Abstract

Info

Title:: Colorectal polyp segmentation model based on frequency-aware and contextual network

Author(s):: JI Changpeng; LIANG Zheng; DAI Wei; School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China

Keywords:: Keywords: colorectal polyp segmentation frequency-aware and contextual network Transformer self-attention mechanism

Abstract:: Abstract: A frequency-aware and contextual network (FAC-Net) is proposed for colorectal polyp image segmentation to address the challenges posed by blurred boundaries and complex shapes that impede accurate localization and compromise segmentation accuracy. A Transformer encoder is employed to construct a feature pyramid for capturing global context through self-attention mechanisms. Subsequently, a global frequency-domain perception module is designed, in which wavelet transform is incorporated to use high-frequency components for boundary enhancement and low-frequency components for auxiliary localization. A multi-scale semantic enhancement module is further developed using grouped features and cross-space learning mechanisms to strengthen the models ability to capture spatial details in lesion areas. Finally, a cross-layer feature aggregation module is implemented with an attention-guided cross-layer fusion strategy to effectively integrate shallow detailed features and deep semantic features, thereby significantly improving segmentation precision. Experimental results on 5 benchmark datasets shows that the proposed model yields Dice coefficients of 0.927, 0.937, 0.808, 0.912, and 0.788 on the Kvasir-SEG, Clinic-DB, Colon-DB, CVC-300, and ETIS datasets, respectively, outperforming generic segmentation models. The evaluation results validate that the FAC-Net achieves superior segmentation accuracy and strong generalization ability.