Dermoscopic Skin Lesion Classification Using Artificial Intelligence

Abstract

Skin cancer remains one of the most prevalent malignancies worldwide, with melanoma accounting for the majority of related deaths. Early and accurate detection is critical for improving patient outcomes, yet traditional diagnostic methods are often subjective and resource-intensive. This study investigates the efficacy of deep learning-based binary classification of dermoscopic skin lesions into benign and malignant categories using the publicly available HAM10000 dataset. Malignant cases were defined to include melanoma and basal cell carcinoma, while all other lesion types were grouped as benign. We systematically compared two convolutional neural network (CNN) architectures, EfficientNetB7 and VGG19, under identical preprocessing conditions, including resizing, normalization, noise reduction, and artifact removal (e.g., hair inpainting). Model performance was evaluated using accuracy, loss, and validation metrics. Experimental results demonstrated that while EfficientNetB7 achieved marginally higher training accuracy (87.04%), VGG19 exhibited superior generalization to unseen data, attaining consistent training and validation accuracy (86.44%) with significantly lower and balanced loss values (0.3972), indicating reduced overfitting. These findings suggest that VGG19 offers greater robustness and reliability for real-world clinical applications. Future work will focus on optimizing transfer learning, advanced data augmentation, and regularization techniques to improve generalizability and interpretability.