Multimodal Large Language Models for Pediatric Bone-Age Assessment: A Comparative Accuracy Analysis

Abstract

Rationale and objectives: Bone age assessment guides endocrine and orthopedic care in children, yet manual Greulich‑Pyle reading is subjective, and commercial deep‑learning tools are costly. We examined whether readily accessible closed‑source multimodal large language models can provide clinically acceptable bone age assessment accuracy.

Materials and methods: A balanced set of 180 left-hand wrist radiographs (ages 0-18 years; RSNA Pediatric Bone Age dataset) was analyzed by three models (ChatGPT‑o4-mini‑high, ChatGPT‑4.5, and Google Gemini 2.5 Pro) using an identical prompt. Reference ages were adjusted bone ages by consensus of radiologists. Accuracy metrics were mean absolute error (MAE) and root‑mean‑square error (RMSE); agreement was assessed with Bland-Altman analysis. Model performance was compared with a Friedman test followed by Bonferroni‑adjusted Wilcoxon signed‑rank tests.

Results: Gemini 2.5 Pro yielded the lowest error (MAE 28.48 months, RMSE 35.23 months), comparing with ChatGPT‑4.5 (MAE 35.38 months, RMSE 43.64 months) and ChatGPT‑o4-mini‑high (MAE 35.87 months, RMSE 45.43 months; both P < .001). Bias was -9.38 months for Gemini, -4.80 months for ChatGPT‑4.5, and +5.26 months for ChatGPT‑o4-mini‑high; 95% limits of agreement exceeded ±4.5 years for every model.

Conclusion: Multimodal Large Language Models can estimate bone age directly from radiographs without task‑specific training, but current errors of 2-3years and wide errors remain unsuitable for clinical or medico‑legal decision‑making. Advances in vision encoder capacity and domain‑specific fine‑tuning are needed before LLMs can replace dedicated BAA software or expert readers.

Keywords: Artificial intelligence; Bone-age; Large Language Models; Radiography.