基于支持张量机算法和T1-Weighted MRI的阿尔兹海默症诊断方法研究

发布时间：2018-05-20 11:07

本文选题：阿尔兹海默症 + 轻度认知障碍　；参考：《南方医科大学》2017年硕士论文

【摘要】：阿尔兹海默症(Alzheimer's Disease,AD)是一种起病隐匿且不可逆的神经系统退行性疾病,其多发于65岁以上的人群,是当今世界最为普遍的一种痴呆症。2016年,全球痴呆患者人数已达4700万人,其中50%-75%为阿尔兹海默症患者。目前,中国的阿尔兹海默症患者人数已居世界第一,同时中国也是全球增速最快的国家之一。然而,阿尔兹海默症的诊疗率却与发病情况呈强烈反差,全球阿尔兹海默症及其它类型痴呆患者中,仅有22%接受过诊断。在中国这个比例更低,有49%的病例被误认为是自然老化,仅21%的患者得到了规范诊断,仅19.6%接受了药物治疗。而且就目前的医疗水平,阿尔兹海默症是一种无法治愈的疾病。因此,对阿尔兹海默症进行早期诊断,早期预防与干预治疗是非常重要的。轻度认知障碍(Mild Cognitive Impairment,MCI),是前人所提出的阿尔兹海默症的一种前驱状态,是介于正常衰老与痴呆之间的一种中间状态。轻度认知障碍可作为阿尔兹海默症的“预报器”,如果能及早发现此状态并给予适当的干预治疗,就可以延缓阿尔兹海默症的进展。所以,正确诊断阿尔兹海默症,尤其是正确诊断其早期阶段的轻度认知障碍,对阿尔兹海默症的预防、早期发现与治疗干预至关重要。为了识别阿尔兹海默症与轻度认知障碍患者,本文提出了一种基于支持张量机(STM)的分类器,以T1加权MRI脑图像灰质灰度为特征的诊断方法。该分类器以三维(3D)的脑灰质图像作为模型输入,用STM迭代算法训练分类器每一模的权向量进而进行分类。采集了 70例AD患者,112例MCI患者(包含在随访中转化为AD的,MCI-C:MCI Converters与未转化为AD的,MCI-NC:MCI Non-converters 各 56 例),以及 70 例正常人(NC)的 T1-Weighted MRI 三维(3D)脑图像。首先提取每个脑图像的灰质来构造每个脑图像的三阶灰质张量,张量大小为95×119×102。采用张量主成分分析法(TPCA)取得三阶灰质张量的低维的主成分张量,并以此主成分张量作为基于STM的分类器的输入进行分类(STM-TPCA)。张量独立成分分析(TICA)也用来提取出三阶灰质张量的独立成分张量以作为基于STM的分类器的输入进行分类(STM-TICA)。考虑到特征之间存在冗余性,因此在支持张量机迭代算法将张量特征转化为向量特征后,递归特征消除法(RFE)用来做特征选择,获得最优特征子集作为分类器的输入进行分类(STM-RFE)。最后,对四组人群进行分类:ADNC,MCINC,ADMCI,MCI-CMCI-NC,此分类模型采用10折交叉验证的方法进行训练测试。对于AD与NC的分类,其正确率最高可达91.19%(敏感性92.86%,特异性89.52%);对于MCI与NC的分类,其正确率最高可达83.15%(敏感性91.67%,特异性69.52%);对于AD与MCI的分类,其正确率最高可达82.23%(敏感性65.71%,特异性92.56%);对于MCI-C与MCI-NC的分类,其正确率最高可达77.08%(敏感性77.38%,特异性76.79%)。此外,本文还结合样本的基本信息(年龄、性别、教育程度)与认知分数(Mini-Mental State Exam,MMSE 分数;Alzheimer's Disease Assessment Scale-cognitive subscale,ADAS-cog分数)进行分类,结果发现结合基本信息与认知分数后分类效果能进一步提升,且对比于Shen与Willette结合多模态数据来作为模型输入的研究,本文方法的分类效果皆更为优异。以上实验结果表明以T1加权MRI脑图像的灰质图像为特征的基于STM的分类器是一种有效的阿尔兹海默症诊断方法;并且基本信息,认知分数与MRI脑灰质图像是相容的,具有很好的互补作用。在实验的过程中,我们发现由于高的张量维数(95×119×102),张量独立成分分析和递归特征消除法的运行速度都比较缓慢,高维特征很大程度上提高了特征提取和特征选择的时间。因此,我们考虑基于13个方向4种距离的灰度共生矩阵(GLCM)的纹理特征(Texture feature)张量(12×13×4)作为基于STM的分类器的输入进行分类(STM-Texture)。此改进方法减少了输入样本张量的维数,从而提升了整个分类模型的运行速度。实验结果表明使用灰度共生矩阵的纹理特征张量作为基于STM的分类器的输入的分类方法,既能保持原本分类方法的优越性,同时也减少了运行时间。
[Abstract]:Alzheimer's Disease (AD) is an insidious and irreversible neurodegenerative disease. It is more prevalent in people over 65 years of age, and is the most common type of dementia in the world today.2016. The number of people with dementia in the world has reached 47 million, of which 50%-75% is Alzheimer's disease. The number of people with Alzheimer's disease is the world's first, and China is one of the fastest growing countries in the world. However, the diagnosis and treatment rate of Alzheimer's disease is strongly contrasting with the incidence of Alzheimer's disease. Only 22% of all Alzheimer's and other types of dementia worldwide have been diagnosed. In China, the proportion is lower, and 49% of the cases are misrecognized. For natural aging, only 21% of the patients received standardized diagnosis, only 19.6% received medication. And at the current level of medical treatment, Alzheimer's disease is an untreatable disease. Therefore, early diagnosis of Alzheimer's disease, early prevention and intervention is very important. Mild cognitive impairment (Mild Cognitive Impair). Ment, MCI), a precursor of Alzheimer's disease proposed by predecessors, is an intermediate state between normal aging and dementia. Mild cognitive impairment can be used as a "predictor" of Alzheimer's disease. It can delay the progress of Alzheimer's disease if it can present this state of early onset and give appropriate pre treatment. Therefore, the correct diagnosis of Alzheimer's disease, especially the correct diagnosis of mild cognitive impairment at its early stage, the prevention of Alzheimer's disease, the early detection and treatment intervention are essential. In order to identify Alzheimer's disease and mild cognitive impairment, this paper proposes a classifier based on the support tensor machine (STM) and T1 weighted MRI Diagnostic methods of cerebral gray matter images gray feature. The classifier with three-dimensional (3D) brain images as the model input, using STM iterative algorithm to train the classifier for each weight vector and then classify the first mock exam. Collected 70 cases of AD patients, 112 MCI patients (included in the follow-up into AD, MCI-C:MCI, Converters with the conversion to AD, MCI- 56 cases of NC:MCI Non-converters and 70 cases of normal human (NC) T1-Weighted MRI three-dimensional (3D) brain images. First, the gray matter of each brain image is extracted to construct the three order gray matter tensor of each brain image, and the tensor is 95 * 119 x 102. to obtain the low dimensional principal component tensor of the three order gray matter tensor by tensor principal component analysis (TPCA). The principal component tensor is classified as the input of the classifier based on STM (STM-TPCA). The tensor independent component analysis (TICA) is also used to extract the independent component tensor of the three order gray matter tensor to be classified as the input of the classifier based on the STM (STM-TICA). Considering the redundancy between the features, the tensor is supported by the tensor iteration. After the algorithm transforms the tensor features into vector features, recursive feature elimination (RFE) is used to do feature selection, and the optimal subset is classified as the input of the classifier (STM-RFE). Finally, four groups of people are classified: ADNC, MCINC, ADMCI, MCI-CMCI-NC, and this classification model is trained by 90% off cross validation methods. In the classification of AD and NC, the correct rate is up to 91.19% (sensitivity 92.86%, specificity 89.52%); for the classification of MCI and NC, the correct rate is up to 83.15% (sensitivity 91.67%, specificity 69.52%); for the classification of AD and MCI, the correct rate is up to 82.23% (sensitivity 65.71%, specificity 92.56%); the classification of MCI-C and MCI-NC is correct. The highest rate was 77.08% (sensitivity 77.38%, specificity 76.79%). In addition, the basic information (age, sex, education level) and cognitive score (Mini-Mental State Exam, MMSE score, Alzheimer's Disease Assessment Scale-cognitive subscale, ADAS-cog score) were also classified, and the results were found to be combined with basic information and recognition. The classification effect can be further improved after the knowledge of the score, and compared with the combination of Shen and Willette multimodal data as model input, the results of this method are better. The experimental results show that the STM based classifier based on the gray matter image of the T1 weighted MRI brain image is an effective Alzheimer's disease diagnosis. And the basic information, the cognitive score is compatible with the MRI gray matter image, and has a good complementarity. In the course of the experiment, we found that because of the high tensor dimension (95 * 119 x 102), the running speed of the tensor independent component analysis and the recursive feature elimination method is relatively slow, and the high dimensional features greatly improve the characteristics. Therefore, we consider the texture feature (Texture feature) Zhang Liang (12 * 13 * 4) based on the grayscale symbiotic matrix (Texture feature) based on 13 directions and 4 distances (STM-Texture) as the input of the classifier based on STM (STM-Texture). This improved method reduces the dimension of the input sample and thus improves the whole classification model. The experimental results show that the texture tensor of the grayscale symbiotic matrix is used as the classification method of the classifier based on STM, which can not only maintain the superiority of the original classification method, but also reduce the running time.
【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R749.16;R445.2

【参考文献】