Progressive Reduction Network (PRNet) for Accurate Polyp Segmentation with Multi-Scale and Fuzzy Boundaries

Existing models for polyp segmentation often struggle with the varying sizes and fuzzy boundaries of polyp regions in images, resulting in low segmentation accuracy. To address this challenge, this paper presents a novel Progressive Reduction Network (PRNet) that utilizes a step-by-step approach to precisely locate and refine polyp regions. PRNet leverages the power of Res2Net to extract features from polyp regions and incorporates a multi-scale cross-level fusion module for accurate initial localization. This strategy enhances the network's ability to handle multi-scale polyp regions effectively. To further improve segmentation performance, PRNet incorporates a feature filtering module to eliminate redundant and interfering information. This module enhances the clarity and precision of extracted features. Additionally, an uncertain region processing module is designed to refine edge details by strategically applying decreasing thresholds from top to bottom. This approach enhances the network's capability to identify subtle boundaries and improve overall segmentation accuracy. Finally, a multi-scale context-aware module is incorporated to explore potential semantic feature information and provide a more comprehensive understanding of the polyp region. This module helps the network to capture contextual cues and further refine the segmentation results. Extensive experiments conducted on Kvasir-SEG, CVC-Clinic, and ETIS datasets demonstrate the effectiveness of PRNet. The proposed algorithm achieves impressive Dice coefficients of 92.09%, 93.05%, and 74.19%, respectively. Notably, PRNet significantly outperforms existing methods in segmenting polyp regions with varying sizes and fuzzy boundaries, showcasing its ability to handle challenging scenarios. The findings highlight the potential of PRNet as a robust and accurate tool for polyp segmentation, paving the way for improved polyp detection and diagnosis in clinical settings.