Recent Projects
MR methods
Visual localizers at 3T
While working in David Heeger's lab, I wanted to investigate the temporal
dynamics of high- and low-resolution gradient echo (GE) BOLD at 3T. GE
BOLD is the standard fMRI method, and it is well known that large draining
veins can dominate regions of the image. Large vein-dominated voxels in
analysis can degrade both spatial and temporal characteristics of the data.
The hope was that using small voxels would let us identify and remove
vein-dominated voxels from the data, and that we would show a measurable
decrease in the delay and dispersion of the hemodynamic response in
high-resolution data. We found little difference in the temporal dynamics
of the low- and high-resolution data, but we did find that using small voxels
greatly increased our ability to accurately localize the cortical response to
visual stimuli. We measured the accuracy of two different kinds of localizer
scans: differential and single-condition. The particular pattern of
results we measured indicates that directional blood flow in large draining
veins can shift the observed BOLD response (not surprising), and that this
results in a mismatch between the ROIs defined by differential and
single-condition localizers. What was surprising was that the ROIs
defined by differential localizers were not a subset of the ROIs defined by the
less spatially restrictive single-condition localizer, particularly with low
resolution imaging. The conclusion I draw from this data is that, while
differential localizers are better at localizing cortical responses, spatial
accuracy does not necessarily lead to functional accuracy, and many analyses
may be better off useing a less spatially accurate ROI. This work has
been submitted for consideration as an abstract at VSS '06, and is in the process of
being submitted for publication.
High resolution imaging
in medial temporal lobe (MTL)
Collaborating with Lila Davachi and Souheil Inati, I
studied the accuracy of signal localization in MTL. This is a region
famous for distortion and signal loss, both due to spurious magnetic field
gradients induced by the air/tissue interface of the auditory canal, as well as
other places in the inferior regions of the brain where materials with
different magnetic suscetibility abut. Using established methods for
quantifying distortion, we found that the short read-out time allowed by zoomed
EPI (using outer-volume suppression) provided images with moderately high
resolution, acceptable signal-to-noise ratio, and little distortion.
Signal drop-out in anterior parahippocampal gyrus was the most significant
problem, and one which varied significantly from subject to subject. We
are in the process of preparing this data for publication.
Point spread function for
BOLD fMRI (spatial accuracy)
Since scanner technologies allow routine acquisition of high resolution
functional images (in-plane resolution ~1 mm), the intrinisic spatial
resolution of the hemodynamic response is becoming a critical question. One way
of quantifying the spatial blurring in a system is to measure the modulation
transfer function (MTF): the magnitude of response as a function of spatial
frequency. To measure the MTF in V1 (primary visual cortex), we present visual
stimuli of increasing spatial frequency, then measure the amplitude of BOLD
response. The inverse Fourier transform of this MTF yields an estimate of the point
spread function (PSF) for BOLD fMRI.
Taking advantage of the ultra-high field scanners at the CMRR, I completed a
set of experiments measuring the MTF/PSF for both gradient echo and spin echo
BOLD fMRI at 7 Tesla, as well as gradient echo BOLD at 3 Tesla. Early analyses
(presented as a poster at the 12th Scientific Meeting of the International
Society for Magnetic Resonance in Medicine) told an interesting story:
- at 3 Tesla, the spatial blurring of the BOLD
signal can be described by a Gaussian blurring kernel with a 3.5 mm full
width at half-maximum (FWHM). This is not significantly different from the
results of Engel et al. (Cerebral
Cortex 7:181) at 1.5 T.
- at 7 Tesla, the gradient echo signal has two
components: a strong signal that is fit by a 3.5 mm FWHM, and a weaker
signal that is described by a 1.8 mm FWHM. This consistent with the
expectation that the ultra high field GE BOLD signal comes from two venous
compartments: medium-sized draining venuoles (~20 micron diameter,
draining ~2 mm of cortical territory), and the capillary bed, offering
higer spatial resolution. poster pdf
- at 7 Telsa, the upper limit for the spatial
blurring intrinsic to the spin echo signal appears to be 1.8 mm.
While these results are consistent with other reports
on the spatial accuracy of BOLD, I believe that the conclusions cited above are
based on a confound in the data, originating in A) experiment design and B) the
fact that the point spread function is not uniform across the cortical
surface. Therefore, these results were not submitted for publication.
Z-shimming
Air-tissue interfaces above the auditory canals and frontal sinuses create
susceptibility-induced artifacts in EPI sequences that lead to either data loss
or poor registration between functional and anatomical images. Real-time
correction for susceptibility-induced artifacts is possbile but time consuming.
However, parallel imaging technologies - particularly advantageous at high
field - buy the necessary time to use z-shimming to improve image quality and
recover data in frontal lobe and inferior temporal cortex. I am in the process
of finishing up a short demonstration of the strengths and limitations of
z-shimming at 7 Telsa.
Presented as a poster at the 11th Scientific
Meeting of the ISMRM, July 2003.
Vision science
Contrast response to broadband images
Perhaps the most basic property of early vision (V1) is the contrast response -
the rate at which neural firing rates increase with increasing image contrast -
and contrast response functions for individual sine wave gratings (narrow
spatial frequency content) are well known. However extrapolating from
isolated patches of sine wave gratings to complicated patterns with broadband
spatial frequency content is not simple. We combined fMRI and
computational modeling to measure and understand the contrast response to
natural images, with typical 1/f amplitude specta, and to natural images with
whitened (flat) spatial frequency spectra. The general result is not
surprising - the contrast response saturates at a rate not much different from
the contrast response to sine wave gratings. However, the subtle
differences measured between the contrast response functions for natural and
whitened images were intriguing. We successfully modeled this with a
class of models which could account for both the contrast sensitivity function
at low contrast and contrast constancy at high contrast. pdf
BOLD response to images formed from random placement of Gabor
patches
This project was motivated by the following question: is specification of
RMS contrast and spatial frequency content sufficient to predict the V1 BOLD
response? A simple thought experiment, illustrated by the images shown
below (which all have the same RMS contrast and spatial frequency content),
gives the answer - no! The problem is in the posing of the question: global
contrast and global BOLD response are the wrong questions to be asking in
V1. At VSS 2004 I presented preliminary work testing our ability to
predict and quantify local BOLD
responses to local image features
(abstract published in JOV);
this is a line of research I plan to continue.
Vision beyond V1
One intriguing question is how the visual system efficiently reconstructs
three-dimensional object and scene perception from two-dimensional retinal
information. The stick figure paper
is a brief investigation into the adaptation of classification image techniques
for the study of three dimensional form perception. In this paper, we
emphasize the importance of using a three-dimensional generative model to
create the stimulus space that is used to probe visual processes, even though
the images are displayed in two dimensions.