Hi Jesper, Many thanks for your kind reply and it's very clearly
explained. I'd like to try registering my patient's fa images to the
icbm-dti 81 fa template as you've suggested. May I ask which way of
registration with fnirt is more suitable. 1st way is with the FA_2_FMRIB58
config file (linear then nonlinear registration, and 2nd way is with
tbss_2_reg -t and tbss_3_postreg ( nonlinear then linear registration).

1.  With the 1st way, I follow the 3rd example on the fnirt webpage and
change the reference template (from FMRIB58_FA to ICBM_DTI81 template).
Will the config file parameters be compatible for my reference template :

flirt -ref ${FSLDIR}/data/standard/FMRIB58_FA_1mm_brain -in my_FA -omat
my_affine_transf.mat
fnirt --in=my_FA --aff=my_affine_transf.mat --cout=my_nonlinear_transf
--config=FA_2_FMRIB58_1mm
applywarp --ref=${FSLDIR}/data/standard/FMRIB58_FA_1mm_brain --in=my_FA
--warp=my_nonlinear_transf --out=my_warped_FA

2.  I read how the FMRIB_58 FA brain is created ( linear registration then
repeated iterative nonlinear registration to MNI brain as you have
mentioned in your email). Since the FMRIB_58 is "already in MNI space",is
the tbss_3_postreg required as an additional step after tbss_2_reg to make
it more accurate to register a subject's FA to the "true" MNI space?

3.If i do this with tbss,
i) do I just type tbss_2_reg -t icbm_dti81_FA_template (which is the
target image), followed by tbss_3_postreg?

4.  These two ways diff in the linear and non-linear steps to register a
subjects fa to the MNI space via an FA template, is that to cater for
different purposes?

From what I can interpret, with 1st way using the config file, I first
perform linear registration of my subject's FA image to FA template brain,
followed by non-linear registration with fnirt to the FA template brain.
With the 2nd way using tbss,  I first perform non-linearly registeration
of my subjects' fa images to the FA template brain in tbss_2_reg -t
command, followed by affine linear registration of this FA template to the
MNI brain in tbss_3_postreg, the subjects are then transformed to the MNI
template via this linear and nonlinear matrices. Hence, the 2nd way would
make the subject's FA in the "true" MNI space to do TBSS related
comparisons?

5. Which one would be a more suitable FA to FA registration for my
purpose, which is to get a good registration to the DTI_81 FA template, so
I can use the inverse matrix to backregister the dti81 rois back to my fa
images)? I read that with the tbss way, the white matter tracts would be a
bit expanded.

6.Would it be right to assume that it would be most accurate to register
my images to the DTI_81_FA template or the MNI space in the same way that
DTI_81 template was created for my above purpose?

Many thanks again
Siewmin



Dear Siewmin,
>
>>  I'm trying to compare how the FA values of the white matter atlas rois
>> in
>> my subjects native FA space compares with the FA values of the white
>> matter tracts/rois that are manually drawn on my subjects native FA
>> space.
>>
>> With the 4th question, I'm starting to doubt if I can use the MNI
>> template
>> for my purpose, and would like to ask some advice here.The white matter
>> rois are drawn on an average DTI map of  81 people registered to the MNI
>> template, and I now want to register these rois to my native subjects FA
>> space.  Can I actually use the MNI152 brain for my purpose above even
>> though the rois are not drawn from an average map of these 152
>> participant
>> but from an average map of a seperate 81 participants whose brains are
>> registered to the MNI 152 brain template? Would it still be valid?
>>
>>  Or is it more accurate to use the ICBM-DTI 81 DWI/FA template to
>> perform
>> the normalisation of my subjects DWI/FA to get the inverse
>> transformation
>> matrix to back transform the rois? The DWI or FAcontrast may not be as
>> good as T1 and MNI but the template is the original template that the
>> white matter rois are drawn from. Which way would sound more right?
>
> I think there are actually two different questions here, and for
> pedagogical reasons I'll try and be explicit about them.
>
> 1. Is it valid to use ROI's defined in one instantiation of the MNI space
> one scans registered to another instantiation of the MNI space? Or, in
> other words, is one instantiation of the MNI space as good as another?
>
> I think the answer to this question will depend crucially on how a
> particular instantiation/template of the MNI space was constructed. The
> "true" MNI space is defined exactly by those subjects that was part of the
> initial template, so the question is how close we can get to that with
> another set of subjects. If we assume that there was nothing "odd" with
> the initial sample I think we can assume that we can get fairly close
> (close enough for our purposes) with another sample that has nothing "odd"
> about them. If we register this sample linearly to the original MNI space
> we should be un-biased (with regards to location) though we may have
> relatively little information about smaller structures (these being
> blurred in the averaging). We can improve on this by performing an
> additional non-linear step, but it is then important to use a method that
> doesn't introduce any bias. I would suggest performing an iterative
> registration to the mean of the new subjects, after each iteration
> updating that mean.
>
> 2. Should you use MNI-T1 or ICBM 81-FA information when registering your
> FA data.
>
> Here I actually think you would be better off using the ICBM 81-FA
> informations (though I must admit to not having looked at that particular
> template myself). It is true that to the eye the T1 scan would seem to
> have much richer information, but for the purpose of registering white
> matter tracts the opposite is very much true.
>
> Good Luck Jesper
>
>>
>> Thanks
>>
>> Siewmin
>>
>>  Dear Siewmin,
>>>
>>>> Hi, I have a few questions about trying fnirt to register fa and other
>>>> scalar maps to the
>>>> MNI template, so I can use the invert transformations to put the rois
>>>> of
>>>> the white matter
>>>> atlas back onto the native fa and scalar maps. Apologies for the long
>>>> questions.
>>>>
>>>> I performed linear 6DOF registration of subjects FA to their T1, and
>>>> linear followed by
>>>> non_linear registration of T1 to MNI152. The fa and scalar maps are
>>>> calculated from 4D
>>>> DWI with B0 unwarping/undistortion performed. These maps are 2mm
>>>> isotropic
>>>> and the
>>>> T1 images of the subjects are 1mm isotropic. The MNI template chosen
>>>> is
>>>> the MNI_linear
>>>> template 1mm. I did this following similiar steps to the 2nd fnirt
>>>> example
>>>> script on the
>>>> fnirt website (i.e fmri to MNI via T1) and with slight modification of
>>>> the
>>>> T1_2_MNI152_2mm.cnf. May I ask if the following commands are the right
>>>> way
>>>> and
>>>> quickest way to "concantenate" the two inverse linear T12FA matrix and
>>>> nonlinear MNI2T1
>>>> warp coefficient to transform binary rois from MNI to the native FA
>>>> space?
>>>>  I have also
>>>> listed my questions below about the choice of registration, template
>>>> and
>>>> using these
>>>> appropriate parameters in the config file:
>>>>
>>>> T1_brain and Image_FA_brain (betted) The Image_FA I have is betted so
>>>> I
>>>> don't have a FA
>>>> image with skull.
>>>>
>>>> flirt -ref T1_brain -in Image_FA_brain -out FA2T1_brain -omat
>>>> FA2T1.mat;
>>>> flirt -ref MNI152lin_T1_1mm_brain -in T1_brain -omat
>>>> my_affine_transf.mat;
>>>> fnirt --in=T1 --aff=my_affine_transf.mat --cout=my_nonlinear_transf --
>>>> config=T1_2_MNI152lin_1mm.cnf;
>>>> applywarp --ref=MNI152lin_T1_1mm --in=Image_FA_brain
>>>> --warp=my_nonlinear_transf -
>>>> -premat=FA2T1.mat --out=my_warped_fa2mni_1mm
>>>
>>> This far all looks good as far as I can see.
>>>
>>>> (applying inverse matrix to place ROI from MNI to FA native space)
>>>> convert_xfm �omat T12FA.mat �inverse FA2T1.mat
>>>> invwarp --ref=T1.nii.gz --warp=my_nonlinear_transf.nii.gz
>>>> --out=nonlinear_MNI2T1
>>>> applywarp --ref=Image_FA_brain --in=ROIs_in_MNI_space
>>>> --warp=nonlinear_MNI2T1 --
>>>> postmat=T12FA.mat out=ROIs_in_FAnative_space --interp=nn
>>>
>>> This bit I would do differently. I would combine the linear FA->T1
>>> transform with the non-linear T1->MNI transform, and then invert that
>>> combined warp. I.e. I would do
>>>
>>> convertwarp --ref=MNI152lin_T1_1mm_brain --out=FA2MNI_transform
>>> --premat=A2T1.mat --warp1=my_nonlinear_transf
>>> invwarp --ref=Image_FA_brain --warp=FA2MNI_transform
>>> --out=MNI2FA_transform
>>>
>>>> 1. Is it ok that I use a betted FA image all the way in these steps,
>>>> as
>>>> long as the T1
>>>> image used in FNIRT is the original T1 with skull on?
>>>
>>> Yes.
>>>
>>>> 2. I use the MNI152_lin_1mm template with slight modifications to
>>>> T1_2_MNI152_2mm.cnf ( renaming it T1_2_MNI142lin_1mm.cnf ).  The
>>>> MNItemplate now
>>>> chosen is a lot smoother, is 1mm and of different intensity to the
>>>> other
>>>> MNI template
>>>> used in T1_2_MNI152_2mm.cnf). Apart from modifying the cnf file by
>>>> changing the MNI
>>>> template to the linear 1mm template, and the corresponding brain mask
>>>> ,
>>>> which other
>>>> parameters would be important to change (my T1 and the MNIlin_1mm are
>>>> both
>>>> 1mm in
>>>> resolution? Would there be any recommendations you suggest for the
>>>> some
>>>> of
>>>> parameters in the config file in this circumstance: The current
>>>> settings
>>>> in the
>>>> T1_2_MNI152_2mm config files are
>>>
>>> I would need to understand what you are attempting to do here. By
>>> choosing
>>> a template with higher resolution (1mm) you will incur longer execution
>>> time and greater memory demands. At the same time you have chosen a
>>> template that has much less relevant information at high resolution
>>> (since
>>> the average of the linearly registered brains is very blurred). What
>>> you
>>> are suggesting makes little sense to me off the top of my head, and
>>> maybe
>>> if I knew your motivation for this choice I could help you better.
>>>
>>>> 3. If I perform registration of image fa -->t1 -->to mni, without
>>>> including the -out in the
>>>> command line, the fa imagehas to be resampled once when nonlinear
>>>> transformation to
>>>> the MNI 1mm space is performed. Alternatively, I can use the inverse
>>>> matrix of
>>>> FA2T1.mat (i.e T12FA.mat) on T1. This will register T1 to FA followed
>>>> by
>>>> nonlinear
>>>> transformation of this registeredT1 to the MNI template to get the
>>>> my_nonlinear_transf
>>>> matrix file of the T1(inFA native space) to MNI, which I can use to
>>>> transform FA to MNI in
>>>> one step. With the 1st method, the rois of the white matter template
>>>> would
>>>> be
>>>> transformed onto the raw FA image using the inverse of
>>>> my_nonlinear_transf
>>>> matrix and
>>>> FA2T1.mat as written in the command line above. With the second
>>>> method,
>>>> only the
>>>> inverse of my_nonlinear_transf matrix would be used, without requiring
>>>> the
>>>> postmat
>>>> T12FA.mat. With the Fa_image contrast and resolution, which way would
>>>> be
>>>> more
>>>> precise/accurate to i) register or normalise ( register and resample
>>>> FA
>>>> images to MNI
>>>> template) and ii) back-register the rois (by neighbouring
>>>> interpolation)from the template
>>>> to the raw space of the FA image?
>>>
>>> See my suggestion above.
>>>
>>>> 4. The rois of the white matter atlas is created when normalising to
>>>> the
>>>> MNI152_lin_1mm
>>>> template and not the MNI152 _1mm (the non linear template which has a
>>>> higher
>>>> resolution). If I want to invert transform the rois of the atlas in
>>>> MNI
>>>> space to the native
>>>> space of the fa images, would it be right to use the MNI152_lin_1mm
>>>> template to get the
>>>> transformation matrices(because of how the rois of the atlas has been
>>>> created), even
>>>> though it is of poorer resolution than the other nonlinear MNI
>>>> template
>>>> ?
>>>
>>> Ohh, I see now why you want to use the linear template. I still think
>>> it
>>> makes more sense to use the non-linear template. That would allow you
>>> to
>>> obtain a better registration of your subject into standard space, and
>>> assuming that there is no bias between the linear and the non-linear
>>> templates it should still be valid to transfer your ROI's from the
>>> linear
>>> template.
>>>
>>>>
>>>> 5. Is there any output from running the flirt and fnirt that can be
>>>> used
>>>> to get a measure
>>>> of the precision in the registration methods (apart from visual
>>>> inspections), or there a
>>>> paper of fnirt that mentioned the precision of fnirt? I read that one
>>>> way
>>>> to quantitate the
>>>> registration quality of the rois apart from visual inspection is
>>>> assess
>>>> the amount of
>>>> displacement of x,y, z coordinated of defined landmarks from the MNI
>>>> space
>>>> when they
>>>> are transferred to the normalised FA images?
>>>
>>> No. If there were a way to assess the quality of registration we would
>>> use
>>> that to drive the registration.
>>>
>>>>
>>>> 6. Lastly, on the fnirt website, it mentioned that fnirt method is not
>>>> diffeomorphic by
>>>> consruction with some explainations of the difference. Would that
>>>> matter
>>>> in my case
>>>> whether I use a diffeomorphic by construction method or not for the
>>>> purpose I'm trying to
>>>> achieve here(i.e to try as best to register binary rois from the atlas
>>>> to
>>>> the native fa
>>>> space)?
>>>
>>> No, not as long as it is projected down onto a diffeomorphic transform
>>> prior to inverting it. This should be done when using the default
>>> settings
>>> of fnirt.
>>>
>>> Good Luck Jesper
>>>
>>
>>
>