Improvement of Bone Age Assessment Using a Deep Learning Model in Young Children: Significance of Carpal Bone Analysis

Background: Deep learning methods used for bone age assessment (BAA) mostly employ the whole hand or regional convolutional neural networks without carpal bones; therefore, their application is insufficient in young children. Objectives: This study aimed to improve the accuracy of BAA in young children by integrating a carpal bone analysis and to achieve a similar BAA accuracy for all age groups. Patients and Methods: A hybrid Greulich-Pyle (GP) and modified Tanner-Whitehouse deep learning model for BAA was trained by integrating an additional carpal bone analysis of an open dataset. A total of 453 hand radiographs from a single institution were selected for external validation. To create the reference standard, three human experts conducted a BAA, based on the GP Atlas, and then, interobserver agreement was evaluated. The model performance was estimated by comparing the mean absolute difference (MAD) and the root mean square error (RMSE) between the two BAA models, including one with a carpal bone analysis (M1) and one without a carpal bone analysis (M2), and the reference standard. TheMADof each model was compared between sex and age groups with respect to four major developmental stages, that is, pre-puberty, early and mid-puberty, late puberty, and post-puberty. Results: The M1 model showed a higher accuracy with a lower MAD (0.366; 95% confidence interval [CI]: 0.337 - 0.395) compared to the M2 model (0.388; 95% CI: 0.358 - 0.418) for all age groups, with a significant difference (P < 0.001). The RMSE values versus the reference standard were 0.483 and 0.505 years for the M1 and M2 models, respectively. According to sex and developmental stage distributions, the M1 model had a greater predictive ability compared to the M2 model for pre-pubertal patients, regardless of sex (P = 0.008 for males and P = 0.022 for females). Conclusion: Based on the present findings, the integration of a carpal bone analysis into the BAA model improved its accuracy, especially in young children.

Automated summary

This study aims to improve the accuracy of bone age assessment (BAA) in young children by integrating carpal bone analysis, resulting in a higher accuracy for all age groups.