AI Foundation Model as an Educational and Interpretation Aid for Bone Age Assessment

Purpose: Assessing skeletal bone age is crucial for diagnosing endocrinological, genetic, and growth disorders in pediatric patients. Current methods for bone age determination are labor-intensive, time-consuming, and subject to observer variability. We describe the use of an AI foundation model as an educational and interpretation aid for pediatric bone age assessment.

Methods/Materials: This retrospective study, approved by the Institutional Review Board (IRB), included 202 children (mean age: 10±4 years; male:female ratio = 100:102). All patients had left-hand radiography for bone age assessment between January and December 2018. Pediatric radiologists’ dictated radiology reports using patients’ age for bone age (years ± standard deviation) represented the ground truth. A general purpose medical image-image search AI foundation model (Microsoft Inc.) was leveraged to index 2 external datasets: GP index (59 GP reference images) and 12K index (12,611 images from RSNA training index). The AI tool's performance was evaluated based on the precision of the first four images to display the correct bone age. A Chi-Square test assessed gender-based differences in AI performance.

Results: The AI foundation model achieved precision rates of 0.76 and 0.86 for the 12K and GP index tools, respectively (TP: AI output within ± 1SD). Gender-wise (females, model precision 0.75; males, model precision 0.77) performance assessment revealed no significant differences (p-values ranging from 0.229 to 0.870). The tool allowed image search with the entire radiograph and within a user-selected region of interest.

Conclusions: The AI foundation model as an educational and interpretation aid can help determine pediatric bone age assessment with high precision. Point-of-interpretation similar image search foundational AI model can help in trainee education and interpretation of imaging studies.