CASIA OpenIR  > 脑网络组研究
面向精神疾病分类诊断的磁共振影像机器学习算法研究
姚东任
2021-05-21
Pages102
Subtype博士
Abstract

核磁共振成像技术(Magnetic Resonance Imaging,MRI)作为一种记录脑结构或功能的无创技术,为研究精神疾病的神经基础提供了有力工具。受制于精神疾病的复杂性和难以明确的发病机制,目前临床诊断主要依据患者主诉症状和定性诊断,总体疗效欠佳,急需客观影像学标记辅助诊疗。本博士课题主要研究面向精神疾病分类诊断和磁共振影像的机器学习算法,在传统算法的特征提取、特征融合,以及流行的深度学习算法的可解释性和多尺度图卷积网络方面提出了新的改进或突破,并将其应用于多种脑疾病与健康对照的区别诊断和交叉验证,提高了基于 MRI数据的分类性能。本文的主要创新工作包括:


(1)提出了一种基于类别相对重要性的包裹特征选择算法。该算法根据特征的相对重要性(Relative Importance)结合设计的包裹特征选择策略,实现对原始高维度空间的精简。与此同时,将四分类任务(MCI的两种亚型eMCI、LMCI,AD患者,正常对照)拆分成多个二分类任务,利用层次化的类别判断得到最终每个被试的类别。该算法应用于ADNI数据上四类被试,每类各100例,选取基于FreeSurfer计算的皮层信息。实验结果表明,基于相关重要性的特征选择算法在极小化特征空间的基础上与常用降维方法相比对,对分类性能起到更好的促进作用。此外,海马、楔前叶、颞叶等脑区以及性别作为特征选择后保留的有效特征,在之前的许多研究中已经证实了可以在一定程度上刻画AD、MCI和正常对照的区别。


(2)提出了一种基于类别相对重要性的集成学习算法(Feature Selection Method based on Relative Importance and weighted Ensemble Learning, FS\_RIEL)。该模型根据不同类型的算法对特征贡献度衡量方式的不同,获得了多角度度量的特征重要性,进而实现特征的前融合;与此同时,为了有效的利用特征选择过程中使用了多个包裹分类器,算法将多种分类器进行了有效地集成学习,完成最终特征选择框架的后融合。该算法应用于成人ADHD、儿童ADHD两类静息态fMRI数据集上均取得了较好的分类结果。此外,在儿童ADHD-成人ADHD的交互预测中,研究发现从儿童中学到的特征空间拥有的泛化性能是优于传统特征选择算法直接在成人数据集中得到的分类结果。


(3)提出了基于多尺度互学习的三元组图卷积神经网络(Mutual Multi-scale Triplet Graph Convolutional Network)。该模型首先根据多种不同尺度的模板为每个被试构造对应的脑功能或者结构连接网络;然后利用图卷积神经网路对被试每个尺度下的脑网络连接进行表示学习,整合从粗到细的空间网络信息,该过程在三元组的约束下根据不同被试的类别进行度量和评估;最后通过互学习机制将不同尺度的结果进行融合后得到最终的判别结果。该算法在三种公开数据集上,与当前最好的算法和本模型变体比均取得更优的分类结果。更重要的是,通过引入互学习机制,不同尺度的网络连接信息能够得到更有效的融合,进一步提高分类性能。


(4)提出了一种新的时域自适应的动态图卷积神经网络(Temporal Adaptive Graph Convolutional Network),能够对静息态fMRI数据的空域和时域信息同时进行挖掘。该模型首先通过滑动窗口将时间序列进行划分;然后通过自适应图卷积模型为每个滑动窗口学习自己的连接网络;接着利用时域卷积挖掘不同滑动窗口间连接网络的变化得到被试最终的类别。实验结果在公开的MDD数据集上取得了领先于当前最优模型的性能。该模型作为基于数据驱动的图拓扑结构学习算法,不仅有效地建模了静息态功能磁共振影像在时域的动态变化过程,而且还克服了常规组水平上需要共享图连接模式的弊端。

Other Abstract

Magnetic Resonance Imaging (MRI), as a noninvasive technique for depicting brain structure or function, provides a useful tool for the study of the neural basis of mental disorders. Due to the complexity of mental disorders, the pathogenesis of various diseases is not clear. At present, diagnosis is mainly based on the symptoms complained by patients and qualitative diagnosis, with poor overall efficacy, and objective imaging markers are urgently needed to assist diagnosis and treatment. This doctoral project mainly studies machine learning algorithms for mental disease classification diagnosis and magnetic resonance imaging. Several new improvements or breakthroughs have been proposed in the aspects of feature extraction and feature fusion of traditional algorithms, as well as the interpretation of popular deep learning algorithms and multi-scale map convolutional network. The proposed algorithm has been applied to the differential diagnosis and cross-validation of various brain diseases and healthy controls to improve the classification performance based on MRI data. The main innovative work of this paper includes:

In the first work, we propose an ensemble learning system for a 4-way classification of Alzheimer's Disease and Mild Cognitive Impairment. The model made hierarchical classification mainly based on the relative importance of features. Four categories of subjects, including two subtypes of MCI (EMCI, LMCI), AD patients, and age-matched healthy controls, were divided into multiple two-category tasks to obtain the final category for each subject. The algorithm was applied to four categories of subjects in ADNI data, 100 cases in each type, and cortical information was generated based on Freesurfer software. The experimental results show that the feature selection algorithm based on relative importance can better promote the classification performance than the standard dimension-reduction methods based on minimizing the feature space. Besides, the hippocampus, precuneus, temporal lobe, etc., as well as scale and gender, are useful features retained after feature selection, which have been confirmed in many previous studies to describe the differences between AD, MCI, and healthy controls to a certain extent.


In the second work, we propose a feature selection method based on relevance importance and weighted ensemble learning. By integrating different algorithms, the model can obtain features measured from multiple angles and realize the pre-fusion of features. Simultaneously, to effectively utilize various package classifiers used in the feature selection process, we effectively integrated multiple classifiers to achieve the feature selection framework's final post-fusion. The algorithm was applied to adult ADHD and childhood ADHD in two types of resting-state fMRI data sets and achieved good classification results. Besides, we reserved the feature space learned from the children's ADHD dataset to apply to the adult dataset. The experimental results show that the feature space we learned from the children has a better generalization performance than the classification results obtained directly from the traditional feature selection algorithm in the adult dataset.


In the third work, we propose a mutual multi-scale triplet graph convolutional network. Firstly, four different scale templates were used to construct a network of brain functional or structural connections for each subject. Then the convolutional network of three graphs was used to measure the connection network at each scale. Finally, the results of different scales were fused by mutual learning mechanism to obtain the final discriminant results. The proposed algorithm achieves better classification results than the current best algorithm on three kinds of public data sets, which provides a new idea and solution for selecting brain templates of different scales.


In the fourth work, we proposed a temporal adaptive dynamic graph convolutional neural network to simultaneously mine the spatial and temporal information of resting-state fMRI data. Firstly, the model divides the time series by a sliding window. Then the adaptive graph convolution model is used to learn its connection network for each sliding window. After that the connection network's change between different sliding windows is mined by temporal 1-D convolution to get the subjects' final category. As an utterly data-driven graph topology learning model, the model effectively models the dynamic changes of resting-state fMRI images in the time domain and breaks through the traditional shared graph connection mode constraints. The experimental results achieve better performance than the current optimal model on the open MDD data set.

Keyword脑影像 连接网路 特征选择 图卷积
Language中文
Sub direction classification医学影像处理与分析
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/44816
Collection脑网络组研究
Recommended Citation
GB/T 7714
姚东任. 面向精神疾病分类诊断的磁共振影像机器学习算法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2021.
Files in This Item:
File Name/Size DocType Version Access License
Dongren__PHD_Thesis.(10847KB)学位论文 开放获取CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[姚东任]'s Articles
Baidu academic
Similar articles in Baidu academic
[姚东任]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[姚东任]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.