Robust skin lesion classification through enhanced minority class recognition using hybrid EfficientNet–Capsule networks

Skin lesion image classification is a challenging problem for conventional Convolutional Neural Network (CNN)-based methods because of the intricate interplay of spatial and semantic features. Because skin lesion datasets are imbalanced, models cannot learn minority-class features efficiently. In multiclass skin cancer classification, diagnostic categories vary widely, making precise classification difficult. This study is the first to utilize EfficientNet-B0 with Capsule Networks (CapsNets) on a combined dataset of HAM10000 and ISIC for the classification of skin cancer images. In this study, we present EfficientCapsNet, a hybrid deep learning architecture designed to classify multiple types of dermoscopic skin lesions using multi-class images. We propose using EfficientNet-B0 to extract lesion features and combining it with CapsNets to preserve their spatial relationships. Based on experimental results, EfficientCapsNet performed well across multiple evaluation metrics with an overall accuracy of 72.75%. Despite the inherent challenges of skin lesion classification, our approach highlights its potential to advance dermatologists' use of image-based diagnosis.