融合注意力机制的多模态脑肿瘤MR图像分割  

Multimodal Brain Tumor MR Image Segmentation Network Fused with Attention Mechanism

作　　者：毋小省[1] 杨奇鸿 唐朝生[1] 孙君顶[1] Wu Xiaosheng;Yang Qihong;Tang Chaosheng;Sun Junding(School of Computer Science&Technology,Henan Polytechnic University,Jiaozuo 454000)

机构地区：[1]河南理工大学计算机科学与技术学院,焦作454000

出　　处：《计算机辅助设计与图形学学报》2023年第9期1429-1438,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：河南省科技攻关项目(212102310084);河南省高等学校重点科研项目(22A520027).

摘　　要：针对在多模态MR图像分割中对不同模态特征间的关联性及全局和局部特征提取考虑不充分,导致分割精度降低的问题,基于注意力机制,提出多模态脑肿瘤MR图像分割方法.首先提出三重注意力模块,用于增强各模态特征间的关联性以及对感兴趣区域的位置和边界信息精确判断;然后设计空间和通道注意力模块,用于双重捕获空间和通道上的全局及局部特征,增强对肿瘤组织结构信息的学习能力.在公开数据集BraTs18和BraTs19上的实验结果表明,分割全肿瘤时,所提方法的Dice系数、精确率、灵敏度和Hausdorff距离分别达到了90.62%,87.89%,90.08%和2.2583,均优于对比的同类方法.For the traditional multimodal MR image segmentation methods,the correlation between different modal features,the global and local features were not fully considered,which leads to the reduction of segmenta-tion accuracy.To solve such problem,a multimodal brain tumor MR image segmentation method was proposed based on the attention mechanism.Firstly,a triple attention module was proposed to enhance the correlation be-tween the modal features and to accurately judge the position and boundary information of the region of interest.Secondly,the spatial and channel attention module was designed to capture the global and local features of the space and channel,and enhance the learning ability of tumor tissue structure information.The experimental re-sults on the public datasets BraTs18 and BraTs19 show that the method achieves 90.62%,87.89%,90.08%and 2.2583 in the Dice coefficient,precision,sensitivity and Hausdorff distance when segmenting the whole tumor,respectively,which are better than similar methods in comparison.

关 键 词：多模态图像 脑肿瘤分割 注意力机制 三重注意力 空间和通道注意力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]