融合Transformer与双图卷积的结直肠息肉分割算法

A colorectal polyp segmentation algorithm integrating Transformer and dual graph convolution

摘要:

针对结直肠息肉图像病灶区域定位不准和分割边界模糊等问题,本文提出一种融合Transformer与双图卷积(dual graph convolutional network,dual-GCN) 的结直肠息肉分割算法。首先利用Transformer编码器解析图像中不同尺度的特征信息,建立像素间的长距离依赖关系。其次设计双图卷积语义细节注入(dual-GCN semantic detail injection,DGSDI) 模块,融合深层特征的空间与结构信息,增强边界表示。然后采用局部金字塔注意力(local pyramid attention,LPA) 模块,对浅层特征进行全局和局部的注意力权重计算,精准定位病灶区域并抑制无用信息。最后引入动态特征融合(dynamic feature fusion,DFF) 模块自适应聚合不同尺度特征,增强对尺度变化大和形状不规则病理性图像的处理能力。本文算法在Kva- sir、CVC-ClinicDB、CVC-ColonDB和ETIS 4个公共数据集上的实验结果显示,Dice系数分别为0.924、0.943、0.817和0.813,平均交并比(mean intersection over union,mIoU) 分别为0.871、0.896、0.733和0.732,验证了该算法的有效性。

Abstract:

To address the challenges of inaccurate lesion region localization and blurred segmentation boundaries in colorectal polyp images,this paper proposes a colorectal polyp segmentation algorithm that integrates Transformer and dual graph convolutional network (dual-GCN) techniques.First,the Transformer encoder is utilized to parse multi-scale feature information within the images,establishing long-distance dependencies between pixels.Second,a dual graph convolution semantic detail injection (DGSDI) module is designed to integrate the spatial and structural information of deep features,enhancing boundary representation.Third,a local pyramid attention (LPA) module is employed to calculate global and local attention weights on shallow features,precisely localizing the lesion regions while suppressing irrelevant information.Finally,a dynamic feature fusion (DFF) module is introduced to adaptively aggregate multi-scale features,improving the handling of pathological images with large scale variations and irregular shapes.Experimental results on four public datasets,Kvasir,CVC-ClinicDB,CVC-ColonDB,and ETIS,show that the Dice coefficients are 0.924,0.943,0.817 and 0.813,respectively,and the mean intersection over union (mIoU) scores are 0.871,0.896,0.733 and 0.732,respectively,verifying the effectiveness of the proposed algorithm.