I am relatively new to SPM and trying to help a friend with a VBM
study. (I am a programmer, not a medical doctor and not a
statistician. )
Currently we are working with so-called "Optimized VBM" on SPM2 but
after reading some comments on the Wiki, I am now eyeing SPM5. The
reason I am writing this mail is that I am interested in streamlining
the current process with regard to performing statistics. (I have a
dream goal of having everything 100% automated. ) The current
methodology strikes me as being quite inefficient but it is how
things were being done at another lab so it has some inertia.
Background: For the sake of example, say the study contains the
scans of 50 subjects, 25 of which are control. There is a giant
spreadsheet that contains various clinical data about each subject.
Step 1: Run the optimized VBM preprocessing script.
Step 2: Think of a statistic you are interested in looking... e.g.
score on some test.
Step 3: Make a new directory identifying that experiment. Copy all
50 smoothed grey matter images in.
Step 4: From the giant spreadsheet, extract the column of interest.
Create a tab delimited text file with two of the columns used to
identify controls. Carefully order the data in this table so that it
will match the way you can select the image list within SPM. Load
that into a matrix X with "load SomeData.txt"
Step 5: Create a basic model (multiple regression with constant).
Carefully select images to match the order of X. Make the first
column X(:,1) control, X(:,2) non-control X(:,3) interesting
statistic, up to 4 columns.
Step 6: Do estimation, view results.
Step 7: Go to step 2.
There are two parts of this process that strike me as evil. The
first is having to copy all of the images into their own directory
every time you want to do a statistic (step 3). The second is having
to make this separate intermediate table (step 4) where the rows are
carefully ordered to match the images you select. With this
procedure you can typically make 100 SPM.mat files and therefore end
up with 100 copies of the original 50 images. It strikes me as being
very error prone.
If you have to do something simple, like add one new subject to the
study, it literally takes weeks to redo because there is so much
manual interaction. (Normally you should be happy when you get new
data, right?) It would be ideal if we could just add (or remove)
subjects or tweak one aspect of the process and turn the crank.
The other thing that boggles my mind is why you get a different set
of blobs every time you run a different statistic with multiple
regression. I mean, they are the same 50 heads, right... the only
thing that has changed is one column of numbers.
Any comments are greatly appreciated!!!
Ray
|