Cross-Level Context Fusion Network for Polyp Segmentation in Colonoscopy Images

Medical image analysis, particularly polyp segmentation, plays a pivotal role in the early detection and prevention of colorectal cancer. The accuracy and precision of polyp identification significantly influence the subsequent diagnostic conclusions and therapeutic strategies. However, polyp segmentation still grapples with several challenges such as considerable variations in polyp size, shape, color, and location, and a high degree of visual similarity between polyps and their immediate tissue surroundings, attributable to factors like light reflection and motion blur during the capture of colonoscopy images. In this paper, we propose a novel Cross-level Context Fusion Network (CCFNet) for polyp segmentation within colonoscopy images. This network capitalizes on cross-level and multi-scale contextual information effectively, thereby enhancing segmentation performance significantly. Within the proposed framework, a High-level Feature Cascaded (HFC) module is presented to integrate the high-level features to produce a coarse segmentation map. This map establishes a global relationship for each pixel, which aids in accurately locating the polyps. In addition, a Cross-level Integration Module (CIM) is proposed to fuse the cross-level features to capture the complementary information from the adjacent layers. Consequently, the module extracts and fuses multi-scale features to learn rich feature representations. Moreover, we propose a Global Context Enhancement (GCE) module to utilize the global map to augment feature representations inside the decoder network. These enhanced features are then harnessed to construct multiple side-out segmentation maps. Extensive experimental results on five publicly polyp segmentation datasets demonstrate that our CCFNet surpasses other comparable methods in improving the accuracy of polyp segmentation.