Hi,
A few options exist:
Liner vs non-linear:
You could use linear or non-linear registration to align a given
subcortical structure. The linear alignment may very well be sufficient at
the level of aligning a single structure. If you are just aligning the
masks, I would tend towards linear since there are no intensities within
the structure to drive the non-linear reg (only boundary); i.e. use the
more constrained registration.
To linearly align the masks you could use FLIRT with least squares cost
function.
If you using 6 or 7 degrees of freedom, run_mesh_utlils command line tool
(in the most recent patch) can be used to calculate the transformation
between surfaces out from FIRST and outputs the transformation matrix;
this could also be used to initialize the FLIRT or FNIRT registration.
Masks versus Masked Intensity:
Depending the subcortical structure, the image resolution, SNR you may
wish to use a combination of intensity and the masks from FIRST to align
the structures. If you're interested I could dig up some scripts for
applying FNIRT to the T1 image masked by the FIRST output (it actually
applies FNIRT to an ROI of the image). The idea being that it could
pick-up the intensity differences within the subcortical structure (e.g.
sub-nuclei of the thalamus). ***This has not been extensively tested and
may or may not provide better results***
Cheers,
Brian
> Hi!
>
> Thanks everyone for your responses. Brian, I would be very much
> interested
> in finding out more details!
>
> Thanks
> John
>
>
> On 6/2/09 1:32 PM, "Brian Patenaude" <[log in to unmask]> wrote:
>
>> Hi,
>>
>> Provided that the diffusion data has been register to the T1, you could
>> potentially use the subcortical masks or surfaces from FIRST to register
>> specific subcortical areas. I could provide more details if you're
>> interested in pursuing this.
>>
>> Cheers,
>>
>> Brian
>>
>>
>>> Dear John,
>>>
>>>> We are looking at subcortical areas in a diffusion study.
>>>> Registration
>>>> has so far proved challenging and;
>>>>
>>>> 1- I was wondering if someone could explain how FNIRT could use
>>>> nearest
>>>> neighbor interpolation during the registration process. I noticed that
>>>> FLIRT can do this, but I don't quite understand how this could apply
>>>> to
>>>> FNIRT. Does nearest neighbor interpolation make sense with a non-
>>>> linear
>>>> registration method?
>>>
>>> There is no option to use nearest neighbor (nn) when estimating the
>>> warps (though you can use nn when resampling your data once the warps
>>> have been calculated).
>>>
>>> It would be very difficult to use nn in an estimation scheme since it
>>> would make the derivatives highly non-linear.
>>>
>>>> 2- Additionally, how else can I control registration so we are not
>>>> risking losing small structures that could be wiped out by the
>>>> smoothing
>>>> process? Can anyone give us advice on how to boost our sensitivity?
>>>
>>> I don't quite understand what you mean here. There shouldn't really by
>>> any smoothing going on here, except for a small amount of smoothing
>>> introduced by the interpolation. And I would be very surprised if that
>>> was enough to wipe out structures.
>>>
>>> Could it be that you have data with higher resolution than your "--ref
>>> space" (as defined by your template)? In that case you may end up not
>>> sampling all points in your original image when creating your
>>> resampled image. The solution to that is to use the --super switch
>>> with applywarp.
>>>
>>> Good Luck Jesper
>>>
>>>
>>
>>
>
>
|