Abstract
Computer-aided segmentation has been widely used in clinic since it helps to quickly locate the region of interest (ROI) in medical images. At present, a large number of segmentation algorithms have been proposed based on convolutional neural networks in past decades. However, they are difficult to popularize in segmentation of complex and diverse targets such as Intestinal polyps. In this paper, an efficient segmentation model of parallel multi-scale networks based on compound extended coding UNet++ is proposed. Specifically, a multi-scale parallel strategy is designed to extract multi-level features from multiple branches of the segmentation network for fusion, in which each branch network is constructed by combining EfficientNet encoder and UNet++ decoder. The proposed model is applied on the public CVC-ClinicDB dataset and achieves the mean Dice of 0.9344, IoU of 0.8440, sensitivity of 0.9113 and precision of 0.9211, exhibiting more promising performance than some state-of-the-art methods in polyp segmentation from colonoscopy images.