A CMT-Based Approach for Skin Lesion Segmentation

The problem of skin lesion segmentation is addressed in this paper. Convolutional neural networks (CNNs) have be applied to deal with the image segmentation problem. However, the CNN-based image segmentation techniques may suffer from the problems of poor generalization and difficulty in capturing long-distance dependencies. In another hand, Transformer has the advantage of a flexible structure, but its ability to learn positional encoding of pixels in an image is often limited when facing fewer medical images. In this paper, a CNNs meets Transformer (CMT)-based approach is proposed for the skin lesion segmentation. The proposed approach's performance is scrutinized and juxtaposed against CNN-based methodologies (e.g., the InceptionNet-based, the ResNet-based, and the DenseNet-based) via extensive experiments on the ISIC2018 dataset. Compared to CNN-based methods, the results reveal that the proposed CMT-based approach enhances the skin lesion segmentation accuracy by a margin exceeding 8%.