Prospective multicenter study using artificial intelligence to improve dermoscopic melanoma diagnosis in patient care

Background: Early detection of melanoma improves prognosis. Retrospective studies suggest AI can enhance melanoma detection, but prospective evidence is limited and often constrained by small, homogeneous datasets lacking rare melanoma subtypes. Methods: We prospectively evaluated ADAE, an open-source ensemble melanoma classifier, against dermatologists across eight hospitals using four camera setups, including rare melanoma subtypes and special anatomical sites. We employed real test-time augmentation (R-TTA) by using multiple real images per lesion (varied angles/positions, polarized and nonpolarized) and averaging predictions to assess generalization. Results: ADAE achieved higher balanced accuracy than dermatologists (0.798, 95% CI 0.779–0.814 vs. 0.781, 95% CI 0.760–0.802) with much higher sensitivity (0.921, 95% CI 0.900–0.942 vs. 0.734, 95% CI 0.701–0.770) at the cost of lower specificity (0.673, 95% CI 0.641–0.702 vs. 0.828, 95% CI 0.804–0.852). Conclusion: On a challenging evaluation set of melanoma-suspicious lesions, AI showed a significant performance advantage and may support dermatologists, particularly in difficult cases.