Investigating self-supervised learning for Skin Lesion Classification

Skin cancer is one of the most common cancer worldwide, and is growing as a rising global health issue due to the damage of the natural protection from harmful ultraviolet radiation. Early diagnosis and proper treatment even for the deadliest malignant melanoma can greatly increase the survival rate. Thus, computer-aided diagnosis for skin lesions has been actively explored and made remarkable progress in medical practices benefiting from the the great advance of the deep convolution neural networks in vision tasks. However, most studies in skin lesion/cancer recognition and detection focus on reconstructing a robust prediction model with the annotated training samples in a fully-supervised manner, and cannot make full use of the available unlabeled data. This study investigates self-supervised learning using large amount of unlabeled skin lesion images to train a good initial network for representation learning, and transfer the knowledge of the initial model to the supervised skin lesion classification task with small number of annotated samples for enhancing the performance. Specifically, we employ a negative sample-free self-supervised framework by leveraging the interaction learning of the online and target networks for enforcing representative robustness with only positive samples. Moreover, according to the observation of the potential variations in the target skin images, we select the adaptive augmentation methods to produce the transformed positive views for self-supervised learning. Extensive experiments on two benchmark skin lesion datasets demonstrated that the proposed self-supervised pre-training can stably improve the recognition performance with different numbers of the labeled images compared with the baseline models.