Clinically Interpretable Deep Learning Model for Bone Age Estimation Based on TW3 Scoring

Background Bone age assessment (BAA) is challenging in clinical practice due to the complicated bone age assessment process. We aim to develop a deep learning-based system to fully automate the assessment of bone age using pediatric hand radiographs, similar to a radiologist’s workflow using the Tanner-Whitehouse 3 method (TW3) method. Methods We present a new experimental framework. The framework consists of three parts. One is hand detection. In this module, a deep convolution object detection object is used to detect the left hand in the image, orientation is also detected. In the key-points detection module, another deep convolution key-point detection network is used to localize the region of interest (ROI) in the hand and get the corresponding maturity levels. Finally, we designed a bone age computer based on TW3 method to determine the final bone age. Results A total of 1309 subjects were enrolled in the study, and 241 subjects were divided into test set. The accuracy of orientation classification on the test set is 100%. The mean absolute error (MAE) of predicted radius, ulna, short bone (RUC) age and Carpal bone age were 0.35 and 0.55 respectively. In addition, the AUC of the remaining maturity level for all bone maturity assessment reached at least 0.85. Conclusion Compared with other automatic bone age assessment methods, our framework shows competitive performance with low model complexity and excellent interpretability. Our framework can deal with different noises and provide a reliable result.