DEMF-Net: A dual encoder multi-scale feature fusion network for polyp segmentation

Abstract

Colorectal cancer is a common malignant tumour of the gastrointestinal tract. Studies have shown that colonoscopy can be an effective screening method for detecting colon polyps and removing them to prevent the development of colorectal cancer. In this study, we propose a new approach called the Dual Encoder Multi-Scale Feature Fusion Network (DEMF-Net). This approach uses a dual-scale Swin Transformer and CNN as an encoder to extract semantic features at different scales. In order to enhance the marginal characteristics of irregular polyps and improve the polyp detection rate, we propose a Dual-Branch Attention Fusion Module (DAF) that captures different shapes of target features through the attention mechanism and assigns higher weights to feature channels with high contributions. Additionally, we use an Advanced Feature Fusion Module (AFFM) to establish long-range dependencies and strengthen the target region to ensure that the high-level semantic features of polyps are not lost. We also propose Characterization Supplementary Blocks (CSB) for colorectal polyp images with irregular shapes and unclear boundaries to capture the structure and details of images and enhance model accuracy. We conducted experiments on five widely adopted polyp datasets and showed that our method achieved superior results in terms of both segmentation accuracy and edge details.