Dear Clark,
It might be worth checking something even easier first.
We have found that the first_flirt registrations are sometimes
not so good in the hippocampal region when there's a lot of
atrophy. Just fixing this often solves the problem and leads
to good FIRST results. So have a look at the registrations and
if they are not ideal around the hippocampus then try redoing
them with flirt (in the normal way, not with first_flirt) and see
if you can improve them. We've found that a hippocampal
weighting mask which excludes the ventricles can help if the
initial flirt result isn't ideal.
There is significant effort in making a model, so it is probably
best to see if you can get this or the fnirt solution working before
exploring anything else.
All the best,
Mark
On 12 Mar 2010, at 00:21, Clark Johnson wrote:
> Wow thanks for the quick response!
>
> -- I'll have to check on the exact calls we used -- not sure about
> "intref" right now.
> -- We are investigating your FNIRT suggestion.
>
>
> Beyond that, in light of the problems / issues you listed, does it
> seem likely that a custom model would be worth the effort -- If yes,
> then can we discuss getting access to the software?
>
>
> I think we have something like 15 hand drawn segmentations right now
> - the pipeline needs to process about 160 brains at several time
> points -- so we really want to find a valid, reliable, and automatic
> solution.
>
> Thanks for your help
> Clark
>
>
>
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
> Behalf Of Brian Patenaude
> Sent: Thursday, March 11, 2010 3:37 PM
> To: [log in to unmask]
> Subject: Re: [FSL] FIRST: Hippocampus & Older adults
>
> Hi,
>
> Are you running run_first_all? If not, are you using the "intref"
> model
> for the hippocampus?
>
> FIRST can have problems with the hippocampus in severe atrophy cases.
> There are some older subjects in the training set but not much with
> severe
> atrophy of the hippocampus. The problem is caused for two reasons. The
> shape variation is not well represented. The second is the intensity
> variation is not well represented in the training data. In particular,
> with significant atrophy the presence of CSF where either the
> Amygdala or
> Thalamus was bordering.
>
>
> The best solution that I've found given the models provided is to use
> FNIRT to warp the image into MNI space. Then run FIRST on the warped
> image, and transform the surface(s) back into the native space using
> `run_mesh_utils --doWarpMesh -i warp_field -m mesh.vtk -o out.vtk`
>
> Yes, FIRST can accommodate custom models. The software to do it has
> not
> been included in FSL.
> It may be possible to provide you with the software if you're
> interested,
> have to check. May I ask how many hand segmentations you have?
>
> Cheers,
>
> Brian
>
>
>
>
>
>> Dear FSL world:
>>
>> Does anyone have experience using FIRST with scans from older
>> individuals?
>> We are finding that the built-in shape/appearance model for
>> hippocampus
>> does not do a very good job, especially when there is significant
>> atrophy.
>> The documentation indicates the existing model was generated using
>> &children and adults, normals and subjects with pathologies&. We are
>> wondering if FIRST needs a model that more closely matches our
>> sample? We
>> assume that FIRST can accommodate user-generated models for specific
>> regions, but we have not been able to figure out how to generate
>> something
>> based on hand-drawn hippocampi from a small group of older
>> subjects. Can
>> anyone give us some help/advice on this issue?
>>
>> Thanks
>>
>> Clark
>>
>
|