IRAPU-Net: Improved Residual Atrous Convolutional Parallel Unit Network with Boundary Recalibration for Colorectal Polyp Segmentation

Accurate colorectal polyp segmentation is crucial for early screening and prevention of colorectal cancer. However, mainstream segmentation methods currently face two core challenges in polyp scenarios: Transformer-based models excel at modeling global context but are insensitive to local details, while CNN-based models perform well in extracting local features yet struggle to capture long-range dependencies due to their limited receptive fields, resulting in insufficient segmentation accuracy for ambiguous polyp boundaries. To address these issues, this paper proposes a novel colorectal polyp segmentation method based on IRAPU-Net (Improved Residual Atrous Convolutional Parallel Unit Network with Boundary Recalibration for Colorectal Polyp Segmentation). The method adopts CAFormer as the Transformer backbone for global context extraction. To resolve the misalignment between global semantics and local details during the feature fusion stage, we design a new Improved Residual Atrous Parallel Convolutional Unit (IRAP), which decomposes the improved Residual Atrous Parallel Convolutional Unit (RAPU) into local-detail and multi-scale-context functional branches and deeply couples them with an incremental scale fusion (ISF) mechanism, thereby enhancing salient local features while improving the model’s multi-scale representation capability. In addition, auxiliary components such as a boundary recalibration (BRC) module and a cross-feature fusion (CFF) module are introduced to specifically optimize multi-scale feature fusion and boundary refinement. Experiments conducted on five publicly available datasets widely used in polyp segmentation, including Kvasir-SEG and CVC-ClinicDB, demonstrate that the proposed IRAPU-Net achieves favorable segmentation performance in terms of DSC, IoU, and other evaluation metrics. All principal results are reported as the mean and standard deviation over five random seeds, and the main comparisons are supported by a paired Wilcoxon signed-rank test. Moreover, it shows competitive and robust cross-dataset generalization and boundary preservation, providing a reliable reference for clinical computer-aided diagnosis.