CRCNet: Global-local context and multi-modality cross attention for polyp segmentation

Abstract

Accurate polyp segmentation is important for the diagnosis and treatment of colon cancer. In recent years, efforts have been made to improve the encoder-decoder framework by using global features and attention mechanisms to enhance feature extraction and help improve the segmentation of diverse polyps. However, few studies have considered the impacts of the polyp size, texture, and complex pathological environments on the segmentation performance. Considering the above challenges, this paper proposes a global–local feature-based encoder-decoder framework, named CRCNet comprising two components: a global–local context module (GLCM) and multi-modality cross attention (MMCA). The GLCM is responsible for capturing global and local information from all deep encoders, enabling accurate weighting of the context feature information for each region in the pathological image. The MMCA is in charge of adding background, boundary, and foreground factors for judgment when merging shallow features while paying more attention to doubtful and complicated regions. We conducted extensive experiments on the Kvasir-SEG and CVC-ClinicDB datasets, CRCNet achieved state-of-the-art results in terms of segmentation accuracy and computational efficiency, with Dice and MIoU of 91.59 % and 90.57 % for Kvasir-SEG, respectively, and 95.02 % and 94.48 % for CVC-ClinicDB, respectively. Thus, CRCNet shows a significant improvement over the state-of-the-art method. The corresponding code is available at: https://github.com/1152067715/CRCNet.