Cutan Ocul Toxicol. 2025 Dec 8:1-8. doi: 10.1080/15569527.2025.2598573. Online ahead of print.
ABSTRACT
INTRODUCTION: The global incidence of skin cancer is rising, emphasizing the need for early detection tools. Artificial intelligence (AI) models, including multimodal systems like ChatGPT-4o, can analyze visual data to assist clinicians in diagnosis. This study evaluated ChatGPT-4o’s diagnostic accuracy in detecting melanoma and non-melanoma skin cancers from macroscopic and dermoscopic images.
METHODS: Ninety patients with histopathologically confirmed lesions were included. For each patient, macroscopic images were first uploaded to ChatGPT-4o, followed by combined upload of macroscopic and dermoscopic images. ChatGPT-4o was instructed to provide a preliminary diagnosis and three differential diagnoses for each lesion. Accuracy was assessed at four levels: Level 1: preliminary diagnosis using macroscopic images; Level 2: preliminary diagnosis using macroscopic and dermoscopic images; Level 3: three differential diagnoses using macroscopic images; Level 4: three differential diagnoses using macroscopic and dermoscopic images.
RESULTS: Overall Level 1 accuracy was 73.3%, with Level 2, Level 3, and Level 4 accuracies of 66.6%, 75%, and 76.6%, respectively. Dermoscopic images improved accuracy for squamous cell carcinoma (72.7% vs 81.8%, p = 1.00), reduced overall and basal cell carcinoma accuracy (73.3% vs 66.6%, p = 0.180 and 79.6% vs 67.8%, p = 0.065, respectively), and did not affect malignant melanoma (84.6% vs 84.6%) or lentigo maligna (0% vs 0%). Statistical analysis revealed that the addition of dermoscopic images did not significantly influence diagnostic accuracy, either overall or within individual diagnostic categories. The model recommended biopsy for all lesions, suggesting potential as a supportive diagnostic tool.
CONCLUSION: ChatGPT-4o showed variable diagnostic accuracy for melanoma and non-melanoma skin cancers. Dermoscopic images reduced performance for certain diagnostic categories. These misclassifications highlight the potential for unnecessary interventions and patient anxiety underscoring that AI-based systems should serve as supportive aids rather than standalone diagnostic tools.
PMID:41355746 | DOI:10.1080/15569527.2025.2598573
