HST.583
LAB
3: Improving fMRI signal detection using
physiological
data: Examples from the auditory system
October
2, 2006
(http://web.mit.edu/~jmelcher/www/HST583/Lab3_handout2006.htm)
Copyright: Jennifer Melcher and Irina
Sigalovsky
Contents:
Introduction
Cardiac Gating
Clustered
Acquistion (CVA)
Lab Overview
Guidelines for Laboratory Report
Part
I: Cardiac Gating
Specific Instructions
Part II:
Experimental Design Using Clustered Volume Acquisition
Part IIa: Combining CVA with Cardiac Gating
Part IIa: Specific Instructions
Part IIb: CVA and Temporal Sampling
Part IIb: Specific Instructions
This Lab examines two techniques that use
physiological data to improve fMRI signal detection. One technique, called cardiac gating, is used to improve
detection of activation in brainstem structures, auditory and
non-auditory. The other, called
clustered volume acquisition (CVA), is widely used in auditory studies to
reduce the effects of scanner acoustic noise on activation.
The Lab will focus on two parts of the
auditory system: (1) the inferior colliculus, a major site of converging projections
from both lower and higher brain centers and (2) auditory cortex, located on
the superior temporal lobe.
Part I of this Lab focuses on cardiac gating.
Cardiac gating is used to overcome a technical difficulty associated with functionally
imaging brainstem structures. This difficulty arises because there is
considerable cardiac-related, pulsatile brainstem motion. Cardiac gating avoids
this problem by synchronizing image acquisitions to the subject's heart beat. In this lab, image signal strength is
then corrected in post-processing to account for the variability in interimage
interval (TR) that results from fluctuations in heart rate (Guimaraes et. al., 1998). (For an alternative gating approach that avoids the need for the
correction, but generates more scanner acoustic noise, see Zhang et al., 2006.)
Clustered
Volume Acquisition (CVA)
Part II of this Lab examines CVA. Unlike
cardiac gating, CVA does not use physiological data recorded during each
experiment to improve signal detection. However, the technique is based
directly on physiological data, specifically general information concerning the
temporal characteristics of fMRI responses (i.e., response latency, duration).
There are two main types of acoustic noise in
the imaging environment (Ravicz et al., 2000; Ravicz and
Melcher, 2001). One is an on-going noise
produced by the pumping of coolant to the magnet. The second, more intense
noise is intermittent. It is produced by the scanner gradient coils each time
an image is acquired. The noise can pose difficulties for studies using sound
stimuli by (1) masking the stimuli, and (2) inducing brain activity that is not
related to the stimuli (this noise-related brain activity acts to suppress the
fMRI signal changes produced by the intended sound stimuli).
CVA provides a way to reduce the effects of
the most problematic noise, namely the noise produced by the gradient coils.
CVA involves imaging a volume of slices in a "cluster" and leaving a
quiet interval between clusters (Edmister
et. al., 1999; Hall et. al., 1999).
With this paradigm, the masking effects of the gradient noise can be avoided by
presenting sound stimuli during the quiet interval. In addition, the
suppressive effect of the gradient noise on auditory activation can be avoided
by (1) making the duration of the image cluster shorter than the onset time of
the fMRI response to the first image in the cluster, and (2) making the time between
clusters (TR) longer than the fMRI response to a cluster.
The benefits of CVA for detecting activation
in auditory cortex were illustrated in lecture. In this Lab, you will examine
how these benefits can be extended to subcortical structures by combining CVA
with cardiac gating. When CVA is used with a long TR (e.g., 8 sec), image
signals are sampled far less frequently than in most fMRI studies. The
implications of this lower temporal resolution for experimental design will
also be examined in this Lab.
The organization of the data for this lab is
as follows:
Part I: Cardiac Gating (data are in the directory Lab3.1)
Part II: Clustered Volume Acquisition (CVA)
a. Combining CVA and cardiac gating (data in directory Lab3.2a)
b. CVA and temporal sampling (data in directory Lab3.2b).
The software tool for this lab,
xds, runs from the command
line in the Athena environment. xds is a package for visualizing anatomical and functional
images, statistical maps, MR signal vs. time, etc. It is described in detail in
the Appendix
– "Software Tools" at the end of this
handout.
Guidelines
for Laboratory Report
Your laboratory report should contain answers
to the questions specified below. Do not repeat the lab instructions and avoid
lengthy introductions. Your report should not exceed 4 pages. Conclude your
report with a few sentences summarizing what you learned in the lab.
This part of the Lab is designed to give you
a physical feeling for data acquired with and without cardiac gating. Many of
the analyses you will perform parallel those of Guimaraes et al. (1998), so you may find their paper useful in completing this part of the
Lab. However, the data for the Lab
differ from those of Guimaraes et al. in that they were acquired in a different
imaging plane and using a 3 Tesla, instead of a 1.5 Tesla, scanner.
To get started, type
attach hst.583
add /mit/hst.583/bin
cp -r
/afs/athena.mit.edu/course/other/hst.583/lab_data/lab3 .
Directory Lab3.1 contains two sets of functional images (A.bshort and B.bshort). Both were acquired using a standard fMRI block paradigm, i.e. a
stimulus (continuous broadband noise) was repeatedly turned on for 30 sec and
off for 30 sec. For both sets of data, a single slice was imaged which passed
through the inferior colliculi, as well as other auditory structures (Figure 1). One data set was acquired using a fixed TR (= 2
sec). The other data set was acquired in the same experiment using cardiac
gating. In this case, the oxygen level of blood in the subject’s finger tip was
was measured (instead of the EKG) and used to trigger the scanner. This oxygen
signal peaks after each heart beat and so, like the EKG, varies in synchrony
with the cardiac cycle. Images were acquired every other heart beat (Figure 2) so that the TR would be approximately the same as
for the "fixed TR" data set (i.e., ~2 sec; Note that a typical heart
rate is ~60 beats/ min = 1 beat/sec). Directory Lab3.1 also contains files A.para and B.para
with information on the experimental paradigm. Specifically, they contain a
code for each successive functional image in *.bshort. The code indicates
whether an image was acquired during the "stimulus on" condition (1)
or the "stimulus off" condition (-1).
Figure 1: T2-weighted anatomical image of the functionally
imaged slice. The slice
intersected the auditory cortex and cochlear nuclei, as well as the structures
analyzed in this part of the Lab: the inferior colliculi.
Figure 2: Timing of image acquisitions relative to pulse
oximeter signal during cardiac gating.
The pulse oximeter noninvasively measures the oxygen level of blood in
the subject’s fingertip.
The files LeftIC_func.ovl, RightIC_func.ovl in Lab3.1
contain the coordinates of voxels overlaying the inferior colliculi. These
overlay files will be used below to locate the anatomical structures of
interest on the low resolution functional images. You can use another set of
overlay files (*_anat.ovl ) to
locate the inferior colliculi on a high resolution anatomical image (Anat.bshort) of the functionally imaged slice using xds.
Example:
To display the anatomical image alone, type:
xds -z3 -b Anat.bshort &
To display an overlay of the left inferior
colliculus on the anatomical image, type:
xds -z3 -b -v LeftIC_anat.ovl Anat.bshort &
Then, hit "o".
The files A.stdev.bfloat and B.stdev.bfloat
contain maps of the standard deviation in image signal over time. Each file
contains two maps. The first is the standard deviation during stimulus ON
conditions. The second is the standard deviation during stimulus OFF
conditions. Specifically, for each voxel and condition, the standard deviation
was calculated as follows:
where Si is the signal for the ith image and N is the number of
images in the condition.
We calculate the standard deviation for the
"ON" and OFF" conditions separately so that signal variability
estimates are not obscured by changes in signal corresponding to activation.
Question:
I.1 - Compare data sets A and B
quantitatively using A.stdev.bfloat and B.stdev.bfloat and the *_func.ovl files
provided.
Example:
To extract standard deviation values for left
inferior colliculus from A.stdev.bfloat:
xds -z3 -v LeftIC_func.ovl A.stdev.bfloat &
hit "o" to display overlay file LeftIC_func.ovl
hit "s" to display standard deviation
for voxels in LeftIC_func.ovl
Also examine plots of image signal vs. time
for different voxels using xds, and play a
"movie" of images using xds. Which data set
was acquired using a fixed TR, and which with cardiac gating? Explain your reasoning.
Directory Lab3.1 also includes the file Intervals. This is a text file containing the time interval (in
msec) between each successive pair of image acquisitions in the gated data set.
Heart rate fluctuates, so the time interval between acquisitions is not
constant (e.g., take a look at the data in the Intervals file). This non-constant interval results in
fluctuations in signal because of the uneven recovery of longitudinal
magnetization from acquisition to acquisition. To remove these fluctuations,
image signal can be corrected to the value it would have been if the interval
between acquistions had been fixed (see Guimaraes et al., 1998). The gated data
(either A.bshort or B.bshort), after correction, are in Corrected.bshort. Standard deviation maps for the gated/ corrected
data are in Corrected.stdev.bshort.
Questions:
I.2 - Summarize the most salient aspects of
the correction algorithm (described in Guimaraes
et. al., 1998).
I.3 - Check whether the gated data were truly
corrected by comparing the pre- and post-correction data. Explain why you think
it was (or wasn't) corrected.
I.4 - Was cardiac gating worth the added
experimental complexity? Compare activation maps calculated using fixed (FixedTR.t.bfloat) and gated/corrected (Gated.t.bfloat) data sets.
To view the activation maps superimposed on
the anatomical image:
xds -z3 -b -A FixedTR.t.bfloat Anat.bshort
xds -z3 -b -A Gated.t.bfloat Anat.bshort
These color-coded activation maps indicate
the p-value result of a statistical test applied to each voxel of the imaged
slice (red and yellow correspond to the lowest
(p = 0.01) and highest (p = 2x10-9) significance levels,
respectively). The statistical test was a t-test which compares the mean image
signal during stimulus "on" conditions to the mean signal during
stimulus "off" conditions.
Part II.
Experimental design using Clustered Volume Acquisition (CVA)
Part
IIa: Combining CVA with Cardiac Gating
Animal work has shown that the representation
of sound in neural firing patterns changes considerably from brainstem to
cortex. Suppose you want to examine these changes in humans using fMRI. In
other words, you want to sample activation in the various auditory cortical
areas that cover the superior temporal lobe (i.e. use multislice imaging) and,
simultaneously, detect activation in subcortical auditory structures. This can
be accomplished by combining cardiac gating (to optimize detection of brainstem
activation) with CVA (to avoid the contaminating effects of the gradient
noise). This part of the Lab considers issues related to combining these two
techniques.
Directory Lab3.2a contains a functional image data set (CVA_0{00-10}.bshort) acquired using cardiac gating and CVA with a long
TR. In this case, 11 slices were imaged in a cluster lasting 800 msec (Figure 3, Figure 4). The acquisition of this 11-slice volume was
synchronized to the subject's EKG. A volume was acquired approximately every 8
sec. The intervals between each consecutive pair of volume acquisitions for
this data set are in Lab3.2a/Intervals. The experimental paradigm is coded in CVA.para.
Figure 3: Imaged slices superimposed on a
sagittal anatomical image located near the midline.
Figure 4: Timing of image acquisitions relative to the
subject's EKG for
the functional image data in directory Lab3.2a.
Questions:
II.1 - Slice CVA_004.bshort intersects the inferior colliculi. These data, corrected
for fluctuations in the subject's heart rate, are in CVA_Corr_004.bshort. CVA_stdev_004.bfloat and CVA_Corr_stdev_004.bfloat contain maps of standard deviation corresponding to
the raw (CVA_004.bshort) and the
corrected (CVA_Corr_004.bshort)
data, respectively. Quantitatively compare signal variability in the inferior
colliculi before vs. after correction using the standard deviation maps and the
overlay files in Lab3.2a (LeftIC_func.ovl and RightIC_func.ovl). Does the effect of the correction (or lack thereof)
make sense? Why? Does this result imply anything about the usefulness of
cardiac gating in long TR experiments?
II.2 – Consider a subject with a heart
rate that varies between 75 and 85 beats/min. Suppose you are imaging this
person using CVA and cardiac gating (TR ~8 sec). The image cluster duration is
800 ms. The delay from the time the scanner is triggered by the EKG to the
beginning of the cluster is D = 150 ms (see Figure 4).
Figure 5.
Consider two possible orders of image
acquisition illustrated in Figure 5. In one,
images are acquired sequentially within a cluster from anterior to posterior
(A). In the other, they are acquired from posterior to anterior (B). To ensure
that the slice containing the inferior colliculi is truly cardiac gated, which
slice order would you use? Why? You may want to draw a picture similar to the
one in Figure 4 (i.e., illustrating the timing of image acquisitions relative
to the subject's electrocardiogram).
Part
IIb: CVA and temporal sampling
While the contaminating effects of acoustic
scanner noise can be reduced using CVA with a long TR (e.g., 8+ sec), the price
is diminished temporal resolution. This part of the lab (a) illustrates a
potential pitfall of this lower temporal resolution, and (b) examines ways this
pitfall can be avoided by controlling the timing between image acquisitions and
the auditory stimulation paradigm.
To begin, you will examine the time course of
activation in auditory cortex at high temporal resolution (~2 sec) for two
example sounds. You will then consider the implications of sampling these time
courses at a lower temporal resolution. The example sounds were trains of
repeated noise bursts (a stimulus commonly used in auditory neurophysiologic
and psychoacoustic investigations). The "noise" of each burst sounds
like the static from a radio that is not tuned to a station. For one sound,
bursts occurred in a train at a low rate (2/sec). For the other, the rate was
high (35/sec). Each burst was ~25 ms long. A detailed
examination of the responses to these stimuli can be found in Harms and Melcher, 2002.
Figure 6. Experimental paradigm
Directory Lab3.2b contains two functional image data sets (LowRate.bfloat and HighRate.bfloat) acquired using cardiac gating (TR ~2 sec), but not
CVA. In these cases, a single slice, rather than multiple slices, was imaged in
order to reduce the effects of acoustic scanner noise without sacrificing
temporal resolution. The slice passed through the inferior colliculi and auditory
cortices. LowRate.bfloat was
acquired using low-rate trains of noise bursts as the stimulus, and HighRate.bloat was acquired using high-rate trains of noise bursts
as the stimulus. These data have already been corrected for variations in
signal due to the variations in the subject’s heart rate. They have also been
interpolated to a constant TR=2s for the purpose of averaging described below.
Both data sets were acquired in a block paradigm, i.e. the noise burst train
was repeatedly turned on for 30 sec and off for 30 sec (Figure 6). The total duration of image acquisition for each
stimulus was ~9 min. Notice, however, that each functional data set contains
only 35 images because the data were divided onto 70 sec.-long blocks of time
(with 10 sec overlap) and, then, these blocks were averaged. Specifically, the
first five images correspond to the off condition, the next fifteen images to
the stimulus "on" condition, and the remaining fifteen images to the
stimulus "off" condition.
Lab3.2b also contains maps of activation to high-rate and
low-rate noise burst trains (i.e., LowRate.t.bfloat and HighRate.t.bfloat). These activation maps were calculated using a
t-test to compare image signal during train "on" and "off"
conditions. You can see the maps superimposed on an anatomical image of the
functionally imaged slice by typing:
xds -z3 -b -A LowRate.t.bfloat Anat.bshort &
xds -z3 -b -A HighRate.t.bfloat Anat.bshort &
Question:
II.3 - Compare the activation detected in
auditory cortex for low- and high-rate trains. Describe the similarities and
differences.
File LowRateStim_func.ovl in Lab3.2b contains the coordinates of voxels that showed
activation in auditory cortex to low rate noise bursts. Use xds to
examine image signal vs. time averaged across these voxels for LowRate.bfloat and HighRate.bfloat (see Appendix).
Example:
Display the activation maps on the functional
images by typing:
xds -z3 -v LowRateStim_func.ovl -A
LowRate.t.bfloat LowRate.bfloat &
xds -z3 -v LowRateStim_func.ovl -A
HighRate.t.bfloat HighRate.bfloat &
To display the overlay file, LowRateStim_func.ovl,
hit "o".
To view image signal vs. time for different
voxels, hit "g" (a new window will appear), then hit "N".
Position the cursor at a location of interest on the functional image. The new
window will display signal vs. time for that location. To view signal vs. time
averaged across the voxels in LowRateStim_func.ovl, hit "s".
Questions:
II.4 - Draw the time course of the response
to low- and high-rate noise bursts. Include an indication of the sound on period.
Describe the similarities and the differences between the time courses for the
two stimuli. Note that the stimulus goes on at image #5 and goes off at image
#20 (Images are numbered from 0 to 34.) Given these time courses, explain the
cortical differences in the activation maps for low- vs. high-rate noise
bursts.
II.5 - Suppose you were to use CVA with a
long TR to examine the activation for 2/sec and 35/sec noise bursts. To best
detect activation, how would you time your volume acquisitions relative to the
stimulation paradigm? Explain. For simplicity, assume a fixed TR of 10 sec
(i.e., no cardiac gating). Also assume a stimulation paradigm as in the above
example (i.e. the noise burst train is turned on and off every 30 sec).
The data for 2/sec and 35/sec noise burst
trains indicates that certain attributes of sound (e.g., repetition rate) are
represented in the time course of fMRI activation. Suppose you want to
understand how this representation varies across auditory cortical areas. You
would need to measure activation with high temporal resolution in multiple
slices while minimizing the effects of acoustic scanner noise.
Question:
II.6 - Design an experiment that uses CVA
with a long TR (10 sec), but allows the reconstruction of activation time
courses with high temporal resolution (e.g., 2 sec). What is the
"cost" (if any) of your proposed approach?
In this laboratory, you will use a package
called xds to visualize anatomical and functional images,
statistical maps and MR signal vs. time. xds can be
run from the command line in the Athena environment.
Usage:
xds [-options] inputfile(s) &
Options:
-A display activation map (e.g. -A map_name)
-z define size of the window displayed (e.g. -z3)
-b perform a bilinear interpolation on input
file(s)
-v display an overlay file (e.g., -v overlay_name)
Using the mouse:
- Left button is used to select voxels for
inclusion in a region of interest (ROI)
- Middle button is used to change contrast (hold
it down and move the mouse horizontally) and brightness (move vertically)
- Right button is used to display a time-series of
images in a movie-like fashion.
Displaying signal vs. time:
- Place the cursor in the image window and hit
"g". A new window will appear containing a plot of image signal vs.
time at the location of the cursor. If you move the cursor to different
locations in the image window, you will see the signal vs. time plot update for
each new location.
- Hit "N" to toggle between "full
range with standard deviation" and "local scale with no standard
deviation" modes.
- Click on a particular place in the signal vs.
time window to display a corresponding image in the image window. Alternatively,
type an image number (from 0 to N-1) and hit "enter". Image numbers
are displayed in the lower left corner of the image window.
Obtaining image statistics:
- Select an ROI by either
- Using the left mouse button
- Loading voxel coordinates from the default.ovl
file by hitting "o"
- Hit "s" (statistics will be displayed
in the command window)
- Hit "O" to save an ROI into
default.ovl (copy default.ovl into filename.ovl after)
- Hit "C" to clear (i.e. delete) an ROI
from the image window
Guimaraes et al. (1998) Imaging
subcortical auditory activity in humans. Human Brain Mapping 6: 33-41. [click
here for pdf]
Hall et al. (1999) "Sparse" temporal
sampling in auditory fMRI. Human Brain Mapping. 7: 213-223. [click here
for pdf]
Edmister et al. (1999) Improved auditory
cortex imaging using clustered volume acquisitions. 7: 89-97. [click
here for pdf]
Harms and Melcher (2002) Sound repetition rate
in the human auditory pathway: representations in the waveshape and amplitude
of fMRI activation. J. Neurophysiol. 88: 1433 - 1450. [click
here for pdf]
Ravicz et al. (2000) Acoustic noise during
functional magnetic resonance imaging. J. Acoust. Soc. Am. 108: 1 - 14. [click
here for pdf]
Ravicz and Melcher (2001) Isolating the
auditory system from acoustic noise during functional magnetic resonance
imaging: Examination of noise conduction through ear canal, head, and body.
109: 216-231.
[click here for pdf]
Zhang et al. (2006) Strategies for improving the detection
of fMRI activation in trigeminal pathways with cardiac gating. 31: 1506-1512. [click here
for pdf]