Abstract
OBJECTIVE: The purpose of this study is to develop an automated model to assist in the detection of substructural injuries of the triangular fibrocartilage complex (TFCC), thereby reducing the reliance on subjective assessment. METHODS: This retrospective study utilized 330 TFCC injured patients and 273 healthy controls from two institutions, only analyzing 2821 coronal fat-saturated T2-weighted imaging slices. From 483 samples (267 injured, 216 normal), 2292 images were processed: 1834 for training, 458 for validation, with an internal test set of 209 images from 47 samples (26 injured, 21 normal). An external test set comprised 320 images from 73 samples (37 injured, 36 normal) at another institution. Radiologists segmented and classified TFCC substructures by consensus. Different YOLO versions were trained and compared, with the optimal model benchmarked against musculoskeletal (MSK) radiologists (Resident1 and Attending2). RESULTS: Among evaluated YOLO versions, the YOLO11l model exhibited the optimal segmentation performance, with mean Dice (mDice) coefficients of 0.82 (internal test set) and 0.77 (external test set). Its classification sensitivity, specificity, and accuracy were 91.67%, 76.11%, and 83.25% on the internal test set, significantly outperforming other versions. On the external test set, corresponding values were 84.68%, 61.22%, and 71.00%, representing the best overall performance. Notably, the diagnostic performance of the YOLO11l model was non-inferior to that of Resident1 (p = 0.41) but inferior to that of Attending2 (p = 0.015). CONCLUSIONS: The YOLO11l model represents a promising approach to aiding the assessment of TFCC injuries. Compared with less experienced radiology residents, this model can provide reliable and reproducible diagnostic support.