A Comparative Study of Artificial Intelligence and Machine Learning Techniques for Skin Disease Detection: Melanoma, Basal Cell Carcinoma, and Melanocytic Nevi
TL;DR
Persistent research gaps are identified, including severe class imbalance, an under-studied melanoma-versus-nevus decision boundary, limited focus on malignancy-specific triage models, sparse use of explainability methods such as Grad-CAM, and weak cross-dataset generalization testing.
Abstract
Skin cancer is among the most common and rapidly increasing malignancies worldwide, with early and accurate diagnosis directly influencing patient survival. Conventional visual examination by dermatologists achieves an average diagnostic accuracy of only about 60%, rising to roughly 89% with dermoscopy, leaving considerable scope for computational decision support. Convolutional Neural Networks (CNNs) have repeatedly matched or exceeded dermatologist-level performance on dermoscopic image classification tasks, motivating growing research interest in automated, image-processing-driven diagnostic systems. This review paper synthesizes current literature on CNN-based skin disease detection, with a focused lens on three clinically significant lesion categories: Malignant Melanoma (MEL), Basal Cell Carcinoma (BCC), and Melanocytic Nevi (NV). The paper surveys benchmark datasets (HAM10000, ISIC 2019, PH2), preprocessing and image-processing techniques (hair removal, segmentation, colour normalization, augmentation), transfer-learning backbones (ResNet, DenseNet, EfficientNet, Inception, VGG), and evaluation practices reported across more than fifty peer-reviewed and pre-print sources. Persistent research gaps are identified, including severe class imbalance, an under-studied melanoma-versus-nevus decision boundary, limited focus on malignancy-specific triage models, sparse use of explainability methods such as Grad-CAM, and weak cross-dataset generalization testing.