Atlas-Assisted Bone Age Estimation from Hand-Wrist Radiographs Using Multimodal Large Language Models: A Comparative...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pediatric Endocrinology

Background:

• Bone age assessment is crucial in pediatric endocrinology and forensic medicine.
• Multimodal large language models (LLMs) show promise in medical imaging but require evaluation for bone age determination.
• The Gilsanz-Ratib (GR) atlas is a standard reference for bone age assessment.

Purpose of the Study:

• To evaluate the diagnostic performance of four multimodal LLMs (ChatGPT-5, Gemini 2.5 Pro, Grok-3, Claude 4 Sonnet) in bone age determination.
• To compare LLM performance against a radiologist's assessment using the GR atlas.
• To determine the clinical utility of these LLMs for bone age assessment.

Main Methods:

• Retrospective analysis of 245 pediatric wrist radiographs (age < 18).
• LLMs used GR atlas-assisted prompting for bone age estimation.
• Performance metrics included Mean Absolute Error (MAE), Intraclass Correlation Coefficient (ICC), and Bland-Altman analysis.
• Comparison with an experienced radiologist's GR atlas-based assessment.

Main Results:

• ChatGPT-5 achieved the lowest MAE (1.46 years) and highest ICC (0.849), demonstrating superior performance.
• Gemini 2.5 Pro showed moderate performance (MAE: 2.24 years).
• Grok-3 (MAE: 3.14 years) and Claude 4 Sonnet (MAE: 4.29 years) exhibited error rates too high for clinical application.

Conclusions:

• Significant performance variations exist among multimodal LLMs for bone age assessment.
• Only ChatGPT-5 met criteria for clinical usefulness, suggesting potential as an auxiliary tool or for educational support under supervision.
• Current multimodal LLMs, excluding ChatGPT-5, lack the reliability for clinical bone age determination.