Files in this item

FilesDescriptionFormat

application/pdf

application/pdfNAKASHIMA-THESIS-2015.pdf (2MB)
(no description provided)PDF

Description

Title:Classification of tinnitus versus non-tinnitus hearing impaired subject fMRI data
Author(s):Nakashima, Corinne Cameron
Advisor(s):Baryshnikov, Yuliy
Department / Program:Electrical & Computer Engineering
Discipline:Electrical & Computer Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:M.S.
Genre:Thesis
Subject(s):clustering
k-means
tinnitus
classifiers
functional magnetic resonance imaging (fMRI)
Abstract:Diagnosis of tinnitus, the perception of sound minus external stimulus, depends on self-report and there is currently no objective way to determine if a person has the disorder. Previous examinations of resting-state brain scan data combined with a priori knowledge in the form of suspect regions found from previous research to show differing activity for tinnitus versus non-tinnitus regions of interest have consistently found spatial and temporal differences in electrical activation patterns between tinnitus subjects and controls, implying that development of an objective diagnostic method is possible. However, a whole-head, data-driven, non-a priori method appropriate for diagnostic use has yet to be developed. In the past, clustering algorithms have been used on resting-state data for diagnosis of other conditions, by dividing data into as many groups as desired, and determining each datum's or feature's membership in a cluster group by how similar the feature is to all other features in the cluster, then iterating through subsequent reapplication of this process to the data, until cluster membership converges towards a static set of groups; in the case of this study, the data that has been clustered is resting state functional magnetic resonance imaging (fMRI) data, which consists of brain voxel time-courses, essentially three-dimensional pixel information corresponding to relative levels of electrical activity within the data point over time. Then a linear or non-linear classifier, an algorithm that trains itself by building a probability profile for how likely each feature will be to belong to a cluster group, given a target or control class (in the case of this study, tinnitus or non-tinnitus hearing impaired), is built based on feature cluster label data from each of a subset of subjects from both classes. This classifier is then tested on the remaining subject fMRI data to determine the efficacy of the clustering + classifier system being tested. After a variety of clustering methods and classifiers as well as differing methods for applying these strategies had been tested, the best results indicate that a k-means clustering algorithm paired with a Gaussian maximum-likelihood classifier yields an average classification success rate of close to 60 percent. While this result is not yet good enough to be applied with expectation of a reliable result, it is significantly better than chance, demonstrating that whole-head, (mostly) unsupervised classification is possible, and indicating a direction for further research into the development of a useful tinnitus diagnostic tool.
Issue Date:2015-12-10
Type:Thesis
URI:http://hdl.handle.net/2142/89093
Rights Information:Copyright 2015 Corinne Nakashima
Date Available in IDEALS:2016-03-02
Date Deposited:2015-12


This item appears in the following Collection(s)

Item Statistics