Trochanteric region fracture classification network (TRFC-net): a deep learning solution to increasing inexperienced residents' trochanteric region fracture classification accuracy on X-rays.

Background: Accurate classification of trochanteric region fractures on X-rays is crucial for optimal clinical management but poses certain difficulties for inexperienced physicians, with interobserver variability in fracture subtype identification being particularly challenging. Therefore, we aimed to construct a trochanteric region fracture classification network (TRFC-net) based on deep learning algorithms to provide auxiliary diagnostic support for clinicians. Methods: The TRFC-net was designed to include three task modules for the proximal femur detection module, a fracture image screening module to distinguish fractured from nonfractured proximal femur images, and a fracture classification module to categorize A1, A2, and A3 fracture subtypes. The RetinaNet object detection algorithm detects the proximal femur region on X-ray and crops it. Subsequently, the Swin transformer classification network screens the trochanteric region fracture images from the cropped images and subsequently classifies the fracture image types. Gradient-weighted class activation mapping (Grad-CAM) was incorporated to visualize the identified fracture features. To evaluate the practical performance of TRFC-net in classification assistance, we compared the classification accuracy of three attending physicians and resident physicians with and without the use of TRFC-net. Results: TRFC-net achieved an overall classification accuracy of 0.902, and its F1-scores for non-fracture, A1, A2, and A3 classifications were 0.964, 0.875, 0.867, and 0.848, respectively. Attending physicians achieved an overall classification accuracy of 0.939±0.014; the overall accuracy between TRFC-net and attending physicians was not significantly different [95% confidence interval (CI): 0.099-0.111; P=0.105], whereas that between the attending physicians and resident physicians was (P<0.001). TRFC-net provided greater improvement in the classification accuracy of residents than of attending physicians: the overall accuracy of resident physicians increased from 0.817±0.013 to 0.924±0.010, and the difference in overall accuracy between resident physicians assisted by TRFC-net was not significantly different from that of attending physicians (95% CI: 0.782-0.798; P=0.790). This suggests that TRFC-net may provide clinical utility in patient populations similar to those examined in this study. Conclusions: Within the dataset used in this study, TRFC-net effectively assisted inexperienced residents in approaching the classification accuracy of attending physicians; however, the diagnostic performance for A3 fracture subtypes was relatively unsatisfactory, and further validation with larger-scale, multicenter cohorts is needed.