Residual-Gated Attention U-Net with Channel Recalibration for Polyp Segmentation in Colonoscopy Images

Abstract

This study proposed a modification to the Attention U-Net architecture by integrating a Residual-Gated mechanism and Squeeze-and-Excitation (SE) Block-based channel recalibration within the Attention Gate to enhance feature selectivity in polyp segmentation. This integration reinforces both spatial and channel attention, enabling the model to better highlight polyp regions while suppressing irrelevant background features. Experiments were conducted on three colonoscopy datasets, CVC-ClinicDB, CVC-ColonDB, and CVC-300, using IoU and DSC metrics. Compared to the Attention U-Net baseline, the proposed model achieves noticeable improvements, with performance gains of mIoU 0.0043 and mDSC 0.0094 on CVC-ClinicDB, mIoU 0.0012 on CVC-ColonDB, and a larger margin of mIoU 0.0224 and mDSC 0.0127 on CVC-300. The best results were obtained on CVC-ClinicDB (mIoU 0.8889, mDSC 0.9412). Although the absolute scores on CVC-ClinicDB and CVC-ColonDB are lower than those reported in several recent studies, these datasets contain higher variability in polyp size, boundary ambiguity, and illumination, contributing to more challenging segmentation conditions. Visual evaluation further shows smoother and more coherent boundaries, especially on small or low-contrast polyps. Overall, the integration of the residual-gated mechanism and SE block within the attention gate effectively improves model accuracy and generalization, particularly in challenging scenarios.