Abstract
Background: Melanoma is the deadliest form of skin cancer, for which early detection is crucial and can lead to positive survival outcomes. Advances in AI, particularly large language models (LLMs) such as GPT-5, present promising opportunities to support melanoma early detection, but their performance in this domain has not been systematically assessed. Objectives: Assess GPT-5's diagnostic performance on dermoscopic images. Methods: GPT-5 was evaluated on two public benchmark datasets: the ISIC Archive and HAM10K, using 100 and 500 randomly selected dermoscopic images, respectively. Via the OpenAI Application Programming Interface (API), GPT-5 was prompted to perform three tasks: (1) top-1 or primary diagnosis, (2) top-3 differential diagnoses, and (3) malignancy discrimination (melanoma vs. benign). Model outputs were compared with histopathology-verified ground truth, and performance was measured by sensitivity, specificity, accuracy, F1 score, and other metrics. Results: GPT-5 achieved modest accuracy in top-1 or primary diagnosis but markedly improved performance in top-3 differential diagnoses, with sensitivity > 93%, specificity > 86%, accuracy ≥ 92%, and F1 score > 91%. For malignancy discrimination, GPT-5 showed more balanced sensitivity and specificity than GPT-4-based models (GPT-4V, GPT-4T, and GPT-4o), resulting in more reliable classification overall. Conclusions: GPT-5 outperformed GPT-4 and its derivatives, particularly in differential diagnosis, highlighting its potential for clinical decision support and medical education. However, GPT-5 also showed a tendency to misclassify melanoma as benign, underscoring the need for cautious clinical interpretation and refinement.