Semi-supervised generative adversarial networks for imbalanced skin lesion diagnosis with an unbiased generator and informative images

Skin cancer remains a significant global health challenge, necessitating effective early detection methods. Traditional supervised learning approaches for skin lesion classification often require extensive labeled datasets, which are costly and time-consuming to obtain. This study addresses the limitations of semi-supervised learning in skin cancer diagnosis, particularly issues related to classification bias toward majority classes and the impact of incorrect pseudo-labels from low-confidence unlabeled images. We propose an unbiased bad generator within the weighted bad semi-supervised generative adversarial network (WB-SGAN), which integrates self-social consistency regularization and a new weighted inversed cross-entropy loss function for informative images. Additionally, a weighted cross-entropy loss function is used to reduce the bias of the classifier's predictions on labeled and unlabeled images. This framework enhances the generation of informative fake samples, thereby reducing bias in pseudo-labels and improving classification performance. Our experiments demonstrate that WB-SGAN outperforms existing state-of-the-art (SOTA) methods, achieving balanced accuracies of 80.35 % and 74.19 % on the ISIC-2018 and PAD-UFES datasets, respectively, even with just 5 % labeled data. This approach highlights the visual and volumetric aspects of training images and their labeling, offering a solution for skin lesion classification in semi-supervised domains with limited and imbalanced datasets.