effectively mitigating the risk of overfitting in small-scale medical data. This study provides a decision support tool that balances parametric efficiency and diagnostic performance for ONFH early diagnosis and treatment.