Time frequency analysis of functional MR images
Date of Issue2014
School of Computer Engineering
Functional Magnetic Resonance Imaging (fMRI) technique is used to evaluate and visualize human brain activity in a non-invasive manner. It measures blood-oxygenation level dependent (BOLD) changes induced by neural activity resulted in due to external stimuli. FMRI experiments result in time-series of BOLD signals from brain voxels, collected over stimulated and rest conditions. Artifacts related to subject’s head movement and physiological processes confound the signal and low signal-to-noise ratio (SNR) of fMRI makes its analysis difficult. Temporal autocorrelation of fMRI time-series and the contextual dependencies on brain activation also requires proper handling. The motivation of this thesis is to propose frequency and time-frequency domain techniques to properly handle temporal autocorrelation, to enhance the detection of activated brain voxels and resting state brain networks. This thesis makes three major contributions and introduces: (i) a mixed spectrum analysis (MSA) method to remove the temporal autocorrelation effect in voxel time-series, in a model-free manner, while detecting activated voxels. It can also detect resting-state networks and physiological signals in fMRI data; (ii) two frequency domain methods, using higher harmonics of stimulus frequency, to enhance activation detection in event-related data. One of these selectively includes higher harmonics in analysis and the other is additionally extended in scale-space domain, and (iii) a time-frequency domain method to detect default-mode network in resting state fMRI data. Chapter 1 gives an introduction to the thesis, motivation for this work, briefly describes the previous work and provides a summary of our major contributions. In Chapter 2, we aim to remove the temporal autocorrelation effect from voxel time-series, in a model-free manner, while calculating the statistical significance of activation. A fixed model for temporal autocorrelation is restrictive because temporal autocorrelation structure is known to vary across brain regions. We propose a novel MSA method based on Priestley’s P(ω) test that removes the temporal autocorrelation effect in voxel’s time-series by determining its spectrum from appropriately truncated autocovariance series. Spatial correlation of activations is incorporated using a conditional random field (CRF). Subsequently, we apply Priestley’s principle in detecting default-mode network in resting state fMRI data. This new perspective uses temporal autocorrelation values instead of the normally used low-frequency amplitude and seed-region time-series. Exploring further, we test the proposed method to detect other frequencies of interest in fMRI data, such as physiological process related frequency components. This method is easy to use as compared to the present methods that tend to be exploratory and cumbersome requiring good knowledge of anatomical brain regions. Additionally, we apply the random field principle to efficiently regularize the AR(1) parameters in model-based prewhitening methods. It can easily be extended beyond AR(1) model. In Chapter 3, we intend to improve activation detection in periodic event-related data. We propose two ways of using higher harmonics of stimulus signal to achieve this. The first presented method, based on χ2 random field, uses the fundamental frequency and its first higher harmonic. The second method, based on Hartley’s F-test, evaluates the improvement in statistical significance values before including higher harmonics along with the fundamental frequency. We also extend the scale-space method based on χ2 random field, to include higher harmonics, for providing important information on spatial extent of activation. In Chapter 4, we aim to improve default mode network detection in resting state fMRI data. We propose a novel method based on Cohen’s class of distribution for time-frequency analysis of the data’s low-frequency components. It is thus able to use the variation in the amplitude of various low-frequency components along the time and also the mutual information contained in these low-frequency components to its advantage. The proposed method is also less tedious than the other commonly used seed-based time-series analysis method, as it does not require pre-identification of brain seed regions. We also present in this chapter the results of using earlier mentioned Priestley’s principle in detecting resting state networks. The presented methods are extensively tested on both synthetic and real fMRI data and the results are also compared with the standard methods in each application.
DRNTU::Engineering::Computer science and engineering::Computer applications::Life and medical sciences