CSNet: Colonic Polyp Segmentation for Assisted Diagnosis in Colonoscopy

Accurate colonoscopic diagnosis is essential for the early diagnosis and treatment of colorectal cancer, where the identification and segmentation of the polyp lesion area is extremely important. Therefore, accurate and efficient polyp segmentation techniques are of great research and clinical interest. In recent years, most methods have been based on traditional U-shaped structures and feature extraction by stacking network layers. However, such operations generate a lot of redundant information and ignore the correlation between layers, resulting in inaccurate and ambiguous polyp segmentation. To address these issues, we propose a new framework, CSNet, for colonoscopic polyp segmentation. It consists of two components, Multi-scale Enhancement Module (ME) and Aggregation Enhancement Decoder (AED). The ME captures the differential feature information in the layers during downsampling through the proposed Enhancement Unit (EU), and uses pyramidal structure arrangement and Feature Aggregation Complementarity (FAC) to enhance the scale features of different layers with a view to obtaining local information and long-lasting dependency between layers. The AED is responsible for aggregating high-level feature information and generating global maps to capture global features. For qualitative and quantitative evaluation tests on five publicly available datasets, our CSNet achieved state-of-the-art results, with the Kvasir dataset achieving 91.62% and 85.76% for Dice and MIoU respectively, and 97.22% accuracy.