[Insight-users] Deformable registration for large volumes

[Luis Ibanez]

Hi Somi,

Unfortunately, the registration framework does not currently offer
a direct mechanism for dealing with images that are larger than
the memory of your computer.

Kishore's suggestion of subsampling the images is a reasonable one.

Your suggestion of performing local registrations is also a good
option, and it is one that could be implemented in a consistent
way with a BSplines deformable registration.

Option (a) from your email, is probably the best thing to try first.
Essentially, you could divided your large images into smaller
cubes and perform a TranslationTransform registration between
the cubes of the Fixed image and the cubes of the Moving image.

Note however, that for this to be easy, the Fixed and Moving
images should be quite close in space.  This is also a standard
requirement for a deformable registration anyways...

Once you have the translation that is computed for each one
of the cubes, you could set them as the parameters values at
the nodes of a BSplineDeformableTransform.

Note that this is not fully equivalent to performing a BSpline
deformable registration from the beginning, but at least it is
something that will be possible to do with a restricted amount
of memory.


     Regards,


          Luis


---------------------------------------------------------------------------------------------
On Thu, Nov 19, 2009 at 12:40 PM, somi <seesomi at gmail.com> wrote:
> Hi Kishore,
> My images have resolution of 2 microns isotropic voxel size (typically brain
> MR images have resolution of 1 to 2 mm I suppose).....to load them into
> memory I would have to sub-sample them substantially, unfortunately there
> are some structures which are visible only at the highest resolution...I
> plan to deformably register
> all the images to an atlas.
>
> I vaguely remember reading somewhere that ITK was designed to process
> volumes even when it was not possible to load the volume into
> memory...though I am not sure...
>
> In case it is not possible what is the best approach (other than
> subsampling), for example would any of the following work :
>
> a) Do overlapping block wise deformable registration and then apply some
> smoothing/regularization to the deformation field
>
>
> b) In BSpline deformable registration for example, precompute the similarity
> measures in some neighborhood around gridpoints
>
> or any other approach
>
>
> Regards,
> Somi
> On Thu, Nov 19, 2009 at 12:00 PM, Kishore Mosaliganti
> <kishoreraom at gmail.com> wrote:
>>
>> Hi Somi,
>>
>> I do not know the answer to your question but usually subsampling
>> affects the final transform in a very small way and speeds up the
>> convergence.
>>
>> Kishore
>>
>> On Thu, Nov 19, 2009 at 11:01 AM, somi <seesomi at gmail.com> wrote:
>> > Hi,
>> > Is there a way in ITK to perform deformable registration between two
>> > volumes
>> > if the volumes are very large and cannot be loaded into memory ?
>> >
>> > Regards,
>> > somi
>> >
>> > _____________________________________
>> > Powered by www.kitware.com
>> >
>> > Visit other Kitware open-source projects at
>> > http://www.kitware.com/opensource/opensource.html
>> >
>> > Kitware offers ITK Training Courses, for more information visit:
>> > http://www.kitware.com/products/protraining.html
>> >
>> > Please keep messages on-topic and check the ITK FAQ at:
>> > http://www.itk.org/Wiki/ITK_FAQ
>> >
>> > Follow this link to subscribe/unsubscribe:
>> > http://www.itk.org/mailman/listinfo/insight-users
>> >
>> >
>
>
> _____________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html
>
> Kitware offers ITK Training Courses, for more information visit:
> http://www.kitware.com/products/protraining.html
>
> Please keep messages on-topic and check the ITK FAQ at:
> http://www.itk.org/Wiki/ITK_FAQ
>
> Follow this link to subscribe/unsubscribe:
> http://www.itk.org/mailman/listinfo/insight-users
>
>