Scholarlink

Objective <p>Rapid advances in artificial intelligence have increased interest in using large language models (LLMs) for medical education and clinical applications. This exploratory study evaluated the ability of three multimodal LLMs, ChatGPT 5, Gemini 2.5 Flash, and Grok 4, to identify anatomical structures in cross-sectional images of the upper and lower limbs.</p> Results <p>Twenty cross-sectional images, each highlighting a single anatomical structure, were presented to the models with standardized prompts specifying the anatomical region. Accuracy was scored for each model. ChatGPT 5 correctly identified 9 of 20 structures (45%, 95% CI: 23.1–68.5%), Gemini 2.5 Flash 5 of 20 (25%, 95% CI: 8.7–49.1%), and Grok 4 4 of 20 (20%, 95% CI: 5.7–43.7%). A qualitative error analysis revealed common misclassification patterns. These results indicate modest accuracy under the tested conditions and highlight areas for model improvement.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Critical evaluation of large language models for human cross-sectional anatomy identification: implications for collaborative intelligence

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月21日

查看全文
  • Aryan Kermansaravi,
  • Elenasadat Tonekabonipour,
  • Mohamed Abdelhalim,
  • Salman Farooq Dar,
  • Bassem Amr

摘要

Objective

Rapid advances in artificial intelligence have increased interest in using large language models (LLMs) for medical education and clinical applications. This exploratory study evaluated the ability of three multimodal LLMs, ChatGPT 5, Gemini 2.5 Flash, and Grok 4, to identify anatomical structures in cross-sectional images of the upper and lower limbs.

Results

Twenty cross-sectional images, each highlighting a single anatomical structure, were presented to the models with standardized prompts specifying the anatomical region. Accuracy was scored for each model. ChatGPT 5 correctly identified 9 of 20 structures (45%, 95% CI: 23.1–68.5%), Gemini 2.5 Flash 5 of 20 (25%, 95% CI: 8.7–49.1%), and Grok 4 4 of 20 (20%, 95% CI: 5.7–43.7%). A qualitative error analysis revealed common misclassification patterns. These results indicate modest accuracy under the tested conditions and highlight areas for model improvement.