Trustworthy AI for Skin Cancer Image Classification

Abstract

Skin cancer constitutes a growing global health burden, with melanoma representing one of its highly aggressive and potentially lethal subtypes. Dermoscopy enhances early diagnosis by improving the visualisation of lesion structures. However, interpretation remains subjective and dependent on clinical expertise. Deep learning methods have demonstrated strong potential for automated dermoscopic image classification. However, many studies lack standardised evaluation protocols and transparent analysis of model behaviour. This study presents a systematic evaluation of multiple architectures for multi-class skin lesion classification using the official ISIC-2018 dataset splits. We analyse the impact of class imbalance handling and assess ensemble learning techniques. Explainability is achieved using Gradient-weighted Class Activation Mapping, which visualises the lesion regions that influence the model’s predictions and enables qualitative assessment of the model’s reliability. Experimental results demonstrate that EfficientNet-B5 achieved the best individual performance with 86% ± 0.008% accuracy and a macro F1-score of $0.79 \pm 0.02$, while the ensemble model improved overall accuracy to 87% ± 0.008% and macro F1-score to 0.80 ± 0.009, demonstrating improved robustness for minority lesion classes. Explainability analysis highlights clinically relevant attention patterns and common failure cases. This work promotes reproducible benchmarking and interpretable deep learning practices for dermoscopic image classification.