中山大学学报自然科学版 ›› 2018, Vol. 57 ›› Issue (2): 52-60.

• 论文 • 上一篇    下一篇

基于支持张量机算法和3D脑白质图像的阿尔兹海默症诊断

徐盼盼1,杨宁2,李淑龙1   

  1. 1. 南方医科大学生物医学工程学院∥广东省医学图像处理重点实验室,广东 广州 510515;
    2. 广东省第二人民医院,广东 广州 510317
  • 收稿日期:2017-10-17 出版日期:2018-03-25 发布日期:2018-03-25
  • 通讯作者: 李淑龙(1981年生),女; 研究方向: 影像组学,医学图像分析;E-mail: shulong@smu.edu.cn

Diagnosis of Alzheimers disease based on Support Tensor  Machine and 3D brain white matter image

XU Panpan1, YANG Ning2, LI Shulong1   

  1. 1. School of Biomedical Engineering, Southern Medical University∥Guangdong Provincial Key Laboratory of Medical Image Processing, Guangzhou 510515, China;
    2. GuangDong No. 2 Provincial People's Hospital, Guangzhou 510317, China
  • Received:2017-10-17 Online:2018-03-25 Published:2018-03-25

摘要:

结构磁共振成像(sMRI)本质上具有三维张量结构,而传统的向量空间机器学习方法将其展开成向量进行建模,这破坏了数据的内在结构信息的完整性,降低了机器学习性能。为了克服数据向量化的弊端,提出了一种基于支持张量机(Support tensor machine, STM)的以3D T1加权MR脑白质图像为输入的阿尔兹海默症诊断算法。首先用SPM8软件将采集的MRI数据进行预处理,分割为灰质、白质、脑脊液3部分,提取脑白质各体素的灰度值构建三阶灰度张量,然后用递归特征消除(Recursive Feature Elimination, RFE)法结合支持张量机进行特征选择,最后用支持张量机进行分类。在阿尔兹海默症患者(AD),轻度认知障碍患者(MCI)(包括转化为AD的MCI-C和未转化的MCI-NC)以及正常对照(NC)4组人群中进行实验测试,并用10折交叉验证方法获得验证结果。用ROC曲线下面积AUC、分类准确率、敏感性、特异性这4个指标评价分类器的性能,AD vs NC组分别达到99.1%、 97.14%、95.71%、98.57%; AD vs MCI组分别达到88.29%、84.07%、78.57%、91.07%;MCI vs NC组分别达到89.18%、87.91%、93.75%、78.57%;MCIC vs MCINC组分别达到87.5%、82.08%、80.36%、82.14%。算法保持了原始图像的张量结构,提高了分类器的性能,实验结果表明此算法是一种有效的阿尔兹海默症诊断方法。

关键词: 阿尔兹海默症, 3D脑白质图像, T1加权 MRI, 递归特征消除, 支持张量机

Abstract:

Structural magnetic resonance imaging (SMRI) method has an intrinsic thirdorder tensor structure. Traditional vectorbased machine learning methods unfold the 3D images as vectors to carry on the modeling, which break the natural 3D structure of data so that some useful information underlying the neuroimaging data is missing. Therefore, a novel classification method based on the Support Tensor Machine (STM) is proposed to overcome these drawbacks. The T1 weighted MRI images are first preprocessed and segmented into gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF) regions using SPM8 tool. The third-order grey tensors are then constructed for each partition. Recursive feature elimination (RFE) method coupled with Support Tensor Machine (STM) is used to select the optimal features subset for classification using the STM-based classifier. The proposed algorithm perform the classification on four cases including the patients of Alzheimers disease and Mild Cognitive Impairment (including patients were converted to AD, MCI-C; and patients were not converted to AD, MCI-NC) and normal controls (NC), 10-fold cross validation is employed to assess the classification performance. In terms of AUC, classification accuracy, sensitivity and specificity, the case AD vs NC archive 99.1%, 97.14%, 95.71%, 98.57% respectively; the case AD vs MCI archive 8829%, 84.07%, 78.57%, 91.07% respectively; the case of MCI vs NC archive 89.18%, 87.91%, 93.75%, 78.57% respectively; and the case MCI-C vs MCI-NC archive 87.5%, 82.08%, 80.36%, 82.14% respectively The proposed method keeps the natural tensor structure of the original data and improves the performance of the classifier. The experimental results indicate that the proposed algorithm is effective for the diagnosis of Alzheimers disease.

Key words: Alzheimer's disease, 3D brain white matter image, T1 weighted MRI, Recursive feature elimination, Support Tensor Machine

中图分类号: