Construction of artificial intelligence system of carpal bone age for Chinese children based on China-05 standard.

ABSTRACT

PURPOSE: The purpose of this study is to construct an automatic carpal bone age evaluation system for Chinese children based on TW3-C Carpal method by deep learning and to evaluate the accuracies in test set and clinical test set. METHODS: A total of 8184 radiographs of Chinese Han healthy children (2-14 years old) were collected from the 2005 China Skeletal Development Survey data and from the accumulated radiographs in bone age studies over the years. Three experienced radiologists and the China-05 standard maker jointly evaluated each bone development stage, and the consensual stage was decided as the reference standard. According to each epiphysis development stage, 10% of them were derived by stratified random sampling as test sets, and the remaining radiographs were used as training sets and validation sets. Furthermore, the overall performance of the model was estimated by comparing the mean difference, 95% limits of agreement, mean absolute difference (MAD) and root mean square (RMS) between the model predictions and the reference standard. RESULTS: The percentage agreement of ratings in each epiphysis in the test set ranged from 82.82% to 90.06%, with an average of 86.94%. Compared with the reference standard, the automated bone age system has a mean difference of 0.01 years old, ± 0.45 years old in 95% confidence interval by single reading, an 85.93% percentage agreement of ratings, a 90.5% bone age accuracy rate, 0.20 years old of MAD, and 0.32 years old of RMS in the clinical test set. CONCLUSIONS: The automatic bone age system for Chinese children has a comparable accuracy and stable determination compared with experienced radiologists.