Leveraging AI models for lesion detection in osteonecrosis of the femoral head and T1-weighted MRI generation from radiographs.

Abstract

This study emphasizes the importance of early detection of osteonecrosis of the femoral head (ONFH) in young patients on long-term glucocorticoid therapy, including those with acute lymphoblastic leukemia, lupus, and other diagnoses. While X-ray and magnetic resonance imaging (MRI) are standard imaging methods for staging ONFH, MRI can be costly and time-consuming. The research focuses on utilizing artificial intelligence (AI) to enhance the evaluation of radiographic images for ONFH detection. The study involved analyzing X-ray and MRI from 102 control hips and 104 ONFH-affected hips at Association Research Circulation Osseous (ARCO) Stage II and IIIa. We employed transfer learning with the YOLOv8 model for object detection, using 80% of the data for training and 20% for validation, then assessed detection accuracy through mean average precision (mAP) and a precision-recall curve. Additionally, AI generated synthetic MRI (sMRI) from X-ray images using a Generative Adversarial Network (GAN) and evaluated their similarity to original MRI. Results showed that the mAP for ONFH detection was 0.923 for the YOLOv8n model and 0.951 for YOLOv8x. The GAN-generated sMRI exhibited lower image quality compared with originals but maintained potential for lesion assessment. Intrarater reliability among evaluators was high. The findings indicate that AI techniques, particularly YOLOv8 for object detection and GAN for image generation, can effectively assist in ONFH screening, despite some limitations in the generated MRI quality.