CCJ-SLC: A Skin Lesion Image Classification Method Based on Contrastive Clustering and Jigsaw Puzzle

Self-supervised learning has been widely used in natural image classification, but it has been less applied in skin lesion image classification. The difficulty of existing self-supervised learning is the design of pretext tasks, intra-image pretext tasks and inter-image pretext tasks often focus on different features, so there is still space to improve the effect of single pretext tasks. We propose a skin lesion image classification method based on contrastive clustering and jigsaw puzzles from the perspective of designing a hybrid pretext task that combines intra-image pretext tasks with inter-image pretext tasks. The method can take advantage of the complementarity of high-level semantic features and low-level texture information in self-supervised learning to effectively solve the problems of class imbalance in skin lesion image datasets and large intra-class differences in skin lesion images with small inter-class differences, thus improving the performance of the model on downstream classification tasks. Experimental results show that our method can achieve supervised learning classification results and outperform other self-supervised learning classification models on the ISIC-2018 and ISIC-2019 datasets, moreover, it performs well on the evaluation metrics of unbalanced datasets.