A Hybrid Approach to Mitigating Class Imbalance in Skin Cancer Classification using Residual Multihead Convolutional Neural Networks

Skin cancer is one of the most common cancers, characterized by a wide variety of lesion textures. Accurate classification is crucial for early detection and improving patient survival rates. The HAM100 dataset was released to address key challenges in the automated diagnosis of skin lesions; however, this dataset faces the challenge of extreme class imbalance, which introduces bias in conventional models. This study proposes the Residual Multihead Convolutional Neural Network (ResMHCNN) architecture, which has been successfully implemented in previous research to address the class imbalance problem. The ResMHCNN architecture uses multiscale features and channel attention for feature extraction with a lightweight model. Three variants of ResMHCNN - LCCnet, BTnet, and BCnet - were evaluated across four scenarios: baseline, oversampling with augmentation, application of Focal Loss, and a hybrid combination of oversampling, augmentation, and Focal Loss. The hybrid strategy yielded the most significant improvement, increasing BTnet's recall from 0.273 to 0.466 and achieving the highest F1-Score of 0.444. These results demonstrate that the combination of oversampling, augmentation, and Focal Loss is effective for skin cancer screening, minimizes false negatives, and mitigates the impact of class imbalance.