Abstract

Objectives: To evaluate and compare the diagnostic performance of GPT‑4o and Gemini Flash 2.0 in identifying and subtyping acne and rosacea from clinical photographs. Introduction: Large language models (LLMs) are increasingly accessed by patients for dermatological self‑diagnosis, yet their performance for common inflammatory conditions such as acne and rosacea is poorly defined. This study tested two widely available LLMs for real‑world diagnostic accuracy. Materials / method: Between December 2021 and December 2024, 43 clinical images (33 acne, 10 rosacea) from 31 patients with histologically or clinically confirmed diagnoses were evaluated. Images were submitted to GPT‑4o and Gemini Flash 2.0 using standardized prompts simulating patient queries. Two dermatologists provided reference diagnoses, with consensus adjudication by a third. Sensitivity, specificity, PPV, and NPV were calculated for overall diagnosis and subtyping. Results: GPT‑4o achieved 93% overall diagnostic accuracy with sensitivity of 93.0% and specificity of 97.7%. For acne, sensitivity was 90.9% and specificity 100%; for rosacea, sensitivity was 100% and specificity 97.7%. Subtype accuracy was lower: acne subtypes, sensitivity 54.6%, specificity 89.9%; rosacea subtypes, sensitivity 50.0%, specificity 80.0%. Gemini Flash 2.0 generated diagnoses in only 21% of cases, precluding meaningful statistical analysis. Conclusion: GPT‑4o substantially outperformed Gemini Flash 2.0 in diagnosing acne and rosacea, showing high accuracy for primary classification but limited ability for subtyping. These results highlight both the potential and current limitations of LLMs in dermatology and emphasize the need for dermatologist oversight as patients increasingly consult AI tools.