摘要
针对肠息肉图像分割结果存在伪影、边界模糊及精度低等问题,提出一种基于混合反向注意力机制的息肉分割网络。该网络使用U型网络结构,设计多尺度并行空洞卷积注意力模块,以更加细粒度的多尺度特征减少下采样细节的损失。采用密集连接和特征融合的方式设计跨阶段局部模块,减少上下文之间的语义差异,补充细节特征。利用位置注意力和通道注意力同反向注意力相结合策略,构建区域与边界关系的同时学习位置和通道的特征,进而清晰分割出息肉与正常粘膜。实验结果表明,该网络提高了分割精度,消除了边界外部的部分伪影,在一定程度上改善了边界模糊的问题。
To address the challenges of artifacts,blurred boundaries,and low accuracy in the segmentation results of polyp images,a polyp segmentation network based on hybrid reverse attention mechanism is proposed.The network incorporates a U-shaped structure and introduces a multi-scale parallel dilated convolution attention module.This module helps preserve finer-grained multi-scale features during down sampling,reducing the loss of important details.Additionally,dense connectivity and feature fusion are employed to cross stage partial module to bridge semantic differences between contexts and enhance detailed features.Furthermore,a combination of positional attention and channel attention,integrated with the inverse attention strategy,is employed to learn location and channel features while establishing the region-boundary relationship for accurate polyp and normal mucosa segmentation.Experimental results demonstrate that the polyp segmentation network based on hybrid reverse attention mechanism improves the segmentation accuracy,reduces artifacts outside the boundary,and mitigates the boundary blurring issues to a certain extent.
作者
兰蓉
孙宇浩
赵凤
郭迪
LAN Rong;SUN Yuhao;ZHAO Feng;GUO Di(School of Communications and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China)
出处
《西安邮电大学学报》
2024年第1期87-95,共9页
Journal of Xi’an University of Posts and Telecommunications
基金
国家自然科学基金项目(62071379,62106196)
陕西省自然科学基础研究计划项目(2021JM-461)
西安邮电大学西邮新星团队计划项目(xyt2016-01)。
关键词
图像分割
息肉分割
注意力
空洞卷积
U-Net
image segmentation
polyp segmentation
attention
dilated convolution
U-Net