Accurate detection and segmentation of polyps during colonoscopy are of great significance for the early prevention and treatment of colorectal cancer. However, due to the considerable variations in polyp size and shape, as well as their blurred boundaries with surrounding tissues, polyps are often difficult to detect, making precise segmentation a challenging task. Although numerous deep learning (DL) based segmentation methods have been proposed in recent years and achieved certain progress, their results remain unstable and often unsatisfactory. To address these challenges, we propose PGMNet, an accurate and efficient network for polyp segmentation, which consists of a PVTv2 encoder, a Global-Local Interactive Relation Module (GLIRM), and a Multi-stage Feature Aggregation Module (MFAM). The PVTv2 encoder is capable of capturing both fine-grained details and global semantic representations, making it well-suited for complex medical image segmentation tasks. GLIRM performs multi-scale information fusion during upsampling to restore fine-grained details and global semantic context, while simultaneously introducing a bit-slice mechanism to effectively suppress noise. MFAM leverages a gating mechanism to efficiently aggregate GLIRM information from different stages, thereby improving the quality of the final predictions.Extensive experiments were conducted on five publicly available polyp datasets, and the results demonstrate that PGMNet achieved very promising performance in terms of segmentation accuracy and generalization ability. In particular, on the challenging ETIS dataset, PGMNet achieved an mDice of 82.33% and an mIoU of 74.29%, highlighting its superior performance.
Keywords: bit plane slicing; medical image segmentation; polyp segmentation; transformer.
© 2026 IOP Publishing Ltd. All rights, including for text and data mining, AI training, and similar technologies, are reserved.