Research Background
Research Accomplishments
Research In-Progress
Research Summary
With all of our technological advances, we still have a relatively poor understanding of how our human brain works and how it is disfunctioning due to disease or injury. Developing technology to understand the functioning human brain is my passion and area of inquisition. To observe the cognitively active brain, functional magnetic resonance imaging (fMRI) is the ideal noninvasive imaging technique.
FMRI is a variant of the standard anatomical MRIs that acquires fast lower resolution MRIs over time while the human brain is functioning. FMRI uses the blood oxygen level dependent (BOLD) signal discovered by Ogawa in 1990 which is a correlate to neuronal firing. When brain neurons fire, the local blood oxygenation increases. When the local blood oxygenation increases, the volume elements (voxels) that contain the increased oxygenated blood have a corresponding fMRI signal increase.
When our brains are inserted into an MRI machine, the hydrogen protons align and precess about the main magnetic field at their resonant frequency. A microwave radiofrequency (RF) pulse is sent from the machine to perturb the precessing hydrogen nuclei projecting a voxels magnetization into the horizontal plane creating a signal. Different tissue types under varying disease or blood oxygenation level have different signal properties due to their specific magnetic properties. The fMRI voxel’s magnitude signal depends upon the transverse T2* and longitudinal T1 relaxation rates of the material being imaged but also spin density ρ and the static magnetic field B0. The fMRI magnitude signal increases with oxygenation due to task. Although fMRI is an amazing technology, there are many limitations such as spatial and temporal resolution in addition to a very high noise level.
Research Accomplishments
The entire fMRI process involves many fields from MRI physics to neuroscience, with computational brain sciences joining the fields together. The computational aspects range from image reconstruction, image and time series processing, voxel and region connectivity, detection of statistically significantly active voxels, to the thresholding of statistical parametric maps. I have published research in and around these computational aspects. I have developed improved temporal resolution and improved reconstruction of task images, examined the effects of processing images, developed new complex-valued (magnitude and/or phase) time series models for brain activation, and evaluated various statistical parametric map thresholding. Note: I would like to particularly highlight that by far nearly all fMRI studies and research projects only use the magnitude of the complex-valued time series and discard the phase half of the data. It has been overwhelmingly demonstrated that complex-valued time series have increased sensitivity and specificity and that the phase depends on the physiology in a voxel. Each voxel’s base phase depends on the base magnetic field B0 and certain voxels phase changes with task similar to the magnitude, dependent on local magnetic field changes. This is a new frontier for neuroscience research and opens up the investivation of vasculature and direct neuronal firing.
Research in Progress
In recent years I have migrated upstream and crossed over to image reconstruction. In MRI and fMRI, the original measurements are not voxel values, but are to a good approximation the Fourier coefficients for spatial frequency cosine and sine plane waves termed k-space. I have published on the signal and noise properties of processed spatial frequencies and on the effects of systematically measuring less data. In fMRI, spatial and temporal resolution is limited by the relatively large amount of time that it takes to measure the k-space data. I am currently pursuing three lines of research to measure less data but still be able to reconstruct a high-quality image for brain activation.
The first subsampling path I am investigating is a Bayesian version of the SENSE image reconstruction technique termed BSENSE. This is the subject of one of my PhD students’ dissertation in which lines of k-space are skipped for multiple sensitivity weighted coils, resulting in aliased complex-valued images. The observed overlapping complex-valued voxels are modeled as a linear sum of sensitivity weighted original non-overlapping unobserved images. The multi-coil aliased voxels are separated and combined into a single complex-valued image. Prior distributions are placed on the unknown true non-overlapping voxel values, the unobserved multiple coil sensitivity weightings, and on the residual variance. Hyperparameters are assessed for the prior distributions using full pre-scan calibration images and marginal posterior mean true non-overlapping voxels (images) are estimated. Increased fMRI activation is computed from these images. Also along this subsampling path, the same student and I have been developing a Bayesian version of a technique caled GRAPPA. Our BGRAPPA technique uses prior information to estimate the missing k-space lines of data, so they can be inverse Fourier transformed and combines similar to non Bayesian GRAPPA.
The second subsampling path I am researching is a variant of a simultaneous multi-slice technique to measure rotating Hadamard pattern complex-valued sums and differences of several slice images. Utilizing non-summed pre-scan calibration images, the remaining patterns are formed and the several aliased complex-valued slice images are separated. This is the work of a current doctoral student’s dissertation and an extension of a previous doctoral students’ work. In this new research is to utilize the CAPRIHIANA technique to artificially shift the objects in the images vertically by adding a small amount of magnetization and to also CAIPI-VAT shift them horizontally using another additional amount of magnetization. Using the one observed Hadamard pattern summed image and artificially summing the pre-scan calibration images to form the other patterns, the images are separated to be as they should and brain activation is computed.
The third subsampling path I am exploring is non-Cartesian radial sampling of k-space where measurements are nonuniformly taken along radial spokes. It is temporally more efficient to measure data along spokes, but the reconstruction is much more difficult. Here a nonuniform inverse Fourier transform is utilized, and the approach will be to assess hyperparameters of the likelihood of the reconstructed images from pre-scan calibration images and estimate a much higher resolution image as the marginal posterior mean. This work will form the basis of a current student’s dissertation. Upon successful completion, this work opens up the possibility of more complicated k-space trajectories such as rotating propeller bands, spirals, and rosettes. Along the way to develop these non-Cartesian methods, we have been developing a more realistic simulated data software program that we plan to release to the public when substantially complete.
Research Summary
My research efforts have involved the theoretical development of new mathematical methods, their statistical characterization by computer simulation, and validation by phantom and human experiments. Developing fMRI technology will enable better understanding of brain function and how it is affected by neurological diseases, mental illnesses, and brain injuries. In the near future I plan to continue researching Bayesian techniques to measure less k-space data in space but also in time, k-t subsampling. In addition, I would like to have a student explore the biological advantages of utilizing phase time series information. Each of my methodological advances will improve fMRI analysis by increased sensitivity, specificity, and localization of fMRI activation.
Return to Professor Rowe's Webpage
|
