PISeg: Polyp Instance Segmentation with Texture Denoising and Adaptive Region

Polyp shape is one of the critical factors in helping categorize colorectal polyps and stratify them according to the risk of colorectal cancer. Unfortunately, there is no data on the recognition of pedunculated and sessile polyps, with very few studies conducted. In this work, we present advanced annotations that benchmark the task of polyp segmentation at the instance level. In addition, we propose a transformer-based Polyp Instance Segmentation network (PISeg) with two novel modules, namely, Texture Denoising and Adaptive Region. The first module aims to remove jamming textures on the polyp surface while also revealing essential polyp features. Meanwhile, the second one enhances the ability to detect regions containing polyp-specific characteristics and, hence, precisely determine each polyp instance’s overall scope. Comprehensive experiments and ablation studies show that our proposed modules help improve the overall performance and outperform state-of-the-art methods in polyp instance and semantic segmentation tasks.