Feasibility study of hallux valgus measurement with a deep convolutional neural network based on landmark detection

Objective To develop a deep learning algorithm based on automatic detection of landmarks that can be used to automatically calculate forefoot imaging parameters from radiographs and test its performance. Materials and methods A total of 1023 weight-bearing dorsoplantar (DP) radiographs were included. A total of 776 radiographs were used for training and verification of the model, and 247 radiographs were used for testing the performance of the model. The radiologists manually marked 18 landmarks on each image. By training our model to automatically label these landmarks, 4 imaging parameters commonly used for the diagnosis of hallux valgus could be measured, including the first–second intermetatarsal angle (IMA), hallux valgus angle (HVA), hallux interphalangeal angle (HIA), and distal metatarsal articular angle (DMAA). The reference standard was determined by the radiologists’ measurements. The percentage of correct key points (PCK), intragroup correlation coefficient (ICC), Pearson correlation coefficient ( r ), root mean square error (RMSE), and mean absolute error (MAE) between the predicted value of the model and the reference standard were calculated. The Bland–Altman plot shows the mean difference and 95% LoA. Results The PCK was 84–99% at the 3-mm threshold. The correlation between the observed and predicted values of the four angles was high (ICC: 0.89–0.96, r : 0.81–0.97, RMSE: 3.76–6.77, MAE: 3.22–5.52). However, there was a systematic error between the model predicted value and the reference standard (the mean difference ranged from − 3.00 to − 5.08°, and the standard deviation ranged from 2.25 to 4.47°). Conclusion Our model can accurately identify landmarks, but there is a certain amount of error in the angle measurement, which needs further improvement.