Automatic diagnosis of osteonecrosis of the femoral head and segmentation of necrotic regions using deep learning at computed tomography: a real-world multicenter study.

BACKGROUND

Computed tomography (CT) is essential for diagnosing osteonecrosis of the femoral head (ONFH) due to its superior visualization of osseous structures, yet CT-based deep learning (DL) models remain underexplored in real-world settings.

METHODS

This retrospective multicenter study developed a two-stage DL model integrating Vision Transformer (ViT) and Swin Transformer (SwinT) for ONFH diagnosis and necrotic segmentation, using coronal CT images from 2,312 patients (53,828 images) across four hospitals from February 2023 to June 2024. The data from three hospitals were split 7:1:2 for training, validation, and internal testing, with the fourth serving as the external test set. Diagnostic performance was evaluated at both the slice and patient levels using the area under the receiver operating characteristic curve (AUC), sensitivity, and negative predictive value (NPV) while segmentation performance was assessed using the Dice coefficient and HD95. Human-computer comparison trials were conducted against three orthopaedic surgeons.

RESULTS

At the slice level, the model achieved robust performance with strong results in internal test set (AUC = 0.98, sensitivity = 98.91%, NPV = 98.66%, Dice = 72.64%, HD95 = 4.51 mm) and external test set (AUC = 0.97, sensitivity = 99.61%, NPV = 99.63%, Dice = 69.11%, HD95 = 5.27 mm). At the patient level, diagnostic performance remained high, with AUCs of 0.98 and 0.97 in the internal and external test sets, respectively. It outperformed surgeons in diagnosis and matched a 10-year-experienced attending in segmentation.

CONCLUSION

Our CT-based two-stage DL model exhibits high accuracy and strong generalizability for ONFH diagnosis and necrotic segmentation, providing a technically robust foundational framework with promising clinical translational potential for ONFH management.

TRIAL REGISTRATION: This study was registered in the Chinese Clinical Trial Registry chictr.org.cn (Registration ID: ChiCTR2300067945, 01/02/2023).