Working with low resolution images

[Tosif Ahamed]

Hi Gary,

I wanted to know to what degree does resolution affects DTITK registration process. I downsampled original DTI images to 64x64x39 and tried to align them, they seem aligned but I can't make out how good the alignment is.

On a related not, can we downsample the deformation field? Let's say I calculated the DTITK deformation field in original space but want to apply the field to a low resolution image. Can I downsample the deformation field to a low resolution?

Thank you.

--
Tosif Ahamed

Research Engineer
Neural Information Analysis Labs
ATR-International,　2-2-2 Hikaridai Seika-cho, Soraku-gun
Kyoto 619-0288, Japan
Tel: +81-774-95-2407

[Hui Zhang]

Hi Tosif,

First of all, thanks for trying DTI-TK.

I wanted to know to what degree does resolution affects DTITK registration process. I downsampled original DTI images to 64x64x39 and tried to align them, they seem aligned but I can't make out how good the alignment is.

In theory, high resolution should potentially allow for higher registration accuracy than using a downsampled version.  It is certainly true that you can get a decent registration at the resolution you tried.  It will preclude some of the finer details to be registered, if they "disappear" from the downsampling.

 

On a related not, can we downsample the deformation field? Let's say I calculated the DTITK deformation field in original space but want to apply the field to a low resolution image. Can I downsample the deformation field to a low resolution?

If you want to apply the deformation field to a low resolution image, you do not need to alter the deformation field itself at all.  I assume you intend to reslice the low resolution image into one that has the original resolution.  In that case, you can simply use the low resolution image as the target volume, via the "-target" flag.

Gary

 

[Tosif Ahamed]

Hi Gary,

Thank you for the reply. That was the first thing I tried, but for some reason the top part of the brain got cut. I think I should mention my problem in a little more detail.

I was wondering if aligning the functional images by using the DTI deformation field would give a smarter alignment. So I want to use the deformation field calculated by DTITK to align and map functional images to standard space. I know this is probably not the task you built DTITK for, but I just wanted to give it a shot.  I have been trying for a couple of days but I keep running into some problems. I am using DTITK alongside SPM.

I first tried coregistering trace images and functional images, which upsamples the functional image to match the resolution of the trace image. Once the trace and functional images are coregistered, I applied the deformation field to the functional images using deformationScalarVolume, in the target I gave a reference image of 64x64x30 (functional image resolution). The resulting images were well aligned between subjects, but the top part of the brain was cut. Is there anything I can do to avoid this?

The other thing I tried was to downsample the trace images first to 64x64x39, coregister functional and trace images and then apply the field. This method does not align the images at all. Is it because the deformation field was calculated using images of a higher resolution?

The only remaining option I have is to work in lower resolution to begin with, calculate the deformation field on the downsampled DTI images and then use those. I haven't tried it yet though, but do you think it is a good idea?

Thank you once again.

[Hui Zhang]

Hi Tosif,

Thanks very much for the context.  DTI-TK can in principle help with the multimodal fusion, even though I have not attempted such applications myself.

If I understand correctly, you want to map your functional images onto the standard space, with the correspondence established with DTI registration.  And you wish the resolution of the resulting images to be similar to the original functional images.  In fact, you want them to have the same voxel spacing as well as the same number of voxels per dimension.

I think this is where your apparent difficulty lies.  If you compute the field of view defined by the functional images and that of the DTI volumes, my guess is that you will find that they are not the same.  Within the current capability of DTI-TK, what you can do is to designate the same voxel spacing as the functional images with the flag "-vsize", something like

deformationScalarVolume -in input.nii -out output.nii -trans input.df.nii.gz -target input.df.nii.gz -vsize 2.5 2.5 2.5

Give this a try and see if it helps.

Gary