AI-ASSISTED DERMOSCOPY FOR EARLY DIAGNOSIS OF MALIGNANT MELANOMA: A CLINICAL VALIDATION

Abstract

Early and accurate detection of malignant melanoma is critical to improving patient survival rates. This study investigates the diagnostic performance of an artificial intelligence (AI)-assisted dermoscopy system powered by a convolutional neural network (CNN) trained on a large, diverse dermoscopic image dataset. The AI model achieved a high diagnostic accuracy of 93.5%, with a sensitivity of 92.1%, specificity of 94.9%, and an area under the ROC curve (AUC) of 0.964, outperforming experienced dermatologists in comparative evaluations. Analysis across Fitzpatrick skin types I–VI revealed consistent accuracy, indicating the model's generalizability across diverse populations. The confusion matrix showed strong discriminatory power, with minimal false positives and negatives. Explainability techniques, including Grad-CAM and SHAP, provided visual and quantitative interpretability of the model’s predictions, reinforcing trust in its diagnostic reasoning. Comparative assessments demonstrated that the AI system surpassed human evaluators in sensitivity and precision, validating its role as a clinical decision support tool. Additionally, SHAP-based feature importance and attention mapping illustrated the system’s ability to focus on clinically relevant image regions, further bridging the gap between AI and human diagnostic logic. Despite its strengths, the study recognizes limitations such as dependence on data quality and the need for robust clinical integration pathways. Nonetheless, the findings underscore the potential of AI-assisted dermoscopy to revolutionize melanoma diagnosis, reduce unnecessary biopsies, and ensure early intervention. Future research should aim to validate the system in real-world clinical trials, integrate it with patient history data, and develop regulatory frameworks to guide ethical deployment.