基于U-Net的特征交互分割方法  

U-Net Based Feature Interaction Segmentation Method

作　　者：孙君顶[1,2] 惠朕堃 唐朝生[1] 毋小省[1,2] SUN Junding;HUI Zhenkun;TANG Chaosheng;WU Xiaosheng(College of Computer Science and Technology,Henan Polytechnic University,Jiaozuo 454000;Provincial Key Laboratory for Computer Information Processing Technology,Soochow University,Suzhou 215006)

机构地区：[1]河南理工大学计算机科学与技术学院,焦作454000 [2]苏州大学江苏省计算机信息处理技术重点实验室,苏州215006

出　　处：《模式识别与人工智能》2021年第11期1058-1068,共11页Pattern Recognition and Artificial Intelligence

基　　金：河南省科技厅科技攻关项目(No.212102310084);河南省高等学校重点科研项目(No.22A520027);苏州大学江苏省计算机信息处理技术重点实验室开放课题项目(No.KJS2048)资助。

摘　　要：针对肝脏分割中存在误分割及小目标漏分割的问题,文中提出基于U-Net的特征交互分割方法,采用ResNet34作为主干网络.为了实现不同尺度间的非局部交互,设计基于转换器机制的特征交互金字塔模块作为网络的桥接器,获得具有丰富上下文信息的特征图.设计多尺度注意力机制替代U-Net中的跳跃连接,关注图像中的小目标,充分获取目标层的上下文信息.在公开数据集LiTS及3Dircadb和CHAOS组成的数据集上的实验证实文中方法能取得较好的分割效果.To address the problems of mis-segmentation and missing segmentation of small targets in liver segmentation,a U-Net based feature interaction segmentation method is proposed using ResNet34 as the backbone network.To achieve non-local interactions between different scales,a transformer-based feature interaction pyramid module is designed as the bridge of the network to obtain feature maps with richer contextual information.A multi-scale attention mechanism is designed to replace the jumping connection in U-Net,considering the small targets in the image and sufficiently acquiring the contextual information of the target layer.Experiments on the public dataset LiTS and the dataset consisting of 3Dircadb and CHAOS demonstrate that the proposed method achieves good segmentation results.

关 键 词：肝脏分割 转换器 注意力机制 特征交互 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]