Automatic estimation of hallux valgus angle using deep neural network with axis-based annotation.

Objectives: We developed the deep neural network (DNN) model to automatically measure hallux valgus angle (HVA) and intermetatarsal angle (IMA) on foot radiographs. The objective is to assess the accuracy of the model by comparing to the manual measurement of foot and ankle surgeons.

Materials and methods: A DNN was developed to predict the bone axes of the first proximal phalanx and all metatarsals from the first to the fifth in foot radiographs. The dataset used for model development consisted of 1798 radiographs collected from a population-based cohort and patients at our foot and ankle clinic. The retrospective validation cohort comprised of 92 radiographs obtained from 92 consecutive patients visiting our foot and ankle clinic. The mean absolute error (MAE) between automatic measurements by the model and the median of manual measurements by three foot and ankle surgeons was compared to 3° using one-tailed t-test and was also compared to the inter-rater difference in manual measurements among the three surgeons using two-tailed paired t-test.

Results: The MAE for HVA was 1.3° (upper limit of 95% CI 1.6°), and this was significantly smaller than the inter-rater difference of 2.0 ± 0.2° among the surgeons, demonstrating the superior accuracy of the model. In contrast, the MAE for IMA was 0.8° (upper limit of 95% CI 1.0°) that showed no significant difference from the inter-rater difference of 1.0 ± 0.1° among the surgeons.

Conclusion: Our model demonstrated the ability to measure the HVA and IMA with an accuracy comparable to that of specialists.