MCGNet: Multi-level consistency guided polyp segmentation

Abstract

Polyp segmentation is challenging due to the varying shapes and sizes, and low contrast, resulting in blurred segmentation boundaries. To address this problem, we propose a multi-level consistency guided network (MCGNet), which performs joint supervision at three different levels: (i) at size level, we obtain two inputs with consistent content and different sizes by applying affine transformation to the input image; (ii) at boundary level, we utilize two modules, Multi-scale Attention (MA) and Feature Similarity Aggregation (FSA), to reinforce the boundary information and learn the boundary consistency in the output layer; (iii) at class activation map level, we follow a consistency regularization approach to restrict the range of class activation maps in the middle layer by Pixel Correlation Attention (PCA) module. Experimental results on 5 widely used datasets show that the MCGNet achieves state-of-the-art performance and exhibits outstanding generality.