[caret-users] Target space

Fri Aug 31 08:30:17 CDT 2007

Hi Jackie,

Apologies for the long reply that follows, but what you're trying to do 
is non-trivial.

I couldn't get your link to work, but I think you mean this Desai paper:

Desai R, Liebenthal E, Possing ET, Waldron E, Binder JR (2005). 
Volumetric vs. surface-based intersubject alignment for localization of 
auditory cortex activation. NeuroImage 26(4), 1019-1029.
http://www.neuro.mcw.edu/~rhdesai/papers/neuroimage05_desai_flatmap.html

I learned a lot from this paper, and I hold Rutvik in the highest 
esteem. But I have some reservations about this approach, which is 
outdated in some respects. Consider this excerpt from Rutvik's methods:

> Working with a patch as opposed to the entire hemisphere has two main 
> advantages. First, since the patch has a limited amount of curvature, 
> it can be flattened without requiring major cuts to overcome intrinsic 
> curvature. This preserves the topology of the region. This does not 10 
> hold true for an entire hemisphere, which requires cuts to flatten 
> without major distortions. The placement of cuts changes the topology 
> of the surface, since adjacent points on the 3D surface are not 
> necessarily adjacent on the flat map. A second advantage is that the 
> computational time required to flatten and warp the patch is 
> significantly less than the time required for performing the same 
> operations on an entire hemisphere.
>
> For displaying group data, we chose the template anatomy provided by 
> Holmes et al. (1998), which represents an average of 27 scans of the 
> same individual. This “N27” or “Colin” anatomy was processed in an 
> identical manner by preparing a surface model and cutting and 
> flattening the perisylvian patch.

1) The patch approach requires defining boundaries. This isn't hard to 
do on any single subject, but defining clear, rater-free guidelines 
across individuals is quite challenging. And the concern about cuts is 
misplaced with spherical registration; while one might flatten as a 
means to generating an easy landmark border drawing substrate, these 
flattening cuts do NOT affect registration, except where they coincide 
with your user-defined landmarks. For example, our core 6 landmarks 
include the medial wall and calcarine sulcus, which also happen to be 
flattening cuts. But other flattening cuts (e.g., frontal, cingulate, 
temporal, and sylvian) do not affect registration in any way. Spherical 
registration operates on the CLOSED topology, which is cut-free. Also, 
now that Caret's 3D border drawing is much faster and easier than it 
used to be, one doesn't even really need to flatten. This document 
drafts a procedure for registering Freesurfer-generated surfaces without 
flattening:

http://pub:download@brainmap.wustl.edu/pub/donna/NIH/JULIA/TUTORIAL/FreeSurfer_to_CARET.doc

2) Computational constraints in 2007 are not what they were in 
2003-2004, when Rutvik probably did most of his work. Most people these 
days choose computationally expensive methods that require less user input.

3) Rutvik chose colin because he didn't have a better alternative back 
then. Dr. Van Essen's PALS paper was published after Rutvik's. You have 
already decided to use PALS_B12 as a target.

You still might have a compelling reason for sticking with this 
approach, but be aware that flat-based registration just isn't much 
anymore -- at least not in this lab.

It's still possible to register subregions on the sphere to the PALS_B12 
atlas, but this can be done by drawing landmarks on the PALS_B12 
inflated or very inflated surfaces, using the average fiducial surface 
and your individual to carefully delineate the subregion boundaries in a 
principled way.

I can't imagine a scenario where it would make sense to generate 
flattened patches of each PALS_B12 subject; one could use the colin flat 
map deformed to PALS_B12 that comes with many of the visualization specs 
in sumsdb, but again -- I would draw in 3D, on the PALS_B12 very 
inflated surface, using PALS_B12 average fiducial surface in a Window 2 
to help sync the source and target region boundaries. Drawing on each 
PALS_B12 subject's surface is not necessary.

Another thing that isn't quite clear to me is the relationship between 
the individual and group results. If you want to compare them in 
volume-land, then yes -- both need to be in a common stereotaxic space. 
What you do then is within your analysis domain. But when mapping it 
onto the PALS_B12 atlas, put some thought into which mapping algorithm 
makes the most sense:

* most wustl.edu users prefer enclosing voxel for most applications, but 
activation that doesn't intersect the surface precisely is missed; it's 
unusual for clusters to miss it altogether, but the peak may be missed
* max voxel or MCW Brainfish might be appropriate if you're trying to 
give the volumetric method the benefit of the doubt, to make sure peaks 
make it

But keep in mind that usually mapping group results to PALS_B12 is for 
visualization -- not direct comparison. It can be used properly for 
comparisons, but caution/thought is needed.

Donna

On 08/30/2007 01:01 PM, Chuan-Chih Yang wrote:
> (My replies are also inline :)
>
> Am 30.08.2007 18:43 Uhr schrieb "Donna Dierker" unter
> <donna at brainvis.wustl.edu>:
>
>   
>> Hi Jackie,
>>
>> See inline replies below.
>>
>> Donna
>>
>> On 08/30/2007 08:43 AM, Chuan-Chih Yang wrote:
>>     
>>> Hi Donna,
>>>
>>> Thanks again for the suggestions. I decide to chagne to PALS_B12.
>>> Is there an PALS_B12 dataset that are in native space (structural volume
>>> scans)? 
>>>   
>>>       
>> Technically, yes, those scans exist and may be part of the OASIS dataset
>> available from Dan Marcus.  But the anatomical volumes used to generate
>> the PALS_B12 surfaces were originally multiple (3-4) MPRAGES
>> coregistered, averaged, and written to 711-2C atlas space in a single
>> resampling, using Avi Snyder's imgreg.  The PALS_B12 surfaces were
>> generated from the resulting atlas space MPRAGE, and these surfaces were
>> subsequently registered to various atlas spaces using the procedures
>> defined here:
>>
>> http://brainvis.wustl.edu/help/pals_volume_normalization/
>>     
>>> I wanna use Freesurfer to generate the surface version of PALS_B12, and then
>>> cut & flatten the patches,
>>>       
>> Why?  I'm afraid I'm completely lost on what you're trying to do.
>>     
>>> and then use CARET to do the landmark-based
>>> warping. (I would like to select some landmarks on the individual patch & on
>>> the PALS_B12 patch)
>>>   
>>>       
>> Many people use Freesurfer to segment their volumes and generate
>> surfaces, and then use Caret to register those surfaces to the PALS
>> atlas.  But it almost sounds like you're trying to recreate the PALS
>> atlas using Freesurfer-generated surfaces, and this goal confuses me.
>>     
>
> Dear Donna,
>
> I want to compare the volumetric & surface-based alignment methods, as Desai
> et. al in the paper in 2005. I wanna use PALS_B12 instead of N27 as the
> template for displaying the group result.
>
> Here is the link 
> http://www.sciencedirect.com/science?_ob=ArticleURL&_udi=B6WNP-4G4XBWC-2&_us
> er=28961&_coverDate=07%2F15%2F2005&_rdoc=1&_fmt=&_orig=search&_sort=d&view=c
> &_acct=C000003618&_version=1&_urlVersion=0&_userid=28961&md5=aa0015307383f25
> e8f568e8ba0675156
>
> The attemp is to use Freesurfer to generate the flatten patch for both
> individual and template. For volumetric alignment, I would like to use the
> individual zstat and group zstat to overlay onto the individual patch & the
> template patch, respectively.(for displaying purpose) Thanks for Desai's
> suggestions and helpful emails.
>
> For the volumetric alignment, I wanna take the zstat from FSL's flirt, (the
> individual zstat is in native space, the group zstat is in mni152, if I
> understanding it correctly.) For that, I will need to transform the
> individual zstat & the individual patch into mni152 ; and directly overlay
> the group zstat (which is already in mni152) onto the PALS_B12 patch (which
> I will also need it to be normalized into mni152)
>
> For the landmark-based alignment, I would need to select the landmarks both
> on the individual and PALS_B12 patch, and then deform them by using CARET.
> The group zstat which is already in mni152 will be mapped onto the PALS_B12
> patch.(I havn't got the idea yet how to do this by using Caret, am still
> newbies :P)
>
> For the spherical alignment, I would need to use all the individual T1 to
> create the sphere and then registrate them to fsaverage by FreeSurfer
> standard pipeline. For transfering the warped spherical activation to the
> template patch, one will then need SUMA to creat the std meshes for
> individual patch & template patch (Desai, et. al)
>
> Then I will need to select some ROIs on the template patches by using CARET,
> to quantitatively compare the activations.
>  
>   
>> I'm pretty sure Fischl, et. al., have put the OASIS dataset through
>> Freesurfer processing stream, so the PALS surfaces probably do exist in
>> Freesurfer form somewhere, but you would need to consult Dan Marcus or
>> Fischl et. al. on getting those surfaces and establishing correspondence
>> between the OASIS ID's and the Buckner case numbers.
>>     
>
> Shall I still need to get the correspondence between the OASIS ID's and the
> Buckner case numbers, if my attemp is to generate the flatten patch of this
> template? Coz' I was thinking if the PALS_B12 is an average volume of twelve
> MR structrual volumes.So I can generate the flatten patch of this template
> directly by Freesurfer without the need of having the info of the ID# and
> case numbers. 
>
> I am appreciate for any suggestions. :)
>
>   
>> I'm probably misinterpreting your intent, so please clarify your goal.
>>     
>>> Cheers,
>>> jackie
>>>
>>>
>>>
>>>
>>>
>>> Am 29.08.2007 15:12 Uhr schrieb "Donna Dierker" unter
>>> <donna at brainvis.wustl.edu>:
>>>
>>>   
>>>       
>>>> Hi Jackie,
>>>>
>>>> See inline replies below.
>>>>
>>>> Donna
>>>>
>>>> On 08/29/2007 07:16 AM, Chuan-Chih Yang wrote:
>>>>     
>>>>         
>>>>> Dear list,
>>>>>
>>>>> I got a question about the selection of target space (mapping volumes onto
>>>>> surface). My attemp is to map the functional zstat.nii of FSL onto N27
>>>>> template by using caret. I found different target spaces that I can choose
>>>>> from in the 'atlas surface selection' window,
>>>>>   
>>>>>       
>>>>>           
>>>> I hope you have a good reason for selecting Colin as the target surface,
>>>> rather than PALS_B12.
>>>>     
>>>>         
>>>>> in my case--- the zstat of individual in not yet normalized to any space,
>>>>> but the group zstat is already normalized to MNI 305,by FLIRT. I am
>>>>> wondering which target space and altas shall I choose, if I wanna normalze
>>>>> the zstat from FSL onto N27?
>>>>>   
>>>>>       
>>>>>           
>>>> We don't have a version of Colin that was normalized using FLIRT (linear
>>>> affine), but we do have versions of colin in SPM99 and SPM2 space.  (See
>>>> http://brainvis.wustl.edu/help/pals_volume_normalization for an
>>>> explanation of how methods and spaces relate in our terminology.)  You
>>>> could use one of these for your group results, but I'd really recommend
>>>> using the PALS_B12 FLIRT surface as your mapping target for those group
>>>> results.
>>>>
>>>> For your non-normalized individual results, you really can't use any of
>>>> our atlas targets -- colin or PALS_B12.  If you want to map the native
>>>> individual results, you'll need to segment his/her structural volume and
>>>> map to the individual's surface.
>>>>     
>>>>         
>>>>> Is there any difference between mapping the individual zstat and the group
>>>>> zstat from FSL onto N27?
>>>>>       
>>>>>           
>>>> Absolutely:  Mapping individual results to colin -- even normalized
>>>> results -- is a bad idea.  Even mapping normalized individual results to
>>>> PALS_B12 surface isn't a great idea.  Mapping individual results to its
>>>> own surface reconstruction is a good idea.
>>>>
>>>> Mapping group results to colin is no longer recommended, because of the
>>>> limitations inherent in a single subject target.  Colin has strange
>>>> anatomy in some places that cause undesirable mapping results in some
>>>> areas (e.g., left posterior STS and IPL).
>>>>
>>>> Mapping group results to PALS_B12 is the recommended option.
>>>>
>>>> If you're interested in seeing how the individual's results differ from
>>>> the group's, then I can think of two ways to accomplish this, but please
>>>> be advised that I'm not experienced in these kind of analyses, so I
>>>> don't know if a reviewer would be happy with them:
>>>>
>>>> * Normalize your individual to the same target used for your group
>>>> results.  Compute the difference in volume-land.  Map the difference
>>>> volume to the PALS_B12 FLIRT surface.  Consider using the MCW BrainFish
>>>> algorithm.
>>>>
>>>> * Normalize your individual to the same target used for your group
>>>> results; reconstruct your individual's surface; map the individual zstat
>>>> to the individual's surface; and register it to PALS_B12.  Map the group
>>>> zstat to the PALS_B12 FLIRT surface.  Use Attributes: Metric:
>>>> Mathematical Operations to subtract one from the other (or
>>>> caret_command's new metric math, if you have the latest and greatest
>>>> command line utility).
>>>>
>>>> Either way, you're guaranteed to get lots of clusters of differences;
>>>> however, I'm not sure how you'd establish significance criteria.  Any of
>>>> the 12 Buckner subjects that contributed to PALS_B12 will have such
>>>> clusters where its individual surface differs from the PALS_B12 group
>>>> target (i.e., normal anatomical variability is pretty high in humans).
>>>>     
>>>>         
>>>>> (is there a availble N27 altas that is FLIRTed into
>>>>> MNI305 ? )
>>>>>   
>>>>>       
>>>>>           
>>>> No.  You may have a good reason for using Colin, but I'd like to know
>>>> what it is.
>>>>     
>>>>         
>>>>> Thanks much for the help!
>>>>>
>>>>> Cheers,
>>>>> jackie
>>>>>       
>>>>>           