Performance Evaluation of Multi-Level Various-Order Functional Connectivity Based Classification Framework Using Connectome-Convolutional Neural Network
- 주제(키워드) Convolutional Neural Network , Functional Connectivity
- 발행기관 고려대학교 대학원
- 지도교수 정지채
- 발행년도 2019
- 학위수여년월 2019. 8
- 유형 Text
- 학위구분 석사
- 학과 대학원 컴퓨터·전파통신공학과
- 원문페이지 63 p
- 실제URI http://www.dcollection.net/handler/korea/000000084645
- UCI I804:11009-000000084645
- DOI 10.23186/korea.000000084645.11009.0000925
- 본문언어 영어
- 제출원본 000045999229
초록/요약
Resting-state functional magnetic resonance imaging (rs-fMRI) has gained in popularity for diagnosing brain diseases, including mild cognitive impairment (MCI). Despite numerous studies on MCI with rs-fMRI, identifying neuroimaging biomarkers for the disorder is yet incomplete. Many researchers have studied on functional connectivity (FC) based classification to represent the abnormalities in brain networks that could be deformed in MCI patients. Although some existing methods are partially successful, they fail to capture more latent and complex information by only considering ‘low-order’ relationships between ‘two’ brain regions. Also, as traditional machine learning algorithms, support vector machine (SVM) is still widely used to classify FCs, the application of other machine learning techniques such as convolutional neural network (CNN) has been applied very recently and has remained largely challenging. In this thesis, we propose and analyzes a novel multi-level various-order functional connectivity (FC) based classification framework using connectome-convolutional neural network (CCNN). The proposal constructs a unified framework of both high-quality various-order FC networks, which can find discriminative brain connectivity patterns between MCI and normal controls (NC), and a state-of-art connectome-convolutional neural network to learn features and classifiers for each MCI and NCs. We assess the diagnostic performances of the proposed framework on a publicly available data-set in terms of accuracy, sensitivity and specificity. We then demonstrate the effectiveness of our framework by comparing it with competing methods. We also show that our unified model is able to combine information ranging from superficial interactions to more latent and complex relationships with multi-level various-order FC networks and the corresponding framework using multi-level various-order FC outperforms other classifiers using either ‘single’-level various-order FC metrics or multi-level ‘same’-order FC networks. From this, our proposed framework can provide instructive biomarkers not only for diagnosis of MCI, but for other mental/neural disorders, such as autism spectrum disorder (ASD)
more목차
Chapter 1. Introduction 1
Chapter 2. Materials and Methods 5
2.1 Data Acquisition and Preprocessing 5
2.2 Functional Connectivity Network Estimation 7
2.2.1 Conventional Low-Order FC Network Construction 7
2.2.2 High-Order FC Network Estimation based on Correlation’s Correlation 8
2.2.3 High-Order FC Network Estimation based on MVND 11
2.2.4 Proposed Multi-Level Various-Order FC Network Construction 14
2.3 Classifiers for Funcional Connectivity Based MCI diagnosis 16
2.3.1 Connectome-Convolutional Neural Network Classifier 17
2.3.2 Simple and Deep Neural Network Classifiers 20
Chapter 3. Experimental Results 22
3.1 Performance Evaluation 22
3.1.1 FC Networks based on Correlation's correlation 23
3.1.2 FC Networks based on MVND 28
Chapter 4. Discussions 39
4.1 Analysis of Performances among Various FC Networks and Classifiers 40
4.2 Effect of Window Length and Step Distance to Performance 42
Chapter 5. Conclusions 44
References 46
Acknowledgement 51
