Automated and quantitative segmentation of colorectal polyps via an information-guided deep learning network with boundary perception.

Background: Although remarkable progress has been achieved in deep learning-based approaches for polyp segmentation, accurate delineation remains a significant challenge due to the intrinsic boundary ambiguity and complex morphological variations of polyps. These factors often lead to imprecise localization and incomplete regional segmentation, thereby degrading overall segmentation performance. To address these limitations, we developed an information-guided deep learning network (IGBP-Net), a novel framework driven by guidance linking spatial localization and boundary perception. Methods: In IGBP-Net, boundary extraction is not treated as an isolated task, and an information aggregation module (AGG) is used to establish initial localization. These localization cues explicitly guide a boundary-perception module (BPM) equipped with multiscale strip convolutions, which can effectively capture the fine-grained details specific to the irregular shapes of polyps. To support this core interaction, intermediate and deep representations are contextually refined via global-local and channel multiscale feature extraction. Finally, a region fusion module (RFM) determines the correlation between boundary and region features, iteratively reinforcing highly uncertain areas to produce the final segmentation maps. Results: Extensive comparative experiments on four benchmark datasets [Kvasir, Computer Vision Center Clinic Database (CVC-ClinicDB), Computer Vision Center Colon Database (CVC-ColonDB), and ETIS-Larib Polyp DB (ETIS)] demonstrated the superiority of IGBP-Net over advanced methods, achieving an average improvement in Dice score of 2.05% and in intersection over union of 2.63%. Specifically, it yielded a Dice score of 0.9326±0.0196 on Kvasir and 0.9452±0.0115 on CVC-ClinicDB. These performance gains were confirmed to be statistically significant according to the Wilcoxon signed-rank test (all P<0.05; P<0.001 on ETIS), with narrower 95% confidence intervals indicating lower experimental variance and higher stability. Qualitative comparisons further revealed that the proposed model can generate highly accurate segmentation maps for polyps with complex boundaries, closely matching the ground truth. Conclusions: These results validate the effectiveness and robustness of the proposed IGBP-Net, which represents a highly competitive solution for overcoming boundary ambiguity in automated polyp segmentation.