Hi,
Under ideal conditions, where the PVE was very accurate
and there was little spatial blurring in the original image (which
generally is impossible due to the finite Fourier samples taken,
as well as T2/T2* blurring) then the "right" thing to do would
be to do the most accurate interpolation for the rotations, etc.,
followed by an integration over the new voxel area, as I
suggested. However, things aren't ideal, obviously, so the
details here make less difference.
You are fundamentally limited in what data you can ever
get out of combining a low resolution image with a higher
resolution one. If you want to know how much gray matter
was in a particular perfusion voxel, then the above method
is likely to do as well as any method can. In the lower
resolution image you have to sacrifice the accuracy of the
boundaries - there is no choice. You can upsample your
perfusion, but each sample in the higher resolution space
will be intrinsically carrying information from the whole
neighbourhood, defined by the size of the lower resolution
voxel.
Hope this helps.
All the best,
Mark
On 19 Aug 2009, at 01:52, Mojabi, Pouria wrote:
> Thank you very much Mark for elaborating on this.
>
> As a general strategy, low resolution perfusion image on one side and
> high resolution T1 image on the other side, what direction would you
> chose, T1 -> perf or perf -> T1
>
> What I said in my earlier email was migrating from perfusion
> resolution
> to T1 resolution; by down sampling our grey matter to perfusion
> resolution we are sacrificing accuracy of the boundaries, are we not?
>
> Your thoughts on this are appreciated
>
> Thanks
>
> -P
>
>
>
>
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
> Behalf Of Mark Jenkinson
> Sent: Tuesday, August 18, 2009 3:24 PM
> To: [log in to unmask]
> Subject: Re: [FSL] Partial volume segmentation and trilinear
> interpolation
>
> Hi,
>
> You will find that when you do substantial downsampling, FLIRT will
> apply a pre-smoothing step which tries to incorporate information
> over the size of the new voxel, although it is done with a Gaussian
> which might be too big in this instance. It is this smoothing (or
> blurring)
> which is critical though - not the interpolation - as you don't really
> want to interpolate single points at the centre of the new voxels,
> you want to know what the total PVE is over the whole voxel volume.
> No interpolation method will give you the latter without doing
> smoothing/blurring.
>
> So if you just want to integrate the intensities from the 1x1x1 voxels
> contained in each of the 4x4x6 voxels then the easiest way would
> be to do the following:
> - apply any transformation (rotation, etc) but use a reference image
> that has 1x1x1 mm voxel sizes and trilinear interpolation
> - convolve this resultant image with a 4x4x6 kernel of all ones
> (you can do this with fslmaths and the "-kernel boxv" option)
> - downsample with -noresampblur and trilinear interpolation using
> flirt
> In both of these cases the use of trilinear interpolation versus any
> other interpolation (besides nearest neighbour) won't make much
> difference.
>
> I'm not sure how much difference doing it this way will have on
> your final results.
>
> All the best,
> Mark
>
>
>
> On 18 Aug 2009, at 20:19, Mojabi, Pouria wrote:
>
>> Hi Deniz,
>>
>> FSL only supports {trilinear,nearestneighbour,sinc} with trilinear
>> being the default. The best interpolation from my perspective is
>> splines, and I switched to SPM in order to get that to work.
>>
>> For bringing low resolution perfusion to T1 resolution you
>> definitely don't want to use tri-linear for interpolation. Try sinc
>> and see how it works
>>
>> The only catch is if you chose interpolation methods other than tri-
>> linear it might introduce negative intensities to your image which
>> is perfectly normal considering the specs of interpolation. So you
>> need to zero them out after interpolation.
>>
>> Also, (I have tested this) I faced better results breaking down my
>> T1 image resolution. Meaning, you can define an intermediate
>> resolution, co-register your perfusion to that then co-register to
>> final T1 resolution, I observed better alignment doing it this way.
>>
>> I mean frankly speaking there isn't so much you could do with those
>> crappy quality perfusion images!
>>
>> Hope this helps
>>
>> -P
>>
>>
>> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
>> Behalf Of Deniz Ozgen
>> Sent: Tuesday, August 18, 2009 9:14 AM
>> To: [log in to unmask]
>> Subject: [FSL] Partial volume segmentation and trilinear
>> interpolation
>>
>> Hi,
>>
>> I have a question about how trilinear interpolation works.
>>
>> Briefly I am working on perfusion images with a voxel size of 4x4x6
>> mm. and T1 images with 1x1x1 mm. I have obtained GM segmentation on
>> T1, I have the PVE map. I would like to obtain the perfusion value
>> for GM. Therefore I obtained the linear tranformation matrix from T1
>> to perfusion image using FLIRT.
>>
>> I am trying to find percentage volume of GM for a given perfusion
>> image voxel. I believe it would NOT be very accurate to transform
>> the PVE map to the perfusion image space using ApplyXFM using
>> trilinear interpolation.
>>
>> I understand how trilinear interpolation works while reslicing a low
>> resolution image to get a high resolution image. But I am not sure
>> how/which the PVE values are combined to obtain the PVE map for a
>> lower resolution image using trilinear interpolation.
>>
>> I hope my question is clear enough.
>>
>> Thank you very much in advance.
>>
>> Deniz.
>>
>>
>
|