Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


36598.pdf (5.08 MB)

ETD Abstract Container

Abstract Header

Group Convex Orthogonal Non-negative Matrix Tri-Factorization with Applications in FC Fingerprinting

Li, Kendrick T
http://rave.ohiolink.edu/etdc/view?acc_num=ucin1592136688034675

Abstract Details

2020, MS, University of Cincinnati, Engineering and Applied Science: Computer Engineering.
Functional magnetic resonance imaging (fMRI) data has been collected and studied in the neuroscience community for more than two decades. Methods have been developed to model fMRI data as a network, referred to as a functional connectivity (FC) network, consisting of a matrix of pairwise correlation of the blood-oxygen-level-dependent (BOLD) signal measured at different brain regions. FC fingerprinting is a recently introduced problem where the goal is to identify subjects based on FC. Given reference fMRI scans from a set of subjects and an unidentified target fMRI scan from a subject in this set, the goal is to identify which subject the target scan was collected from by computing the correlation between the target FC and the reference FCs. The reference FC with the highest similarity is identified as the target FC. FC fingerprinting studies have reported near 100% accuracies, suggesting that the FC captures subject-specific neuronal activity signature effectively. However, recently we and others have shown that fingerprinting accuracy decreases as the number of subjects increases. Our previous work provided multiple contributions in the context of the FC fingerprinting problem. We showed using silhouette analysis that the reason for the decrease in FC fingerprinting accuracy was due to cluttering of FCs in high-dimensional space. We introduced an intelligent feature selection framework which used a subset of the elements in the FC to improve accuracy. We observed that effective features exhibited a block structure where connectivity between certain groups of regions were more effective for fingerprinting than others. In this thesis our objective is to extract a group-level block-structure from FCs where block-level connectivity can then be used for fingerprinting. One promising direction is to use a non-negative matrix factorization (NMF), a data modeling tool used to extract the inherent structure in data. One variant of NMF, fast matrix tri-factorization (Fast-MtF), allows us to extract the block structure from FC. However, Fast-MtF is incapable of utilizing group-level information to discover a group-level block structure for our dataset. Group-MtF methods, such as linked matrix factorization (LMF) and RESCAL, utilize group-level information. Unfortunately, they lack orthogonality and non-negativity constraints for extracting a block-structure. We developed a new problem formulation, Group convex orthogonaL non-nEgative mAtrix tri-factorizatioN (GLEAN), to capture a group-level block-structure for a group of mixed-sign matrices. By enforcing orthogonality and non-negativity constraints, we formulated a group `convex' model which finds the group-level block-structure . We developed a gradient-descent based approach to solve the GLEAN formulation. To evaluate GLEAN, we conducted two experiments: 1) a comparative analysis between GLEAN and state-of-the-art MtF methods on their effectiveness in finding the ground-truth block-structure from a synthetic dataset and 2) a comparative analysis between GLEAN and group independent component analysis (group-ICA) for computing coarse FCs for effective FC fingerprinting. We observed that GLEAN outperformed all other approaches, even with noisy datasets, in finding the ground-truth block-structure. Furthermore, we observed that GLEAN was more effective at capturing subject-specific information compared to the group-ICA method, providing insights on group-level connectivity effective for FC fingerprinting.
Gowtham Atluri, Ph.D. (Committee Chair)
Raj Bhatnagar, Ph.D. (Committee Member)
Boyang Wang (Committee Member)
75 p.
Computer Science
fMRI; Functional connectivity; FC fingerprinting; Matrix tri-factorization; Group matrix tri-factorization

Recommended Citations

Citations

  • Li, K. T. (2020). Group Convex Orthogonal Non-negative Matrix Tri-Factorization with Applications in FC Fingerprinting [Master's thesis, University of Cincinnati]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=ucin1592136688034675

    APA Style (7th edition)

  • Li, Kendrick. Group Convex Orthogonal Non-negative Matrix Tri-Factorization with Applications in FC Fingerprinting. 2020. University of Cincinnati, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=ucin1592136688034675.

    MLA Style (8th edition)

  • Li, Kendrick. "Group Convex Orthogonal Non-negative Matrix Tri-Factorization with Applications in FC Fingerprinting." Master's thesis, University of Cincinnati, 2020. http://rave.ohiolink.edu/etdc/view?acc_num=ucin1592136688034675

    Chicago Manual of Style (17th edition)

ucin1592136688034675
181
© 2020, all rights reserved.
This open access ETD is published by University of Cincinnati and OhioLINK.