Related Experiment Video
Updated: Mar 1, 2026
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
3.3K
Large language models for dermatological image interpretation - a comparative study
Lasse Cirkel1,2, Fabian Lechner1,2, Lukas Alexander Henk2,3
1Institute of Artificial Intelligence, University Hospital Gießen-Marburg, Philipps University, Marburg, Germany.
Diagnosis (Berlin, Germany)
|May 27, 2025
Summary
Large language models (LLMs) show potential for diagnosing common skin conditions from images, with GPT-4o achieving the highest accuracy. However, performance varies, highlighting the need for cautious use and further development for reliable clinical decision support.
Area of Science:
- Artificial Intelligence in Dermatology
- Medical Image Analysis
- Clinical Decision Support Systems
Background:
- Interpreting dermatological findings presents challenges for both the public and medical professionals.
- Large language models (LLMs) offer potential for accessible diagnostic support, but their efficacy in dermatology is not well-established.
Purpose of the Study:
- To evaluate the diagnostic performance of various multimodal LLMs using dermatological images.
- To compare the accuracy of different LLMs in identifying common skin diseases.
Main Methods:
- Fifty0 dermatological images featuring psoriasis, vitiligo, erysipelas, and rosacea were analyzed.
- Seven multimodal LLMs, including GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet, were tested using a standardized prompt for top diagnosis generation.
Main Results:
- GPT-4o achieved the highest overall accuracy (67.8%), followed by GPT-4o mini (63.8%) and Llama3.2 11B (61.4%).
- Diagnostic accuracy differed across conditions, with psoriasis showing the highest mean LLM accuracy (59.2%) and erysipelas the lowest (33.4%).
- Clear disease features improved LLM accuracy, while Llama3.2 90B declined to diagnose images of intimate areas.
Conclusions:
- LLM performance in dermatological image diagnosis is variable, necessitating careful application.
- A free, locally deployable LLM achieved approximately 66% accuracy, indicating potential for secure, accessible clinical tools.
- Improving LLM accuracy and integrating clinical data can enhance diagnostic support systems.
