Hi,
You are certainly correct that if you have a large intensity
difference at the boundary of a region (like your 10000/1000/100
simulations) and/or you have a higher resolution than in the
reference image space (and your excel spreadsheet seems to
indicate about 7 times more voxels for the same area in the
native space) then you will certainly see a difference due to
interpolation method.
It is easy to select the interpolation method with flirt: either
use "-interp nearestneighour" at the command line or use the
Advanced Options tab in the ApplyXFM or FLIRT GUIs.
However, if your native space has higher resolution than your
standard space, then nearest neighbour will be effectively
ignoring some voxels in your native space. The better solution
would be to use nearest neighbour with the 1mm standard
space in this case if you are very concerned about the change
in average ROI intensity. You may find that with conservative
thresholds (say 0.9) for the trilinear masks and using the
1mm space that there is little difference between the two methods.
I hope this helps.
All the best,
Mark
On 2 Feb 2008, at 21:06, Li Jiang wrote:
> Hi,
>
> Thank you very much.My friend told me that if using nearest neighbour
> interpolation, the intensity decrease less than using trilinear
> interpolation. I don't know the principle of these two methods. If so,
> When doing flirt, how to choose the interpolation method?
>
> I attach the email here. Could you help me how to explain it?
> Thank you!、
>
>
> Dear All,
>
> I've looked at this transformation/scaling issue in a more robust way.
> The drop in mean intensity value is *not* due to scaling the voxel
> intensity values.
> The drop in mean intensity value is due to trilinear interpolation.
>
> I am attaching two files:
> (1) PDF outlining what I did to isolate and assess the problem;
> (2) Excel spreadsheet of statistics I calculated to assess the
> problem.
>
> If the PDF does not explain what I did clearly, please let me know and
> I'll try to explain again.
>
> The ordering of the data in the plot, from left to right, is:
>
> Truth = mprage in subject space; mask in subject space
> Tri Vol = mprage in MNI-space with trilinear interpolation; mask in
> MNI-space with trilinear interpolation and thresholded in such a way
> as to approximate the same *volume* in subject-space
> Tri 50% = mprage in MNI-space with trilinear interpolation; mask in
> MNI-space with trilinear interpolation and thresholded at 0.50
> Tri NN = mprage in MNI-space with trilinear interpolation; mask in
> MNI-space with nearest neighbor (NN) interpolation
> NN Vol = mprage in MNI-space with NNinterpolation; mask in MNI-space
> with trilinear interpolation and thresholded in such a way as to
> approximate the same *volume* in subject-space
> NN 50% = mprage in MNI-space with NN interpolation; mask in MNI-space
> with trilinear interpolation and thresholded at 0.50
> NN NN = mprage in MNI-space with NNinterpolation; mask in MNI-space
> with NN interpolation
>
> Li
>
> 2008/1/30, Mark Jenkinson <[log in to unmask]>:
>> Hi,
>>
>> If it is an average then the way you rethreshold your mask
>> after transforming it can be very important. For small binary
>> masks there can be a large difference between rethresholding
>> at low levels (near zero) or high values (near one) as this will
>> include/exclude contributions of voxels just outside the
>> original ROI. For large masks this should be less important.
>>
>> Try rethresholding near 1 (if you are not already) and see if
>> the values are more consistent.
>>
>> All the best,
>> Mark
>>
>>
>> On 29 Jan 2008, at 08:04, Li Jiang wrote:
>>
>>> Hi,
>>>
>>> It is an average over an ROI.
>>>
>>> First I genarate a ROI mask on the subject image. Then apply the
>>> subject to MNI transformation matrix (.mat) to the ROI-to-MNI
>>> volume.Then measure the intensity of ROI on the subject image and
>>> that
>>> of ROI-to-MNI on the subject-to-MNI image with Matlab.
>>>
>>> Thanks.
>>>
>>> Li
>>>
>>> 2008/1/28, Mark Jenkinson <[log in to unmask]>:
>>>> Hi,
>>>>
>>>> This still doesn't answer the important question:
>>>>
>>>> *HOW* are you measuring these intensities?
>>>>
>>>> Is it an average over an ROI?
>>>> Is it from a single voxel?
>>>>
>>>> If it is a single voxel, is it a peak value surrounded by
>>>> much smaller (even negative) values?
>>>>
>>>> If it is an ROI, how are you generating the ROIs and
>>>> creating the average?
>>>>
>>>> All the best,
>>>> Mark
>>>>
>>>>
>>>> On 27 Jan 2008, at 05:11, Li Jiang wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> For example, with the methods I mentioned above, for one
>>>>> subject,the
>>>>> intensity of the subjects data is 52.4554, but that of the
>>>>> subjects
>>>>> data in MNI-space is only 36.1254. For another subject, the
>>>>> intensity
>>>>> is 46.7794 and 43.8556, which is similar. I don't know how to
>>>>> explain
>>>>> the different.I know that interpolation will affect a little.
>>>>> Compare
>>>>> trilinear and nearest neighbour, which is better?
>>>>>
>>>>> Thank you!
>>>>>
>>>>> Li
>>>>>
>>>>> 2008/1/26, Mark Jenkinson <[log in to unmask]>:
>>>>>> Hi,
>>>>>>
>>>>>> Those commands are fine for flirt.
>>>>>> But how are you judging that the intensity is decreased?
>>>>>>
>>>>>> Could it be due to including slightly different voxels on the
>>>>>> edge of your ROI mask? When you transform a binary
>>>>>> mask you need to rethreshold it afterwards to decide
>>>>>> whether to be inclusive of small partial volume overlap
>>>>>> (of the mask and the new voxels) or exclusive.
>>>>>> This could be the source of changes in average intensity
>>>>>> over the ROI. In addition, trilinear interpolation will do
>>>>>> some smoothing of the image a little - which affects sharp
>>>>>> peaks particularly.
>>>>>>
>>>>>> All the best,
>>>>>> Mark
>>>>>>
>>>>>>
>>>>>> On 25 Jan 2008, at 17:24, Li Jiang wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> That's what I have done.
>>>>>>>
>>>>>>> * get the transformation matrix file (*.mat) that transforms
>>>>>>> subject-space CBF
>>>>>>> volume to MNI-space
>>>>>>> * get ROI volume in subject-space
>>>>>>> * transform ROI volume into MNI-space using the same *.mat file
>>>>>>> * use flirt:
>>>>>>>
>>>>>>> flirt -in <ROIVol> -ref <MNIVol> -out <ROIinMNIVol> -
>>>>>>> applyxfm
>>>>>>> -init <matFile>
>>>>>>>
>>>>>>> Thanks.
>>>>>>>
>>>>>>> Li
>>>>>>>
>>>>>>> 2008/1/25, Steve Smith <[log in to unmask]>:
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> I'm still not sure you have a problem. How are you judging
>>>>>>>> that the
>>>>>>>> intensity is decreased?
>>>>>>>>
>>>>>>>> Cheers.
>>>>>>>>
>>>>>>>>
>>>>>>>> On 22 Jan 2008, at 06:28, Li Jiang wrote:
>>>>>>>>
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> Thank you very much. I still have a question. The original
>>>>>>>>> subject
>>>>>>>>> image is epi sequence.After I transform this image to
>>>>>>>>> highres T1
>>>>>>>>> image
>>>>>>>>> and then to MNI-space, the intensity of the transformed
>>>>>>>>> sub2mni
>>>>>>>>> image
>>>>>>>>> decreased compare to the original epi image. When transform,
>>>>>>>>> I use
>>>>>>>>> trilinear interpolation and cost function is correlation
>>>>>>>>> ratio.
>>>>>>>>> How
>>>>>>>>> can I improve this problem? Which interpolation method and
>>>>>>>>> cost
>>>>>>>>> function should I apply?
>>>>>>>>>
>>>>>>>>> Li
>>>>>>>>>
>>>>>>>>> 2008/1/21, Steve Smith <[log in to unmask]>:
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> avscale gives you a whole set of different pieces of
>>>>>>>>>> information
>>>>>>>>>> about
>>>>>>>>>> the spatial affine transform, including the average scaling
>>>>>>>>>> (size)
>>>>>>>>>> change. It does not tell you anything about intensities. On
>>>>>>>>>> average,
>>>>>>>>>> in general, intensities don't change upon resampling,
>>>>>>>>>> though of
>>>>>>>>>> course
>>>>>>>>>> any given voxel will change!
>>>>>>>>>>
>>>>>>>>>> Cheers.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 19 Jan 2008, at 15:12, Li Jiang wrote:
>>>>>>>>>>
>>>>>>>>>>> Dear Steve Smith,
>>>>>>>>>>>
>>>>>>>>>>> I use fsl to process the functional MR data. It is a great
>>>>>>>>>>> software.
>>>>>>>>>>> To process the data, I first transform the EPI (subject
>>>>>>>>>>> image)
>>>>>>>>>>> data
>>>>>>>>>>> to T1 volume then to MNI_space with affine
>>>>>>>>>>> transformation. And
>>>>>>>>>>> I'll
>>>>>>>>>>> measure the absolute value from the images have been
>>>>>>>>>>> coregistered to
>>>>>>>>>>> MNI-space . I learned from the lectures and noticed the
>>>>>>>>>>> avscale
>>>>>>>>>>> for
>>>>>>>>>>> Inter-subject Registration. I wonder what's the scale mean.
>>>>>>>>>>> If the
>>>>>>>>>>> signal intensity of the subject image will change after
>>>>>>>>>>> affine
>>>>>>>>>>> transformation or global intra-subject transformation.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Best regards.
>>>>>>>>>>>
>>>>>>>>>>> Li Jiang
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> -------------------------------------------------------------
>>>>>>>>>> --
>>>>>>>>>> --
>>>>>>>>>> --
>>>>>>>>>> --------
>>>>>>>>>> Stephen M. Smith, Professor of Biomedical Engineering
>>>>>>>>>> Associate Director, Oxford University FMRIB Centre
>>>>>>>>>>
>>>>>>>>>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
>>>>>>>>>> +44 (0) 1865 222726 (fax 222717)
>>>>>>>>>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>>>>>>>>>> -------------------------------------------------------------
>>>>>>>>>> --
>>>>>>>>>> --
>>>>>>>>>> --
>>>>>>>>>> --------
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> ---------------------------------------------------------------
>>>>>>>> --
>>>>>>>> --
>>>>>>>> --
>>>>>>>> ------
>>>>>>>> Stephen M. Smith, Professor of Biomedical Engineering
>>>>>>>> Associate Director, Oxford University FMRIB Centre
>>>>>>>>
>>>>>>>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
>>>>>>>> +44 (0) 1865 222726 (fax 222717)
>>>>>>>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>>>>>>>> ---------------------------------------------------------------
>>>>>>>> --
>>>>>>>> --
>>>>>>>> --
>>>>>>>> ------
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>> <flirt_scale.xls>
|