Development and validation of a U-Net-based deep learning model for automatic measurement of the hallux valgus angle and intermetatarsal angle.

PURPOSE: To develop and validate a U-Net-based deep learning model for the automated measurement of the hallux valgus angle (HVA) and intermetatarsal angle (IMA) on weight-bearing dorsoplantar foot radiographs and to compare model performance with expert reference values.

METHODS: A total of 1227 weight-bearing dorsoplantar radiographs were obtained. The dataset was divided into training (n = 858), validation (n = 184), and testing (n = 185) datasets. Data augmentation was applied to the training set, yielding 10,296 samples. Six anatomical landmark points were labeled to define the axes of the proximal phalanges of the hallux, and of the first and second metatarsal. For temporal validation, 100 additional radiographs obtained at the same institution were evaluated. The U-Net-based model measurements were compared with the median values obtained from three independent experts (two foot and ankle surgeons and one musculoskeletal radiologist).

RESULTS: In the temporal validation cohort, the AI model demonstrated high agreement with the expert reference standard. For the HVA, the mean absolute error (MAE) was 2.118° (95% CI: 1.770-2.475) and the intraclass correlation coefficient (ICC) was 0.961. For the IMA, the MAE was 0.966° (95% CI: 0.809-1.133) and the ICC was 0.919. Subgroup analysis revealed that the accuracy remained consistent across mild, moderate, and severe deformities.

CONCLUSION: The proposed U-Net-based model provided accurate and reproducible automatic measurements of the HVA and IMA on weight-bearing dorsoplantar foot radiographs with high agreement with expert reference values. This approach may serve as a practical tool for standardized radiographic assessment of hallux valgus in clinical practice and research.