Advanced Visual Turing Test For Validation And Responsible Use Of Synthetic Dermoscopic Images Of Melanoma And Atypical Nevi In Medicine

Abstract

Introduction: Dermoscopy is essential for the early diagnosis of melanoma. AI algorithms integrated into videodermatoscopes can support dermatologists in distinguishing atypical nevi from early melanomas in real time, yet their development is consistently hampered by the scarcity of large, annotated dermoscopic datasets. Generative Adversarial Networks (GANs) offer a promising solution through synthetic data augmentation, though their clinical validation remains a significant methodological challenge. Objectives: This study evaluates the morphological realism and diagnostic utility of synthetic images of melanocytic lesions through a structured "visual Turing test" validation protocol. Methods: A testing set of 60 dermoscopic images was assembled: 30 real (15 atypical nevi, 15 early melanomas) and 30 synthetic. Real images were acquired before surgical excision and histological analysis at Siena University Hospital. Synthetic images were generated using the StyleGAN2-ADA architecture at the Biomedical Engineering Laboratory of the University of Siena. The set was derived from a quality-based pre-selection of 100 images by two expert dermatologists, followed by a K-Nearest Neighbours plausibility assessment applied to the synthetic samples. Expert dermatologists affiliated with the European Task Force on AI in Dermatology then evaluated each image through a purpose-built online platform, discriminating against image origin (real vs. synthetic), formulating a clinical diagnosis, and mapping the presence of 16 dermoscopic patterns. Results:  Accuracy in real/synthetic discrimination was 56.67%, approximating chance-level performance. The accuracy values for benign/malignant discrimination were 76,67% for real images and 72.50% for synthetic ones, suggesting that the tested synthetic images effectively preserved the morphological criteria required for clinical application. Conclusions: This protocol the protocol provides a structured approach for expert evaluation of synthetic dermoscopic images and may complement future clinical and regulatory validation procedures.