Matrix Factorization Algorithms for the Identification of Resting-State Networks Using Functional Magnetic Resonance Imaging

• Activity Number: 202
• Type: Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 11:15 AM
• Sponsor: Section on Statistics in Imaging
• Author(s): Karthik Devarajan* and Harvey Hensley
• Companies: Fox Chase Cancer Center and Fox Chase Cancer Center
• Keywords: functional magnetic resonance imaging ; resting-state network ; non-negative matrix factorization ; independent component analysis ; brain imaging ; functional connectivity

Abstract

In resting-state functional magnetic resonance imaging (rs-fMRI) experiments, a series of blood-oxygen level dependent (BOLD) MRI scans are acquired over a period of ~10 minutes in the absence of any specific task. Analysis of these scans has revealed spontaneous low-frequency fluctuations in the BOLD signal that has been used to investigate the functional architecture of the brain. Independent component analysis (ICA) is often employed as an analytical method for this problem. When data occur naturally on the nonnegative scale, as in rs-fMRI studies, it appears more intuitive to apply a method that retains the non-negativity requirement on the resulting components. Non-negative matrix factorization (NMF) approximates a high dimensional non-negative matrix as the product of two non-negative matrices based on an underlying statistical model. NMF allows only additive linear combinations of the data which enables it to learn parts that have distinct physical representations in reality. In this paper, we develop NMF algorithms for handling signal-dependent and correlated noise in rs-fMRI data and compare their performance with ICA in the extraction of underlying components.