Global Attention Mechanism-Enhanced Deep Learning for Multi-Sequence MRI Diagnosis of Early Femoral Head Osteonecrosis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Renmin Hospital of Wuhan University, Wuhan, China.
- Department of Radiology, Tongji Hospital of Tongji Medical College of Huazhong University of Science and Technology, Wuhan, China.
- State Key Laboratory of Intelligent Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan, China.
- Department of Radiology, First Affiliated Hospital of Shihezi University School of Medicine, Shihezi, China.
- The Russell H. Morgan Department of Radiology and Radiological Science, The Johns Hopkins University School of Medicine, Baltimore, Maryland, USA.
Abstract
Accurate diagnosis of early-stage osteonecrosis of the femoral head (ONFH) remains challenging due to reliance on subjective radiological interpretation. This multicenter retrospective study developed an automated, multi-sequence MRI-based diagnostic model for early ONFH using data from 342 ONFH-affected femoral heads (FHs) from 282 patients and 265 healthy FHs from 233 healthy individuals obtained from two geographically distinct institutions. The dataset comprised three MRI sequences (T1WI, FS-T2WI, and Cor STIR) and 7844 annotated FH instances labeled as normal or abnormal. A You Only Look Once (YOLO) segmentation model was trained on both single-sequence and multi-sequence datasets, and a global attention mechanism (GAM) was integrated to enhance diagnostic performance, yielding the GAM-YOLO model. Model performance was compared with that of radiology residents using Fisher's exact test. The multi-sequence YOLO model achieved superior diagnostic accuracy (sensitivity 93.19%, specificity 94.65%, accuracy 94.04%) compared to single-sequence models. Incorporation of GAM further improved performance (sensitivity 97.00%, specificity 97.52%, accuracy 97.30%; p < 0.05). When applied to entire FH images, the GAM-YOLO model achieved sensitivity, specificity, and accuracy of 99.03%, 98.51%, and 98.82%, respectively, showing higher diagnostic performance than residents (p < 0.01). These findings suggest that the GAM-YOLO model offers a promising and objective tool for early ONFH diagnosis, demonstrating potential for clinical application.