Comparative Analysis of Transfer Learning-Based Deep Learning Models for Skin Cancer Classification Using the HAM10000 Dataset

Abstract

Skin cancer is among the most common forms of cancer worldwide, and early diagnosis is essential for improving patient survival rates. Recent advances in deep learning and computer vision have enabled the development of automated diagnostic systems capable of assisting dermatologists in skin lesion classification. This study presents a comparative analysis of transfer learning-based deep learning models for multiclass skin cancer classification using the HAM10000 dataset, which contains dermoscopic images belonging to seven different skin lesion categories. Five state-of-the-art convolutional neural network architectures, namely ResNet50, ResNet101, ResNet152, EfficientNetB0, and EfficientNetB7, were investigated. The models were trained using transfer learning with ImageNet pretrained weights and data augmentation techniques to improve generalization performance.

Experimental evaluation was conducted using accuracy, precision, recall, F1-score, confusion matrix, and ROC-AUC metrics. Among the evaluated models, EfficientNetB0 achieved the highest test accuracy of 75.45% and validation accuracy of 77.56%, while EfficientNetB7 obtained the highest ROC-AUC score of 94.08%, demonstrating excellent discriminative capability. ResNet50 achieved competitive performance with a validation accuracy of 74.83%, whereas ResNet152 exhibited comparatively lower performance. The results indicate that EfficientNet architectures provide superior feature extraction and classification performance for dermoscopic image analysis.

The findings demonstrate the effectiveness of transfer learning for automated skin cancer detection and highlight the potential of deep learning-based computer-aided diagnostic systems in supporting dermatologists for early and accurate skin lesion classification. Future work will focus on class imbalance handling, explainable artificial intelligence techniques, and ensemble learning approaches to further improve classification performance.