RESUMEN
OBJECTIVES: To develop a deep learning model for automated age estimation based on 3D CT reconstructed images of Han population in western China, and evaluate its feasibility and reliability. METHODS: The retrospective pelvic CT imaging data of 1 200 samples (600 males and 600 females) aged 20.0 to 80.0 years in western China were collected and reconstructed into 3D virtual bone models. The images of the ischial tuberosity feature region were extracted to create sex-specific and left/right site-specific sample libraries. Using the ResNet34 model, 500 samples of different sexes were randomly selected as training and verification set, the remaining samples were used as testing set. Initialization and transfer learning were used to train images that distinguish sex and left/right site. Mean absolute error (MAE) and root mean square error (RMSE) were used as primary indicators to evaluate the model. RESULTS: Prediction results varied between sexes, with bilateral models outperformed left/right unilateral ones, and transfer learning models showed superior performance over initial models. In the prediction results of bilateral transfer learning models, the male MAE was 7.74 years and RMSE was 9.73 years, the female MAE was 6.27 years and RMSE was 7.82 years, and the mixed sexes MAE was 6.64 years and RMSE was 8.43 years. CONCLUSIONS: The skeletal age estimation model, utilizing ischial tuberosity images of Han population in western China and employing the ResNet34 combined with transfer learning, can effectively estimate adult ischium age.
