Bone age estimation from chest radiographs using deep neural networks: a proof-of-concept study.

BACKGROUND: Bone age (BA) is the gold standard for skeletal maturity assessment but is not routinely incorporated into pediatric growth evaluation workflow because it requires dedicated hand radiographs and specialist interpretation. OBJECTIVE: To develop an estimation model for BA from chest radiographs using deep neural networks. MATERIALS AND METHODS: We retrospectively analyzed children aged 3-15 years who underwent both chest and hand radiography at a tertiary center over 20 years. Patients with skeletal dysplasia or chest wall abnormalities were excluded. Reference BA was determined from hand radiographs by pediatric endocrinologists using the Tanner-Whitehouse 2 radius-ulna-short bones (TW2-RUS) method. Three pretrained deep neural networks were fine-tuned to estimate BA from chest radiographs using sex-non-considering models and sex-considering models. Model performance was evaluated using the intraclass correlation coefficient (ICC), root mean squared error (RMSE), and related metrics. RESULTS: Of 180 screened patients, 101 were included, yielding 237 chest radiographs. Estimated BA showed good concordance with the reference standard, with ICCs up to 0.81 (95% CI 0.48-0.97) for sex-non-considering models and 0.87 (95% CI 0.63-0.99) for sex-considering models. Corresponding RMSEs were 1.52 (95% CI 0.67-2.20) and 1.30 (95% CI 0.56-1.95), respectively. CONCLUSION: This proof-of-concept study demonstrates the feasibility of estimating BA from pediatric chest radiographs using deep neural networks. These findings suggest potential opportunistic use when dedicated hand radiographs are unavailable.