融合纹理特征和注意力机制的异常脑MRI分割方法  

Abnormal Brain MRI Segmentation Method Combining Texture Features and Attention Mechanisms

作　　者：张付春 李盟 吴凉 王玉文 吴樾 ZHANG Fu-chun;LI Meng;WU Liang;WANG Yu-wen;WU Yue(School of Information Science and Engineering,Linyi University,Linyi 276005,China;School of Control Science and Engineering,Shandong University,Jinan 250061,China;Library,Linyi University,Linyi 276005,China)

机构地区：[1]临沂大学信息科学与工程学院,临沂276005 [2]山东大学控制科学与工程学院,济南250061 [3]临沂大学图书馆,临沂276005

出　　处：《印刷与数字媒体技术研究》2023年第4期203-211,共9页Printing and Digital Media Technology Study

摘　　要：异常脑磁共振图像(MRI)分割是临床应用的一个重要环节。目前,深度学习被广泛应用到异常脑图像分割任务中。然而由于异常脑结构复杂,肿瘤区域面积小,现有分割方法依然存在特征信息提取不充分、一些先验知识难以获得等问题。为了解决以上问题,本研究在U-Net网络的基础上,提出了一种双U-Net(DU-Net)分割模型。该模型首先将预处理后的MRI提取纹理特征,把提供额外边界信息的纹理特征图像和T1图像共同输入到DU-sub1网络中,其中双编码器子网络结合残差模块,并在解码过程中融入多尺度注意力机制模块进行特征还原,关注特征信息。最后通过DU-sub2网络将脑MRI分割为肿瘤部分、脑脊液、脑灰质和脑白质。DU-sub1和DU-sub2在异常脑MRI分割中分别起到粗略分割和精细分割的作用。在BraTS 2020数据集上进行了训练和测试,该模型分割结果中肿瘤部分、脑脊液、脑灰质和脑白质的平均DICE系数分别为0.831、0.917、0.905、0.911。该方法可以为临床诊断提供帮助,对后期的治疗具有一定的指导意义。Abnormal brain magnetic resonance image(MRI)segmentation is an important link in clinical application.Currently,deep learning is widely applied to abnormal brain image segmentation tasks.However,due to the complex structure of the abnormal brain and the small size of the tumor area,the existing segmentation methods still have problems such as insufficient feature information extraction and difficulty in obtaining some prior knowledge.To solve the above problems,a dual U-Net(DU-Net)segmentation model based on the U-Net network was proposed in this study.Firstly,the texture features from the preprocessed MRI were extracted by the model.Secondly,the texture feature image and the T1 image were inputted that provided additional boundary information into the DU-sub1 network,in which the dual-encoder sub-network combined the residual module,and in the decoding process integrated the multi-scale attention mechanism module to restore features and focused on feature information.Finally,the brain MRI was segmented into tumor parts,cerebrospinal fluid,gray matter,and white matter by the DU-sub2 network.DU-sub1 and DU-sub2 played the role of rough segmentation and fine segmentation in abnormal brain MRI segmentation,respectively.Trained and tested on the BraTS 2020 dataset,the average DICE coefficients of the tumor parts,cerebrospinal fluid,gray matter,and white matter in the segmentation results of this model were 0.831,0.917,0.905,and 0.911,respectively.This method can provide assistance for clinical diagnosis and had certain guiding significance for later treatment.

关 键 词：脑磁共振图像 图像分割 多尺度注意力机制 灰度共生矩阵 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TS893[自动化与计算机技术—计算机科学与技术]