基于双路特征多尺度减法的结肠息肉分割研究
Research on colon polyp segmentation based on dual path feature multi-scale subtraction
熊炜 1张丽真 2杨茜 2孟圣哲 2李利荣2
作者信息
- 1. 湖北工业大学电气与电子工程学院,湖北武汉 430068;美国南卡罗来纳大学计算机科学与工程系,南卡哥伦比亚29201
- 2. 湖北工业大学电气与电子工程学院,湖北武汉 430068
- 折叠
摘要
针对结肠息肉大小差异较大、边界不明确、位置分布较散的问题,提出一种基于双路特征多尺度减法的结肠息肉分割方法.主支路通过重构减法单元与注意力模型进行相邻特征图的融合,加强息肉的边界信息以及息肉特征的提取能力,同时引入可学习的视觉中心(learnable visual center,LVC)来聚合输入图像的局部角落关键区域;副支路设计多尺度提取模块与倒置残差上采样(conv-transpose upsample,CT-upsample)模块融合而成的 AGG 模块(aggregation,AGG)进行多尺度大小息肉提取,还原及补充更多的细节信息.提出的方法在4个公共数据集上进行了实验分析,实验结果表明该方法具有良好的息肉分割泛化性能,其中在CVC-ClinicDB数据集上,mDice 和 mIoU 分别达到了 93.28％和 88.98％.
Abstract
A polyp segmentation method based on dual path feature multi-scale subtraction is proposed to address the issues of significant size differences,unclear boundaries,and scattered distribution of colon polyps.The main branch merges adjacent feature maps by reconstruction subtraction units and attention models,enhancing the boundary information of polyps and the ability to extract polyp features.Simultaneously,a learnable visual center(LVC)is introduced to aggregate local corner key regions of the input image.In the sub-branch,a multi-scale extraction module and a conv-transpose sample module are designed to fuse into an aggregation module(AGG)for multi-scale size polyp extraction,restoring and supplementing more detailed information.The proposed method is experimentally analyzed on four public datasets,and the experimental results show that our method has good generalization performance on polyp segmentation,with mDice and mIoU achieving 93.28％and 88.98％on the CVC-ClinicDB dataset,respectively.
关键词
结肠息肉分割/多尺度减法/可学习的视觉中心(LVC)/倒置残差上采样(CT-upsample)Key words
colon polyp segmentation/multi-scale subtraction/learnable visual center(LVC)/conv-transpose upsample(CT-upsample)引用本文复制引用
出版年
2025
NETL
NSTL
万方数据