EMFF-Net: Edge-Enhancement Multi-Scale Feature Fusion Network

Abstract

According to medical research, colorectal polyps are considered typical precancerous lesions, making colonoscopic polyp images crucial for the early diagnosis of rectal cancer. However, variations in polyp size and shape, texture inconsistencies, and boundary ambiguities often pose significant challenges for polyp segmentation. To address these issues, we propose a multi-scale feature fusion network based on edge enhancement. Specifically, we utilize multi-scale feature fusion in each feature layer, where the extracted features are fused. From these extracted features, we generate a global mapping graph as a bootstrap region. Additionally, we introduce Spatial Channel Convolution (SCEConv) and Reverse Gated Channel Transformer (RGCT) to incorporate boundary information into the segmentation network. This approach enhances the layered features and produces a more refined segmentation map. Extensive qualitative and quantitative experiments on five benchmark datasets and two private datasets demonstrate that the EMFF-Net proposed in this paper significantly improves segmentation accuracy across six metrics. This represents a clear advantage over traditional CNNs and existing SOTA techniques, Especially, we achieved 81% mDice and 74% mIoU on the CVC-ColonDB dataset.