Accuracy of deep learning based bone age assessment system of children and adolescents in Guizhou / 中国医学影像技术

ABSTRACT

Objective: To explore the clinical applicability of a deep learning based bone age assessment system of children and adolescents in Guizhou. Methods: The left hand-wrist radiographs of 148 children and adolescents aged from 2 years to 17 years were assessed independently by three experts who were trained with the CH 05 RUS-CHN method, their mean estimates results were used as the reference standard. The estimates of the deep learning model (model group) and two residents (control group) were evaluated compared with the reference standard, respectively. mean absolute error (MAE) of bone age estimates and the percentage of samples with absolute error (AE) ≤1.0 year were calculated. Results: MAE of the model group was 0.295 [95%CI (0.238, 0.352)] years, with absolute error ≤1 years of 93.92% (139/148). Doctor A of the control group recorded MAE was 0.438 [95%CI (0.369, 0.508)] years, with 89.19% absolute error ≤1.0 years of 89.19% (132/148); doctor B recorded MAE of 0.360 [95%CI (0.295, 0.425)] years, with absolute error ≤1.0 years of 89.86% (133/148). The MAE of model group was significantly lower than that of doctor A (t=-3.071, P=0.002), but not for the doctor B (t=-1.563, P=0.120). Conclusion: When bone age assessed with the CH 05 RUS-CHN method for Guizhou children and adolescents, the deep learning model can estimate bone age with accuracy similar to or even better than that of control group radiologists.