A Modified U-Net for Skin Lesion Segmentation

In this paper, for skin lesion segmentation, we propose an encoder-decoder structure based on U-Net, combining dilated convolution and pyramid pooling module (PPM). The dilated convolution computes the feature maps with a high spatial resolution instead to down-sampling feature maps, and the aim of pyramid pooling module is to obtain more contextual information (multi-scale context information with multi-scale pooling). on the official test set of ISBI 2016, and in terms of three evaluation metrics, Our proposed model is tested and achieved better performance over U-Net and another published method with (JC=82.7, DC=89.6, SE =92.0).