Deep Learning Approach to Improve Skin Lesion Classification for Early Skin Cancer Detection

Abstract

Skin cancer is a common and deadly disease, highlighting the need for early detection. This study evaluates five CNN architectures (DenseNet201, EfficientNetB0, XceptionNet, ResNet50, InceptionV3) and several Vision Transformer (ViT) models (ViT, Swin Transformer V2, DINOv2, PVT, ViT Hybrid) using the HAM10000 dataset. A 5-fold cross-validation assesses performance, and two Weighted Voting ensemble methods-one for CNNs and one for ViTs-are applied to enhance accuracy. Results are compared with transfer learning on ResNet50 and EfficientNetB0, showing that ensemble methods improve classification performance for early skin cancer detection.