Automated bone age assessment in rare pediatric growth disorders: a comparative study using Deeplasia

Area of Science:

• Pediatric Endocrinology
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Bone age (BA) assessment is crucial for growth monitoring and treatment guidance.
• Deep learning (DL) models offer automated BA prediction but face generalizability challenges in complex pediatric populations.
• Validating DL tools on diverse patient groups is essential for clinical adoption.

Purpose of the Study:

• To validate the Deeplasia open-source deep learning system on external data from pediatric patients with syndromic, endocrine, and lysosomal storage disorders (LSDs).
• To compare the accuracy and consistency of Deeplasia against multiple expert human raters for bone age assessment.
• To evaluate the generalizability of DL-based bone age prediction in rare and complex pediatric cases.

Main Methods:

• Retrospective analysis of 1,138 hand radiographs from pediatric patients with various endocrine, syndromic, and LSDs.
• Bone age reference standards established using the Greulich and Pyle method by multiple expert raters.
• Model performance evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and 1-year accuracy, with comparisons against individual human raters.

Main Results:

• Deeplasia achieved MAE of 5.95 months and 89.9% 1-year accuracy for endocrine/syndromic conditions (Cohort 1).
• For LSDs (Cohort 2), Deeplasia achieved MAE of 7.13 months and 81.2% 1-year accuracy.
• Deeplasia demonstrated superior accuracy and consistency compared to individual human expert raters when tested against remaining experts.

Conclusions:

• Deeplasia is a validated, consistent, robust, and reliable tool for bone age assessment in complex pediatric cases.
• The DL system shows superior accuracy compared to individual human raters.
• Deeplasia has the potential to assist clinicians in bone age evaluation, especially in challenging diagnostic scenarios.