Progressive Reduction Network (PRNet) for Accurate Polyp Segmentation in Colonoscopy Images

Colonoscopy is a crucial procedure for detecting and treating abnormalities in the colon, including polyps. However, accurately identifying and segmenting polyp regions in colonoscopy images remains a challenge due to their varying sizes and blurry boundaries. Existing models often fail to provide satisfactory segmentation accuracy, prompting the development of a new approach called Progressive Reduction Network (PRNet).

PRNet aims to locate and refine polyp regions step by step, utilizing Res2Net to extract features of the polyp region. These extracted features are then combined with a multi-scale cross-level fusion module to pinpoint the exact location of the polyp. This module enhances the network's ability to extract features from multi-scale polyp regions, thereby improving segmentation accuracy.

Furthermore, PRNet employs a module for filtering features to eliminate redundant and interfering data from the encoded features. An uncertain region processing module is also included to enhance edge detail feature recognition, progressively decreasing thresholds from top to bottom. Finally, a multi-scale context-aware module is designed to further explore potential semantic feature information.

The experimental results on the Kvasir-SEG, CVC-Clinic, and ETIS datasets demonstrate that PRNet achieves Dice coefficients of 92.09%, 93.05%, and 74.19%, respectively. PRNet exhibits significant improvements in segmentation accuracy for cases with multi-scale lesion areas and fuzzy lesion boundaries.

Overall, PRNet presents a promising approach for accurately identifying and segmenting polyp regions in colonoscopy images. Its multi-scale cross-level fusion module and other features allow for improved feature extraction and better segmentation accuracy. This development could lead to better detection and treatment of abnormalities in the colon, ultimately improving patient outcomes.