DGPS-Net: Domain Generalizable Polyp Segmentation Model

초록

Colonoscopy is a screening technique to identify and eliminate colorectal polyps before they advance into cancerous tumors. While colonoscopy remains the gold standard for polyp detection, it is not without limitations, including a confined time window and susceptibility to human errors. Numerous deep learning models have emerged to aid and assist doctors in detecting polyps during procedures. However, polyp segmentation task can be challenging due to polyps' variation in shape, color and size and their similarity to the surrounding mucosal area. To address this challenge, we propose DGPS-Net (Domain Generalizable Polyp Segmentation Network) to effectively detect polyps that vary in shapes by efficiently fusing feature maps produced with different receptive fields. Additionally, we utilized attention modules including self-attention to attend to long-range pixel relations and enhance boundary detection. Our network demonstrated exceptional generalizability with the average improvements on F1-score, Recall, and mIoU by 2.71%, 2.49%, 2.69%, when tested on four datasets: Kvasir-SEG, CVC-ClinicDB, CVC-ColonicDB, and ETIS-LaribPolypDB.