On The Impact Of Self-Supervised Learning In Skin Cancer Diagnosis

Deep neural networks (DNNs) are the standard approach for image classification. However, they require a large amount of data and corresponding annotations. Collecting medical data is a difficult task, due to privacy restrictions. Moreover, it is even harder to obtain the clinical labels, since these must be provided by specialists. Self-supervised learning (SSL) has emerged as a possibility to overcome this issue, since it uses non-annotated data to pre-train the DNN. Recently SSL has been applied in the context of skin cancer. However, the results were not conclusive. Moreover, a proper analysis of the impact of different SSL approaches is still missing. In this paper we investigate two SSL approaches: Rotation and SimCLR. Our results highlight the benefits of applying self-supervised learning to the classification of dermoscopy images. Additionally, we demonstrate that these approaches learn different and complementary features.