Some months
back while we were analyzing our data, we started getting very weird results.
Our hypothesis was not turning out the way we had hoped it would. In addition,
we were getting inconsistent results between participants who were scanned
earlier and those who were scanned later.
It took us a
while to figure out what the problem was. We started with observation that in
results section, when we wanted the whole brain activation (regardless of the
threshold (uncorrected 1)), we could find that the top portion of the images
was gone!! That is to say, the parietal region looked like it was just scooped
up.
This was a
bit of surprising result for us, as we had always assumed that our software
package (spm) did everything right as long as we gave the correct sequence of
preprocessing and statistical steps. So, we went back to our steps trying to
figure out what could be the issue. Consulting with some lab seniors and based
on mailing list archives, we thought that probably it’s a normalization error,
where the brains of the participants are not being correctly normalized to the
standard MNI space and hence the sort of deformations that we are observing in
our results.
This lead us
to believe that there must be something wrong with the segmentation procedure, the
program probably is not able to properly classify the white and grey matter
leading to wrong transformation files. However, a check on the segmentation
files showed that they were really fine. Also the T1 (structural) file was also
fine. All this lead to check the very basic file that we had never looked – the
EPI image!! And there it was, a huge crater sitting right at the top of
parietal cortex. So it was clear that there was nothing wrong with the analysis
but the image acquisition itself. How do you get images like that? This was new
to us as we had been following the standard parameters for the acquisition as
followed in the lab.
After
extensive testing and parameter optimization supplemented by reading literature
on image acquisition parameters, we were able to pin down certain factors that
might have caused the signal dropout like the one we observed.
Parameters
like Echo time, slice thickness seemed to be the culprit. We found that in a
3-Tesla MRI scanner, an echo time of 30-35ms is an optimum as well as a slice
thickness of 3 - 3.5 mm is best for lesser signal dropouts.
The physics
behind this being very simple (as I understand it), just as when light travels
from one medium to another, it refracts or bends, in the same way, when RF
pulse changes tissue, there is a dropout because of different tissue type. The
way it changes is controlled by the Echo time TE. So in simple terms, smaller
TE, lesser dropout. Also, dropout is affected by the slice thickness that you keep.
Finally, we
were able to get nice EPI images with lesser dropouts and better resolution
(although we had to increase our repetition time and change the entire
paradigm, but it was all for the better). The incident reminded me that it’s so
essential to know everything you are doing in research and not take any factor
or parameter lightly.
The websites
that helped us understand our problem include the following:
2. http://www.ece.unm.edu/~vcalhoun/courses/fMRI_Spring07/Lecture02_IntroTofMRI.pdf : a very nice pdf for parameter
optimization.
No comments:
Post a Comment