Unlabeled skin lesion classification by self-supervised topology clustering network

Abstract

As one of the most widespread cancer, skin cancer can be initially diagnosed by visual observation, following up with a series clinical check by dermoscopic analysis, histopathological assessment, and a biopsy. The initial visual observation provides the possibility of using artificial intelligence, which is confronted with various challenges on account of the discrepancy between different skin images. Though some deep learning methods have achieved improvements on this task, they needs a large mount of skin images annotated by well-trained professional doctors, which is very expensive to label so much data in reality. Furthermore, they maybe deteriorated into a new category with an unseen disease symptom, and existing skin lesion classification methods can not solve this problem. Here, this paper proposes a Self-supervised Topology Clustering Network (STCN) by a transformation-invariant network with self-supervised maximum modularity clustering algorithm following topology analysis principle. This approach can automatically classify unlabeled medical images without prior class number, and we implement sufficient validating experiments to certify that our STCN model can effectively solve the unlabeled medical image classification task.