Print

Print
If you have perfusion images from your ASL data that have good grey-white matter contrast in them, i.e. they are not especially noisy in appearance, then I would use that as the basis of your registration to the structural image using the BBR cost function. This can be achieved fairly robustly using epi_reg, something that is built into the new version of asl_reg and available in the pre-release of BASIL (link to appear on wiki soon I hope).

I am likewise surprised if you are getting better appearing results with the linear structural to standard transformation, but my guess would be that this is to do with propagating errors from the asl to structural transformation.

For masking, I would generally recommend that you transform you regions into the native space of the data and then threshold to create the mask and apply. YOu could also apply a GM mask in this space, using the PV estimates from fsl_anat, this will make a big difference given the different perfusion and grey and white matter. The challenge is that it is hard to get voxels in ASL data with a high proportion of grey matter due to partial volume effects - you can probably only use a threshold of 70% max when working in native space (this isn’t an issue in structural or standard space - expect that the partial volume effects are still there, but hidden, a good reason to work in native space). Another alternative is to apply some sort of partial volume correction - we still don't have a lot of experience with interpretation of results after partial volume correction, but I am encouraging people to give it a go (whilst also doing a conventional analysis) to see if we can understand better to what degree it helps.

Michael


On 25 Apr 2017, at 11:41, Pieter Vandemaele <[log in to unmask]> wrote:

Hi Michael and ASL'ers,

I have some questions on image registration and normalisation and masking.

I almost finished my ASL processing pipeline (which I am proud of ;-) ).

1. The last step is to register the ASL data to the structural to be able to transform to MNI.
I read some of the posts on registration of ASL data to the structural and found 2 contradicting statements:
In the first post (23 January 2014) on the use of asl_reg:

You might try registration with the perfusion image directly, my experience is that this is less reliable, but you might give it a try

https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1401&L=FSL&P=R108084&1=FSL&9=A&J=on&K=1&d=No+Match%3BMatch%3BMatches&z=4

In the second post (23 Jan 2017) and exactly 3 years after the first post (looks like a conspiracy):

My current impression (and something that I am building into the new version of BASIL) is that a good final perfusion image is actually a useful basis for registration - since it inherently contains white/gray contrast - which the control or calibration images tend not do.


https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1701&L=FSL&P=R134794&1=FSL&9=A&I=-3&J=on&d=No+Match%3BMatch%3BMatches&z=4

Could you clarify this more?

2. The preprocessing of the structural scans is done with fsl_anat which gives linear and nonlinear transformations to standard space.
I can use these transformations combined with the asl2struct transformation from asl_reg to transform the perfusion map to standard space.
I ran some tests and it seems the best results are generated by using flirt with the linear transformation matrix instead of using applywarp with the nonlinear warp field which is a bit unexpected. The former results in more brain coverage and symmetry while the latter results in some voids and a asymmetrical image.

Do you have an idea why there is a difference?

3. An off-topic question but an important one.
We would like to extract mean perfusion in brain areas defined by the Harvard Oxford Atlas and compare the results within and between subjects.
What is the best way to do this?

The pipeline
1. Create for each region a mask using thresholding with fslmaths?
2. Mask the perfusion map with this mask.
3. Optional: mask with gray matter map from fsl_anat to eliminate non-gray matter voxels
4. Optional: normalize to 1
5. Calculate statistics?

I am not sure about steps 3. Would this be a good idea?

Best regards,

Pieter
--
Re: Netwerk MR afdeling UZGent
<gifmi logo small.png>
Pieter Vandemaele, MSc-Ing
GIfMI Site Manager
Ghent Institute for Functional and Metabolic Imaging
MR Department -1K12
Ghent University Hospital
De Pintelaan 185
9000 Ghent - BELGIUM
[log in to unmask] tel: +32 (0)9 332 48 20
http://gifmi.ugent.be fax: +32 (0)9 332 49 69
ORCID ID 0000-0002-4523-2476

<pleaseconsidertheenvironmentbeforeprinting.gif>


---
Michael Chappell MEng DPhil
    T: +44 1865 617657
Associate Professor, Institute of Biomedical Engineering, University of Oxford.
Director of Training, EPSRC-MRC CDT in Biomedical Imaging
Governing Body Fellow, Wolfson College, Oxford.
    http://www.wolfson.ox.ac.uk        
Research Fellow, FMRIB Centre