Decision Support System for Skin Lesion Diagnosis Using Deep Learning

Skin lesion and cancer is a prevalent global health issue, and still continually escalate due to the damage of the natural defense against harmful ultraviolet radiation. Early-stage detection and effective treatment significantly enhance the survival rates, even for the most lethal malignant melanoma. The development of computer-aided diagnosis for skin lesions and cancers to support clinical doctors’ decision has been widely explored, and the incorporation of the recent advanced deep learning has demonstrated substantial progress in medical practices. This chapter presents our recent deep learning-based works for skin lesion classification and recognition. Firstly, we integrate the self-supervised learning method to pre-tarin an initial model by leveraging the large amount of available unlabeled data, and then further construct the skin lesion classification model using the conventional full-supervised deep learning procedure with small scale of labeled training samples. Benefitting from the maintained knowledge in the pretrained network, the finally constructed model even with a limited number of annotated samples, can be expected to produce promising classification performance. Secondly, motivated by the great success of the transformer as an alternative tool of convolution operation in the vision tasks in the recent years, we propose hybrid convolution and transformer network to make full use of the respective advantages of both convolution and transformer for improving the skin lesion classification performance. Specifically, we devise an enhance transformer module by incorporation the encoded position information to both query and key vectors before by performing self-attention for balance learning, and implement a position enhanced transformer block. By incorporating the position enhanced transformer block into the late-stage of the ResNet, we propose a novel bottleneck transformer model for skin lesion diagnosis. Extensive experiments on several benchmark skin lesion datasets have been conducted and demonstrated our proposed methods achieve superior classification performance over the existing deep learning-based methods.