Construction and Validation of A Deep Learning-based Bone Age Prediction Model for Children Living in Both Plain and Highland Regions

Objective To construct and validate a deep learning-based bone age prediction model for children living in both plain and highland regions.

Methods A model named "ethnicity vision gender-bone age net (EVG-BANet)" was trained using three datasets, including the Radiology Society of North America (RSNA) dataset training set(n=12 611), validation set (n=1425), test set (n=200), the Radiological Hand Pose Estimation (RHPE) datasettraining set (n=5491), validation set (n=713), test set (n=79), and a self-established datasettraining set (n=825), test set (n=351), and it was validated using an external test set. Self-established dataset retrospectively recruited 1176 left-hand DR images of children from Peking Union Medical College Hospital (n=745, all were Han) and Tibet Autonomous Region People's Hospital (n=431, 114 were Han, 317 were Tibetan). External test set included images from People's Hospital of Nagqu (n=256, all were Tibetan). Mean absolute difference (MAD) and accuracy within 1 year were used as indicators.

Results EVG-BANet exhibited MAD of 0.34 and 0.52 years in RSNA and RHPE test sets, respectively. In the self-established test set, the model achieved MAD of 0.47 years (95% CI: 0.43-0.50) with accuracy within 1 year of 97.72% (95% CI: 95.56-99.01%). For the external test set, MAD was 0.53 years(95% CI: 0.48-0.58), with accuracy within 1 year of 89.45% (95% CI: 85.03-92.93).

Conclusion EVG-BANet demonstrated high accuracy in bone age prediction, and therefore can be applied in children living in both plain and highland.