Publicly Available Large Language Models for Trichoscopy: A Head-to-Head Comparison with Dermatologists

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Trichoscopy is crucial for diagnosing hair and scalp disorders, demanding significant expertise.
• The diagnostic utility of publicly available large language models (LLMs) in trichology remains unexplored.
• This study assesses LLMs' accuracy in interpreting trichoscopic images against human dermatologists.

Purpose of the Study:

• To evaluate the diagnostic accuracy of four LLMs in trichoscopic image interpretation.
• To compare LLM performance against dermatology residents, board-certified dermatologists, and trichology experts.
• To determine the potential of AI in assisting trichological diagnoses.

Main Methods:

• A prospective comparative study using a preprocessed set of structurally transformed trichoscopic images.
• Fifteen dermatologists (residents, board-certified, experts) provided suspected and differential diagnoses.
• Four LLMs (ChatGPT-4o, Claude Sonnet 4, Gemini 2.5 Flash, Grok-3) evaluated images under identical conditions.

Main Results:

• Human dermatologists achieved an overall diagnostic accuracy of 58.1% for suspected diagnoses and 68.3% for suspected + differential diagnoses.
• AI models achieved lower accuracy: 18.2% for suspected diagnoses and 44.4% for suspected + differential diagnoses.
• Gemini 2.5 Flash showed the highest AI accuracy (62.5% for suspected + differential diagnoses), yet all AI models significantly underperformed human experts (p < 0.001).

Conclusions:

• Publicly available LLMs currently underperform human experts in trichoscopic diagnosis, particularly for single correct diagnoses.
• AI models demonstrate moderate to good agreement among themselves but show only slight to fair agreement with dermatologists.
• Specialized training and further development are essential for LLMs to be reliable tools in routine trichological care.