Comparative Evaluation of Convolutional Neural Network Architectures for Automated Skin Cancer Classification: A Study on the ISIC 2018 Dataset

Abstract

The increasing global incidence of skin cancer, particularly lethal malignant melanoma, necessitates the development of robust, automated diagnostic tools to assist dermatologists in identifying subtle pathological markers. In this study, we provide a rigorous comparative evaluation of four state-of-the-art convolutional neural network (CNN) architectures, ResNet-50, DenseNet-169, Inception-v3, and EfficientNet-B0, using the ISIC 2018 (HAM10000) dataset. Our standardized experimental pipeline utilized stratified sampling to address class imbalance, alongside a meticulous preprocessing strategy and data augmentation to ensure model generalization. Quantitative results demonstrate that EfficientNet-B0 outperformed other models, achieving a peak accuracy of 91.84% and a superior F1-Score of 0.8429, despite possessing the most compact parameter footprint of 4.02M. While ResNet-50 exhibited lower diagnostic precision, it offered the fastest inference speed (0.359 ms), highlighting a critical trade-off between accuracy and real-time operational latency. Furthermore, visual validation through Grad-CAM++ confirmed that successful predictions were driven by relevant morphological hallmarks rather than dataset artifacts. Our findings suggest that architectural optimization through compound scaling is more effective than raw model depth for dermatological tasks. Collectively, this work provides a comprehensive framework for selecting deep learning backbones for clinical triage, balancing high-precision diagnostic support with the computational constraints of real-world medical deployment.