I actually abandoned this for a while so that I could get some real  
work done... but now I'm coming back to it.

If I do the registration step I can't get an identity matrix... it  
always wants to nudge my data a little one way or another. Not a huge  
deal, but noise is noise. I tried initializing with an identity  
matrix, and also specified no search and 3 dof. I tried dof=0, and  
that seems to do a dof=12!

Is there a way to lie to feat? E.g. just populate the reg directory  
with appropriate .mat files or similar. I'm actually writing some  
python scripts to do all this stuff, so if I can set it up in a way  
that's reliable, it's OK if it's brittle... as the scripts will take  
care of any necessary precision.

Cheers,
Dav

On Jan 12, 2009, at 1:45 PM, wolf zinke wrote:

> Hi Dav,
>
> Apparently there was no reply yet, so I give it a try, especially  
> because I think that your point of the group analysis without  
> coregistration to a standard template might be an interesting point.
>
> Dav Clark wrote:
>> Currently, FSL is not doing a very good job coregistering my  
>> various images using FLIRT.  (And also, I can't figure out how to  
>> tell FLIRT to motion correct an EPI BOLD sequence to a frame from  
>> another EPI!).
> There have been several discussions about improving flirt results.  
> Overall I have to say that flirt is doing quiet a good job. However,  
> some of the useful options are available via the command line call.  
> It is important to use an appropriate cost function. Depending on  
> the data type and quality you might get better results by changing  
> the degrees of freedom or by applying a fieldmap undistortion prior  
> coregistration if this data is available. It also might help to get  
> the images roughly aligned before doing Flirt (check out the Nudge  
> tool) in case the data is not already in the same orientation. If  
> you reorient the data several times I would recommend to combine  
> transformation matrices and apply only one final image  
> transformation in order to avoid multiple interpolations (check out  
> the convert_xfm tool for this).
> Usually you could use mcflirt for motion correction, since flirt is  
> coregistering only 3d volumes. However, it is possible to split your  
> 4D epi timeseries (fslsplit) and run flirt on each volume  
> separately. If you search earlier discussions about this topic you  
> might find some suggestions how to implement such a loop in BASH.
>>
>> So specifically - is there a way I can give FEAT some already  
>> motion corrected EPI, do low-level analyses on these, and then do a  
>> higher-level analysis without doing any registration (i.e. - tell  
>> FEAT that they are already registered).
> It should be quiet simple to use motion corrected EPI data for a  
> feat analysis, you just select it as input data and uncheck the  
> mcflirt option for he preprocessing. If you still want to include  
> motion correction parameter for the glm analysis you could create  
> extra regressors for each parameter or use the new nice option of a  
> confound matrix. If you did use flirt on a splitted timeseries you  
> could derive the motion parameter easily from the resulting  
> transformation matrix with the avscale function (using the -- 
> allparams option). Again you will find quiet a bit more information  
> about these issues in earlier discussions.
>
> For the higher level analysis it is important that your data is in a  
> kind of common space, so it would not make much sense to run higher- 
> level analysis without any coregistration to a standard space.  
> However, if you collect data from some kind of single subject where  
> you could assume that after alignment to a reference scan all data  
> is in a common space you might consider to skip the coregistration  
> within feat. I do not know if something like this is possible, but  
> it might help in some cases (for example mine). Anyway, if the data  
> is already aligned, you could run a 'fake' coregistration, where you  
> specify the reference scan again, but you do a 3 dof transformation  
> (translation only) and use the no search option. additionally you  
> can specify an identity matrix ($FSLDIR/etc/flirtsch/ident.mat) as   
> initialisation transform. This is one of the nice hidden features  
> you find when looking behind the GUIs. You can specify such a matrix  
> at the end of the design.fsf file (following the comment # Now  
> options that don't appear in the GUI) and run feat after editing the  
> file.
>
>>
>> More generally - the HTML documentation is incredibly sparse... is  
>> there a meatier source of documentation somewhere?
> I guess you can find far more information in this mailing list,  
> especially if you are able to ask specific question where your  
> problems are described in a well understandable manner. I think  
> there are many people out there happy to help and share their  
> knowledge and experience. You also could check the slides from  
> previous FSL courses and the technical reports to find some more  
> details. Attend the next FSL course. I also find it very helpful to  
> check al the log files to see what actually is done in each step.  
> Many command line tools provide some information (ok, sparse again)  
> about there meaning and usage if you just type them in. You can use  
> for example ls  $FSLDIR/bin/ to get a list of available tools. You  
> also could open bash scripts in the $FSLDIR/bin directory to see  
> what's going on there or check the source code.
>
> I hope I was  able to give you some hints,
> wolf
>>
>> Thanks!
>> Dav Clark
>> -- 
>> CognAc (aka Ivry) Lab
>> UC Berkeley