CASIA OpenIR  > 毕业生  > 博士学位论文
基于多中心脑影像的阿尔茨海默病异常模式及分类研究
金丹
2020-05-28
页数158
学位类型博士
中文摘要

阿尔茨海默病(Alzheimer’s diseaseAD)是一种区别于正常衰老过程,以记忆损害、认知功能下降为主的,不可逆转的神经退行性疾病,是老年痴呆最常见的病因。我们对AD的病因及发病机制的了解仍十分有限,目前还没有有效的手段可以延缓或停止AD的疾病进展。脑成像技术,尤其是多模态磁共振成像技术的快速发展,使研究人员能够无创地、便捷地研究大脑的功能和结构,为进一步了解AD的发病机理和AD的早期诊断提供了新手段。近年来,模式识别与机器学习方法在神经影像领域的成功应用,为利用脑影像技术实现脑疾病的计算机辅助诊断提供了可能。因此,通过磁共振成像技术和机器学习方法的结合,研究AD稳定、可靠的早期诊断标记物十分重要。本论文利用多站点、大样本的结构和功能磁共振影像数据,分别从脑结构和功能的角度,系统地研究了 AD患者的大脑异常,并结合机器学习方法进行了AD的计算机辅助诊断,重点评估了结构和功能影像特征作为AD早期诊断的生物标记物的可靠性和泛化性。论文的主要工作和创新点归纳如下:

1. 基于深度学习的AD结构磁共振影像标记研究

为了推进结构磁共振影像在AD临床诊断中的应用,研究人员利用机器学习方法研究了结构影像作为AD诊断标记物的可能性。虽然深度学习方法通过大样本的自动优化学习可以取得很好的分类效果,但是常用的深度学习方法无法获得哪些影像特征在分类中发挥了重要作用以及这些影像特征和AD临床特征的关系。针对目前研究的局限性,本研究创新地结合注意力机制模块,提出了一种新的三维注意力网络,在实现AD的计算机辅助诊断的同时给出大脑中最具有区分能力的脑区位置,提高了模型的可解释性并提升了深度学习模型在AD临床辅助诊断中的应用价值。研究利用两个大样本、多站点的独立数据集,全面评估了模型的分类性能、泛化性以及与临床指标的关系。实验表明,本模型在跨数据集和数据集内部的交叉验证中都取得了较高的分类准确率,具有较好的泛化性能。另外,通过注意机制模块,研究发现了对分类重要的大脑区域,主要位于颞叶、海马、海马旁回、扣带回、丘脑、楔前叶、岛叶、杏仁核和梭状回,并且这些脑区的重要性和萎缩程度显著相关。最后,研究发现模型输出的决策值与被试的认知能力、病理学改变以及转换为AD的时间有显著的相关关系,进一步证明了AD中脑结构异常作为AD辅助诊断的影像标记物的潜在效用。

2. 脑功能网络指标的可重复性研究

静息态功能磁共振影像和图理论方法的结合有助于我们挖掘潜在的网络指标作为生物标记物。然而,利用脑网络指标发展疾病诊断的生物标记物的重要前提是网络指标在相同被试者中重复测量的可靠性。以往的研究大多是从静态脑网络的角度在小样本数据集上进行指标的可靠性分析。然而,静息状态下的自发脑活动是一个动态过程。所以,针对目前研究中存在的局限性,本文利用大样本、重复扫描的被试数据,从动态脑网络的角度,在不同时间尺度上,系统地评估了脑网络指标的可靠性,并进一步研究了扫描持续时间和网络密度对指标可靠性的影响。研究发现:1)功能连接强度、物理连接距离、度和全局效率比其他指标具有更好的可靠性。2)相比于皮下核团,大脑皮层的功能连接和脑区的功能网络具有更高的可靠性。3)扫描时间间隔对网络指标的可靠性有影响,然而网络密度对大多数网络指标的影响很小。4)网络指标的可靠性随扫描持续时间的增加而增加,扫描时间大于12分钟可以作为临床应用的首选。本研究为之后基于功能磁共振影像寻找疾病诊断的生物标记物的研究工作提供了基础。

3. AD脑功能网络异常及其辅助识别的可重复性和泛化性研究

基于静息态磁共振成像的研究表明AD的脑功能存在普遍的改变。由于缺乏大样本、多站点的数据集,与AD脑功能有关的异常模式还没有得到一个可重复的、清晰一致的结论。因此,本研究利用多中心、大样本的影像数据,使用荟萃分析的方法找到了AD中稳定的、可重复的功能异常模式,并进一步分析了这种功能异常与AD认知和生物学指标的关系。在此基础上,本文利用多变量分析方法和独立数据集,系统地评估了AD的功能异常模式作为早期诊断的影像指标的可靠性和泛化性。研究发现了AD中一致的脑功能异常主要位于默认网络、扣带回、基底神经节以及海马区域,并且功能异常的严重程度与患者的认知损伤和β淀粉样蛋白累积程度显著相关。基于异常的功能指标,本研究在多站点数据集上对个体诊断状态和临床评分的预测取得了很好的效果。在独立数据集上的可重复性分析进一步证明了研究结果的可重复性和泛化性。

4. 基于功能影像特征的AD亚型研究

之前的研究表明轻度认知障碍和AD的临床诊断显示出显著的表型异质性。这种变异性在一定程度上反映了潜在的遗传、环境和神经病理学差异。所以,描述这种异质性对于精确诊断、个性化预测非常重要。以往的研究大都是从神经解剖学和病理学的角度研究AD的异质性。本研究首次使用功能磁共振影像数据,在上一个工作的基础上,利用非负矩阵分解的方法在大样本、多中心的数据集上确定了AD的四种亚型,并进一步研究了AD患者亚型间的大脑结构差异和认知功能差异。在独立数据集上的可重复性分析进一步证明了研究结果的可重复性。

英文摘要

Alzheimer's disease (AD) is a kind of irreversible neurodegenerative disease, which is different from the normal aging process. It is the most common cause of dementia. Our understanding of the etiology and pathogenesis of AD is still very limited, and there are no effective means to delay or stop the progression of AD. With the rapid development of brain imaging technology, especially multimodal magnetic resonance imaging (MRI) technology, researchers can study the function and structure of brain noninvasively and conveniently, which provides new means for further understanding the pathogenesis and achieving early diagnosis of AD. In recent years, the successful application of pattern recognition and machine learning in many fields of medical image analysis provides the possibility of computer-aided diagnosis of brain diseases by using brain imaging technology. Therefore, it is very important to study stable and reliable early diagnostic markers of AD by combining MRI and machine learning. In this paper, we use multisite, large sample structural and functional MRI data, from the perspective of brain structure and function, systematically study the brain abnormalities of AD patients, and combine machine learning method to carry out the computer-aided diagnosis of AD, focusing on the evaluation of the reliability and generalization of structural and functional image features as biomarkers for early diagnosis of AD. The main contributions are summarized as follows:

1. A study of structural magnetic resonance imaging markers in Alzheimer's disease based on deep learning

In order to promote the application of structural MRI (sMRI) in the clinical diagnosis of AD, researchers used machine learning methods to study the possibility of sMRI as a diagnostic marker of AD. Although deep learning method can achieve good classification performance through automatic optimization learning based on large samples, the common deep learning method can not obtain which imaging features play an important role in classification and the relationship between these image features and clinical features of AD. In view of the limitations of previous studies, this research creatively combines the attention mechanism module, and proposes a new three-dimensional attention network, which can realize the computer-aided diagnosis of AD, and at the same time give the location of the most distinguishable brain regions, improving the interpretability of the classification model and enhance the application value of the deep learning model in the clinical aided diagnosis of AD. In this study, two large samples, multisite independent datasets were used to comprehensively evaluate the classification performance, generalization and relationship with clinical measures of the model. Experiments show that the model has good generalization performance, which achieves good classification accuracies in the inter-dataset and intra-dataset cross-validation. In addition, through the module of attention mechanism, we found the important brain regions for classification, mainly located in the temporal lobe, hippocampus, parahippocampal gyrus, cingulate gyrus, thalamus, precuneus, insular lobe, amygdala and fusiform gyrus, and the importance of these brain regions is significantly related to the degree of atrophy. Finally, we found that the class score of the model output also had significant correlations with the cognitive ability, pathological changes and the time of conversion to AD, which further proves the possibility of structural abnormality in AD as an imaging marker of assisted diagnosis.

2. A study of test-retest reliability of brain functional network metrics

The combination of resting-state functional MRI (rs-fMRI) and graph theory helps us to identify potential network measures as biomarkers. The important premise of using brain network measures to develop biomarkers for disease diagnosis is the reliability of the repeated measurement of network measures in the same subject. Most of the previous studies analyzed the reliability of measures in small datasets from the perspective of the static brain network. However, the spontaneous brain activity in the resting state is a dynamic process. Therefore, in view of the limitations of previous studies, this study systematically evaluates the reliability of brain network measures from the perspective of the dynamic brain network, using large sample rs-fMRI data, and further studies the impact of scan duration and network density on the reliability of measures. The results show that: 1) functional connection strength, physical connection distance, degree and global efficiency have better reliability than other measures. 2) Compared with the subcortical nuclei, the functional connections and the functional networks of regions in the cerebral cortex are more reliable. 3) The scanning interval has an effect on the reliability of network measures, and the network density has little effect on most network measures. 4) The reliability of network measure increases with the increase of scan duration. Scan duration greater than 12 minutes may be the first choice for clinical application. This study provides a foundation for the next research work of finding biomarkers for disease diagnosis based on fMRI images.

3. A study of the brain functional network abnormality and the generalizability and reproducibility of diagnostic classification in Alzheimer’s disease

fMRI studies in AD indicate that the disease is associated with widespread disruptions in brain functional networks. Due to the lack of large sample and multisite datasets, the abnormal patterns related to the brain function of AD have not yet reached a repeatable, clear and consistent conclusion. Therefore, this study uses multisite, large sample rs-fMRI dataset to find a stable and repeatable pattern of functional abnormality in AD based on a meta-analysis method, and further analyzes the relationship between functional abnormality and cognitive and biological indexes of AD. On this basis, we further use multivariate analysis methods and independent dataset to systematically evaluate the reproducibility and generalization of functional abnormality pattern as an imaging biomarker for early diagnosis. The study found that the brain dysfunction pattern of AD was consistent, mainly located in the default mode network, cingulate gyrus, basal ganglia and hippocampus, and the severity of these abnormalities was significantly related to cognitive impairment and Aβ burden of AD patients. Based on the abnormal functional measures, the predictions of individual diagnosis status and clinical score on multisite datasets achieve good performances. The analysis on independent dataset further proves the reproducibility and generalization of the results of this study.

4. A study of subtypes of Alzheimer’s disease based on functional imaging features

Previous studies have shown significant phenotypic heterogeneity in the clinical diagnosis of mild cognitive impairment and AD. This variability may reflect potential genetic, environmental, and neuropathological differences to some extent. Therefore, describing this heterogeneity is very important for accurate diagnosis and personalized prediction. Previous studies have investigated the heterogeneity of AD based on anatomical and pathological features using sMRI and Positron Emission Computed Tomography (PET) images. Based on our previous work, four subtypes of AD are determined by using the method of nonnegative matrix decomposition on a large sample and multisite dataset, and the differences of brain structure and cognitive function between the subtypes of AD are further studied. The analysis on independent dataset further proves the reproducibility of the results of this study.

关键词阿尔茨海默, 磁共振成像,计算机辅助诊断, 脑结构和功能,荟萃分析
学科领域人工智能其他学科
学科门类工学 ; 工学::控制科学与工程
语种中文
七大方向——子方向分类脑网络分析
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/39025
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
金丹. 基于多中心脑影像的阿尔茨海默病异常模式及分类研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2020.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA Thesis-金丹.pdf(8361KB)学位论文 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[金丹]的文章
百度学术
百度学术中相似的文章
[金丹]的文章
必应学术
必应学术中相似的文章
[金丹]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。