I did get FLOBS working in FSL ver 3.3, which we also had running on our
analysis machine. I had two conceptual questions about the mechanics of
using a set of custom basis functions that I could not gleam from the cited
paper.
First, based on the range of m1, m2, etc. one specifies in FLOBS, are
different HRF shapes (delays, widths, etc.) fit to different voxels in the
GLM analysis? If so, is there a way to generate a map of which shape was
best fit to each voxels
The half-cosine HRF model with parameters m1, m2 etc. along with specified ranges on those parameters is used to generate the basis set. This basis set is optimised to well-represent the HRF shapes that can be created from the half-cosine HRF model with the specified parameter ranges. It is this basis set that is then fit to the data at each voxel, and not the half cosine HRF model.
In fact what gets fit to the data is a design matrix containing what I will refer to as a basis function regressors, where each basis function regressor is the experimental stimulus convolved with a different HRF basis function.
For example, let's assume we have a basis set with 3 basis functions, h_1, h_2, h_3 (note that typically with FLOBs it turns out that these would approximate the canonical HRF (the mean HRF shape), the temporal derivative, and the dispersion derivative), and one experimental stimulus, S. Our design matrix, X, therefore contains three basis function regressors, X_1, X_2, X_3:
X_1=h_1 conv S
X_2=h_2 conv S
X_3=h_3 conv S
where conv is convolution.
In our example B will contain 3 parameter estimates at each voxel, B_1, B_2 and B_3 - one for each of the basis function regressors. And so we can use these to reconstruct the best HRF shape fitted to each voxel. It is simply a case of applying these parameter estimates to the basis functions themselves, e.g.:
In practice you can find the parameter estimates in the stats directory inside the feat directory (the param estimate for regressor 1 in the design matrix is pe1.nii.gz), and the HRF basis functions are in an ASCII file in the flobs directory, e.g. /usr/local/fsl/etc/default_flobs.flobs, and are called hrfbasisfns.txt. So you can load them into something like matlab and calculate the HRF shape at any voxel.
To get the temporal resolution of the HRF basis functions that are in hrfbasisfns.txt, look at to see what the --res option is set to in the logfile in the flobs directory (units are seconds). Typically this should be 0.05secs.
Second, is using only one basis function with a range of timing (m1, m2,
etc.) values a valid HRF?
If you create a basis set that contains a single basis function then this will give you an HRF equal to the mean shape that can be generated from the half-cosine HRF model with the specified parameter ranges. The shape is valid, but with only one basis function this HRF shape is then fixed to be the same at every voxel with no flexibility.
If not, what is the best way to combine, in a
higher-level analysis, the individual contrast activation maps generated for
each basis function when several are used? When using a custom basis set,
the design matrix seems to appear automatically with only one contrast per
basis function and no way to specify a different design matrix, such as
creating a contrast that averages the basis functions. Could all the COPE
images for the different basis functions for all subjects be averaged
together in a single group analysis, or is that statistically inappropriate
for some reason?
Some insight on these issues would be greatly appreciated.
The design matrix and contrasts are generated automatically, but you can go in and change them if you desire - just change the button at the top of the GLM contrasts setup gui from "Original EVs" to "Real EVs" - although there is rarely a good reason to do so.
There are three options you might consider to do group analysis with basis functions:
(1) Pass up the regression parameter estimates for all basis functions into the higher-level group analysis, obtain the group average for each basis function separately and then perform an F-test across them at the group level. However, it is not clear what benefit there would be to doing this. The non-canonical basis functions, such as the temporal and dispersion derivatives, tend to average out to zero at the group level due to the different subject HRF shape variations.
(2) Subsequently, the typically taken approach is to only pass up to the group level the canonical HRF regression parameter estimates. This makes for a simple group analysis, and the benefits of including the basis function at the first level are still felt in terms of accounting for HRF variability that would otherwise cause increased noise in the first level analysis.
(3) Another option that can be taken is to calculate a size summary statistic from the single session analyses (e.g., the root mean square of the basis function regression parameter estimates), and pass that up to the group level. However, it is important to note that this would then require non-parametric (e.g. randomise) methods at the group level than is generally used, as the population distribution of such a summary statistic is likely to be non-Gaussian.
Most people use option 2.
Cheers, Mark.
----
Dr Mark Woolrich
EPSRC Advanced Research Fellow University Research Lecturer
Oxford University Centre for Functional MRI of the Brain (FMRIB),
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
Thank you,
~Keith
-----Original Message-----
Of Steve Smith
Sent: Thursday, October 30, 2008 5:28 AM
Subject: Re: Error loading FLOBS directory in GLM
Hi - yes, this is a bug that was reported last week - we'll announce
the fix as soon as it's sorted.
Cheers.
On 29 Oct 2008, at 15:07, Keith Vogt wrote:
Hello,
I am attempting to create a set of optimal basis functions using
FLOBS. The
flobs report seems to be created OK and a directory with 10 files is
made.
However, when I try to load this directory for convolution as
an "Optimal/custom basis function" in the GLM window, I get an error:
wrong # args: should be "feat5:checkbfcustom w i"
More details are copied in the attached log file. If I ignore this
error and try to
proceed, there is an error in processing the model: "F-test 1 isn't
valid...",
even though I haven't explicitly set an F-test. This error seems to
consistently occur regardless of the number of basis functions used
in FLOBS,
the number of EVs set up, or the basic shape specified. This is my
first time
using the FLOBS utility, so any help is greatly appreciated.
I'm running FLS version 4.0.3, with FEAT version 5.92 on a Linux
platform. I
also tried this on a Windows PC with VMware running the same version
of FSL
and the same result occurred.
Thanks,
~Keith
<FLOBS_in_GLM.log>
---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
---------------------------------------------------------------------------
