Abstract
BACKGROUND: Hallux valgus (HV), also known as bunion deformity, is one of the most common forefoot deformities. Early diagnosis and proper evaluation of HV are important because timely management can improve symptoms and quality of life. Here, we propose a deep learning estimation for the radiographic measurement of HV based on a regression network where the input to the algorithm is digital photographs of the forefoot, and the radiographic measurement of HV is computed as output directly. The purpose of our study was to estimate the radiographic parameters of HV using deep learning, to classify the severity by grade, and to assess the agreement of the predicted measurement with the actual radiographic measurement. METHODS: There were 131 patients enrolled in this study. A total of 248 radiographs and 337 photographs of the feet were acquired. Radiographic parameters, including the HV angle (HVA), M1-M2 angle, and M1-M5 angle, were measured. We constructed a convolutional neural network using Xception and made the classification model into the regression model. Then, we fine-tuned the model using images of the feet and the radiographic parameters. The coefficient of determination (R(2)) and root mean squared error (RMSE), as well as Cohen's kappa coefficient, were calculated to evaluate the performance of the model. RESULTS: The radiographic parameters, including the HVA, M1-M2 angle, and M1-M5 angle, were predicted with a coefficient of determination (R(2))=0.684, root mean squared error (RMSE)=7.91; R(2)=0.573, RMSE=3.29; R(2)=0.381, RMSE=5.80, respectively. CONCLUSION: The present study demonstrated that our model could predict the radiographic parameters of HV from photography. Moreover, the agreement between the expected and actual grade of HV was substantial. This study shows a potential application of a convolutional neural network for the screening of HV.