A dermoscopic image segmentation algorithm based on U-shaped architecture

The dermatological images have complex characteristics such as irregular shape, large scale variation, insufficient contrast and blurred boundary. In this paper, multiscale atrous convolutions pooling module (MSACP) is proposed to learn rich feature information of target regions from multiple perceptual fields, which can fully identify irregularly shaped regions and solve the problem of feature loss due to uneven scales. And attention refinement of contextual information (ARCI) is proposed to extract the shallow detailed feature information and pass it to the decoding part to solve the boundary blurring. Therefore, an improved u-net medical image segmentation algorithm based on MSACP and ARCI is proposed. The algorithm is compared experimentally at ISIC2018 and the results show that it outperforms other algorithms in dermatological medical image segmentation.