PRFNet: Progressive Region Focusing Network for Polyp Segmentation

In clinical practice, colonoscopy serves as an efficacious approach to detect colonic polyps and aids in the early diagnosis of colon cancer. However, the precise segmentation of polyps poses a challenge due to variability in size and shape, indistinct boundaries, and similar feature representations with healthy tissue. To address these issues, we propose a concise yet very effective progressive region focusing network (PRFNet) that leverages progressive training to iteratively refine segmentation results. Specifically, PRFNet shares encoder parameters and partitions the feature learning process of decoder into various stages, enabling the aggregation of features at different granularities through cross-stage skip connections and progressively mining the detailed features of lesion regions at different granularities. In addition, we introduce a lightweight adaptive region focusing (ARF) module, empowering the network to mask the non-lesion region and focus on mining lesion region features. Extensive experiments have been conducted on several public polyp segmentation datasets, where PRFNet demonstrated competitive segmentation results compared to state-of-the-art polyp segmentation methods. Furthermore, we set up multiple cross-dataset training and testing experiments, substantiating the superior generalization performance of PRFNet.