Deep learning versus manual measurement of hallux valgus angle and intermetatarsal angle on Weight-Bearing X-rays in hallux valgus

This study aimed to develop two deep learning (DL)-based automated measurement systems (AMSs) for hallux valgus (HV) angles estimation, namely axis segmentation-based and landmark-based AMS, and to evaluate their reliability, accuracy, and efficiency relative to manual measurements. AMSs were developed using weight-bearing dorsoplantar radiographs (1,683 images from 1,232 patients with HV). An evaluation cohort (208 feet: 121 from patients with HV and 87 from controls) underwent manual and automated measurements of the hallux valgus angle (HVA) and intermetatarsal angle (IMA). Intraclass correlation coefficients (ICCs), Pearson correlation coefficients (r), mean absolute error (MAE), root mean square error (RMSE), Bland-Altman plots, and measured average processing time were calculated. The axis segmentation-based AMS outperformed the landmark-based AMS in both groups, with stronger correlations for HVA (HV: r = 0.96 vs. 0.95; controls: r = 0.77 vs. 0.74) and IMA (HV: r = 0.83 vs. 0.78; controls: r = 0.57 vs. 0.45; all P < 0.0001). In both AMSs and groups, HVA measurements exhibited lower errors and stronger correlations than IMA measurements. Both models significantly distinguished patients with HV from controls (P < 0.0001). The average time required was 3.5 min for manual measurement versus 6.3 s (axis segmentation-based) and 6.5 s (landmark-based) for automated measurements. Our findings indicated that DL-based AMSs provided reliable, reproducible, and accurate HVA and IMA measurements from HV radiographs, with the axis segmentation-based AMS outperforming the landmark-based AMS.