Automated detection of early-stage osteonecrosis of the femoral head in adult using YOLOv10: Multi-institutional validation

Objectives: To develop a deep learning model based on the You Only Look Once version 10 (YOLOv10) for detecting early-stage ONFH in adult using radiographs. Methods: A retrospective database study enrolled patients with ONFH classified as the stage I-II by the Association Research Circulation Osseous (ARCO) staging system based on MRI, and with Kellgren-Lawrence (KL) grade <= 1, as the positive group. In negative group, femoral head exhibited normal or KL grade 1 changes. The model was developed by using internal dataset from one institution between November 2008 and June 2024, with patients were divided into training and internal validation sets in an 8:2 ratio. External test sets were enrolled from two independent institutions between December 2021 and June 2024. Intersection over Union (IoU) was utilized to assess accuracy of bounding box placement and inter-observer consistency. Classification performance was evaluated using the area under the curve (AUC). Results: A total of 2321 patients (mean age, 51 years +/- 14 [SD]; 961 female) with 3970 unilateral hip joint radiographs were evaluated. The model achieved accuracies of 0.91, and 0.89 with IoU scores of 0.95 and 0.96 in two external test sets. The model outperformed the radiologists: for the external test set 1, AUC was 0.93 (95 % CI 0.88-0.97) versus an average AUC of 0.83 among radiologists (range: 0.78-0.88); for the external test set 2, AUC was 0.94 (95 % CI 0.90-0.98) versus an average AUC of 0.79 (range: 0.74-0.85). Conclusions: The YOLOv10 model excelled in detecting early-stage ONFH in adult using radiographs, and outperforming radiologists with varying experience.