The impact of initialization in optimization of independent components in functional magnetic resonance imaging Pubblico
Yang, Zixi (Spring 2019)
Abstract
MELODIC software is widely used to conduct independent component analysis (ICA) of functional magnetic resonance imaging (fMRI), in which the components correspond to resting-state networks and image artifacts. However, the objective function in ICA is nonconvex and typically contains local optima. Consequently, multiple initial values improve the ability to find the global optima. The current version of MELODIC software does not allow an evaluation of the impact of multiple initializations. The goal of this paper is to examine the impacts of initialization in MELODIC. To clarify the effects, we applied MELODIC with multiple seeds to two datasets: the tutorial data set available with MELODIC software and a subject from the Autism Brain Imaging Data Exchange. We examined the variability between components estimated from 100 seeds. In both datasets, there were a number of components that exhibited high variability between seeds, especially with components classified as unknowns. These components tended to have lower kurtosis, which may be related to estimating too many components. Some components that were sensitive to initialization contained spatial features indicative of signal or artifacts. We conclude that independent components estimated with MELODIC are sensitive to initialization.
Table of Contents
1 Introduction 1
2 Method 2
2.1 Data set description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.1 Tutorial Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.2 ABIDE Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 ICA Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Results 5
3.1 Multiple initializations for the tutorial data set . . . . . . . . . . . . . . . . . . . . . . . . 5
3.1.1 Local optima with 100 seeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.1.2 Correlation matrix for differences with 100 seeds . . . . . . . . . . . . . . . . . . . 6
3.1.3 Differences among 159 components . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.4 Orthographic views of signals and artifacts . . . . . . . . . . . . . . . . . . . . . . 7
3.1.5 Time courses and spectral density of signals and artifacts . . . . . . . . . . . . . . 8
3.1.6 Correlation matrix for signals, artifacts and unknowns . . . . . . . . . . . . . . . . 10
3.2 Multiple initializations for the ABIDE data set . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2.1 Local optima with 100 seeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2.2 Correlation Matrix for differences with 100 seeds . . . . . . . . . . . . . . . . . . . 12
3.2.3 Differences among 80 components . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.4 Orthographic views of signals and artifacts . . . . . . . . . . . . . . . . . . . . . . 14
3.2.5 Time courses and spectral density of signals and artifacts . . . . . . . . . . . . . . 15
3.2.6 Correlation matrix for signals, artifacts and unknowns . . . . . . . . . . . . . . . . 16
4 Discussion 17
About this Master's Thesis
| School | |
|---|---|
| Department | |
| Subfield / Discipline | |
| Degree | |
| Submission | |
| Language |
|
| Research Field | |
| Parola chiave | |
| Committee Chair / Thesis Advisor |
Primary PDF
| Thumbnail | Title | Date Uploaded | Actions |
|---|---|---|---|
|
|
The impact of initialization in optimization of independent components in functional magnetic resonance imaging () | 2019-04-09 16:05:33 -0400 |
|
Supplemental Files
| Thumbnail | Title | Date Uploaded | Actions |
|---|