Pseudo Skin Image Generator (PSIG-Net): Ambiguity-free sample generation and outlier control for skin lesion classification

Automated skin lesion classification in dermoscopic images is essential for improving diagnostic performance and reducing melanoma-related deaths. Although deep learning models for image classification have made breakthroughs in classifying skin lesions with intra-class and inter-class similarity issues, some ambiguous features of original samples in the training data still remain unresolved. The ambiguity of the samples remains a challenge for the model's performance on overlapping training data. To address these challenges, we introduce a novel framework called Pseudo Skin Image Generator network (PSIG-Net). The framework not only removes the ambiguous samples but also generates new pseudo samples to enrich the training data. Our approach targets the low-variation samples by generating new pseudo samples using a generative adversarial network (GAN) model that emulates the characteristics of the original respective classes. We regulate the pseudo samples through a series of processing stages to control their similarity and exclude outliers. We employ a Siamese-based network to control the distance between these pseudo samples and the original clusters. The assessment of density-based distance is utilized to select only the closest relationships with the original samples. Through a series of experiments, we observed the significant improvement of our proposed PSIG network, achieving competitive results when compared to similar methods on two challenging skin lesion datasets.