W-net and inception residual network for skin lesion segmentation and classification

摘要

Melanoma is a serious skin disease. Automatic recognition of this lesion by dermoscopic images is a difficult task. Recently, the deep learning paradigm has become a good alternative to detect different types of diseases. In this article, we propose a new deep learning system for melanoma detection. The system consists of three steps: pre-processing, segmentation and classification. The particularity of this work is the introduction of two new deep learning network architectures, W-net and Inception-Resnet, to solve, respectively, the problem of segmentation and the problem of classification. W-net is composed of a ResNet Encoder-Decoder, a ConvNet Encoder-Decoder and a Feature Pyramid network. The use of two concatenated encoder-decoder architectures has significantly improved the segmentation results. Inception-Resnet includes an Inception-Resnet block, which is a fusion of inception with the residual neural network. With this architecture, classification is much more robust. We evaluated the proposed system on the PH2 dataset and on the International Skin Imaging Collaboration datasets (ISIC 2016, ISIC 2017 and ISIC 2018). The results are discussed in terms of accuracy, sensitivity, specificity, dice coefficient and precision. The comparison of the proposed approach with other related work confirmed the advantages of our technique.