Skip to main content
Sign In

MR Physics


Methodological Development of fMRI Techniques

Dr. Du has developed a volume-selective z-shim technique to reduce the susceptibility artifacts in fMRI studies. In this technique, z-shim compensation is applied only to a few selected slices in a whole brain fMRI study. As a result, the time penalty for multiple acquisitions at the z-shim slices is greatly reduced. Using this technique, BOLD activations can be effectively detected at the orbito-frontal cortex in whole brain fMRI with a TR of 2 seconds (see Fig. 1). This technique has currently being used in the fMRI studies of drug addiction, alcoholism, and behavioral disorders.

Quantitative Mapping of Myelin Water Fraction (MWF) in the brain

Healthy neuron axons in the brain are wrapped by myelin sheath, which is formed by fatty layers of myelin. The myelin sheath is critical for the communication and coordination among different regions of the brain. The destruction of the myelin sheath, a.k.a. dymyelination, can severely disrupt the communication of brain signals and impairs normal brain functions. Multiple sclerosis (MS) is a chronic neuro-degenerative disease in which the body's immune system attacks the myelin sheath. Demyelination is considered as a primary cause of dysfunction in MS patients. Recent literatures show that demyelination is also linked to Alzheimer’s disease, schizophrenia disease, and autism. Dr. Du has developed an innovative multi-gradient-echo MRI technique to acquire 4-dimensional (x,y,z,t) data set on an MR scanner. Multi-compartment analysis with a 3-pool model is used for 3-dimensional (3D) whole-brain mapping of MWF in the white matter of the brain by optimal fitting of the time course of the signal decay in each voxel. The MWF maps of a fixed MS brain show great details of the myelin distribution in normal appearing white matter and demyelination in MS lesions (see Fig. 2). Our current research focus is on in vivo MWF mapping of the brain.

Artifact reduction in Echo-Planar Spectroscopic Images (EPSI)

EPSI sequence produces high temporal resolution MRI images, from which the characteristics of a specific tissue can be obtained by its spectrum and Free-Induction-Decay (FID) analysis. Since it takes much longer time to acquire a single complete k-space data than other sequences, it is more prone to artifacts caused by movements, breathing, physiological fluctuations, etc. We focus on reducing these artifacts by developing different k-space sampling schemes and postprocessing methods.

Improving the accuracy of functional maps

It is well known that fMRI data has a strong spatial dependence, which can be exploited by simultaneously considering a collection of neighboring voxels instead of treating each voxel individually. It is then possible to assess complicated spatio-temporal evoked responses with a potential increase in sensitivity. Consideration of these issues leads to the use of multivariate statistical analysis. Dietmar Cordes and Rajesh Nandy (now at UCLA) have developed canonical correlation analysis (CCA), a multivariate extension of ordinary single voxel correlation analysis. Using real and simulated data, this method has been demonstrated to be clearly superior to the conventional univariate methods in detecting activations from fMRI data with low CNR However, conventional CCA is prone to an artifact known as bleeding artifact leading to a loss of specificity. As a further improvement, Dr. Cordes and Dr. Nandy have introduced a new method known as adaptive CCA to correct the bleeding artifacts. With the use of novel Receiver Operating Characteristic (ROC) curves, they have demonstrated that the new method results in an increase in sensitivity and specificity of the activation maps compared to both univariate analysis and conventional CCA. This method will be a valuable tool for investigators using fMRI.

Novel ROC Methods for Assessing Processing Techniques

A popular and useful tool for such an assessment is a method based on ROC. However, due to the lack of knowledge about the actual locations of the truly active voxels, in the past ROC methods in fMRI have been restricted to simulated data, where the timecourses of the voxels are treated individually assuming spatial independence among neighboring voxels. However, in real fMRI data, the assumption of spatial independence is not valid and it is difficult to simulate fMRI data accurately reflecting the true spatial dependence. This essentially restricts the use of ROC methods using simulated data to single voxel analyses. Furthermore, even in a single-voxel analysis, the actual response for some complex tasks (as in cognitive tasks involving explicit and implicit memory functions) may be difficult to model. To solve these problems, Dr. Cordes and Dr. Nandy have developed several groundbreaking modifications to ROC methods that allow for the implementation of ROC methods using real fMRI data. The use of real data makes these methods extremely versatile with a wide range of applications including multivariate statistical analysis. The newly introduced ROC methods will be extremely useful to assess the performance of any hypothesis based fMRI post processing algorithm.

Exploratory Analysis Using ICA

The objective in the analysis of fMRI data is to identify the anatomical locations of neuronal activation during the performance of a task based on MRI image signal fluctuations. Fluctuations directly related to neuronal activity, however, are obscured by fluctuations from other sources, such as CSF and blood pulsation and subject motion. Independent Component Analysis (ICA) is a popular signal processing method that has been recently applied to the analysis of fMRI data. ICA is capable of isolating the MRI signal fluctuations caused by neuronal activity alone, allowing for a more accurate mapping of neuronal activity to anatomical location. ICA, however, is often computationally intractable due to the large size of fMRI data sets. Unfortunately, the size of fMRI datasets often renders this technique computationally intractable without any data reduction in a preprocessing step. Data reduction is directly connected to the problem of estimating the intrinsic dimensionality. Dr. Cordes’ group is currently investigating this problem using information-theoretic criteria and a method based on testing the equality of the smallest eigenvalues of the covariance matrix. Conventionally, data reduction is performed by Principal Component Analysis (PCA). It can be demonstrated that the accuracy of the ICA sources extracted depends on the relative amplitude of the activation. Data reduction to 40 to 60 components by PCA preprocessing can be beneficial and provide better ICA source estimates for activations with relative amplitude larger than 1%. But PCA preprocessing can also be detrimental if the activation is weak (relative amplitude < 1%). Furthermore, it can be shown that information-theoretic such as the AIC and MDL criterion (derived from uncorrelated Gaussian distributions) significantly overestimate the number of sources in fMRI data. Current research is directed towards the development of better dimensionality reduction techniques.