Custom (study specific) templates

[chrisgorgo]

Custom (volume and surface) templates derived from a particular dataset have been shown to improve overlap between participants. It is also the recommended way to deal with smaller head sizes in developmental populations. Both ANTs and FreeSurfer support building such templates.

[chrisgorgo]

I learned a few interesting things about custom templates recently:

• At NIH @jfrenchjr introduced me to an AFNI tool for making custom templates called toMNI_Qwarpar. One thing that struck me as ingenious is the fact that they use the MNI atlas as the coregistration target in the very first iteration (as opposed to using an average of all input T1s). This first iteration is also just an affine transformation. This way the new custom template will resemble the MNI space in terms of location and scaling which is very useful for comparing results across studies.
• At MIT @satra shared some of his experience in dealing with brain scans coming from child and adolescent populations. In his opinion, the in populations older than 4 years cortical folding is fully developed and there is no special need for using custom templates. Those brains could be coregistered to standard MNI templates. Populations younger than 4 years require other special consideration because of changing tissue contrast.

[satra]

@chrisfilo - regarding custom templates - it's still a function of the application, but as group sizes get bigger the need for custom templates reduce. also it's mostly below 2 years of age that tissue contrasts become an issue. (was great to have you here!).

also the mri cvs stream (although long, has been quite useful for kids).

[chrisgorgo]

Thanks for chiming in. It was great to visit your lab! Some of followup questions:

• If you were to design a heuristic to decide if a custom template is needed what would be the number of subjects cutoff? (maybe @binarybottle can chime in here as well)
• Are there any procedures to "normalize" the tissue contrast to values common for older brains (to make standard tools work?)
• When you use the CVS stream on infants do you use the template provided by FreeSurfer? Isn't tissue contrast a problem?

Thanks!

[satra]

@chrisfilo - ages < 2 require a fair bit of custom work, ages > 2 typically don't (other than motion issues, coil issues - e.g., coil too large for head).

[jfrenchjr]

It was very nice meeting you, @chrisfilo.

@toMNI_Awarp is used for the first-pass affine alignment tool before iterative nonlinear transformations are done with @toMNI_Qwarpar. This has been used to create the Haskins pediatric brain atlas, which can be found on the AFNI website. Bouncing off your second question above, tissue normalization happens as the first step in @toMNI_Awarp (by calling 3dUnifize). I haven't worked with older brains, but in pediatric brains it does a nice job.

[zuoxinian]

@chrisfilo @satra nice posts on this topic what I am looking for! Is there any papers published regarding these recommendations or seems you guys pretty concluding this no special need of custom templates for school-age children. I am still thinking of the need of further investigation if we need a customized template for Chinese population, particularly for surface-based processing. Any thoughts? Thanks!

[chrisgorgo]

Check out

Klein, A., Ghosh, S. S., Avants, B., Yeo, B. T. T., Fischl, B., Ardekani, B., Gee, J. C., et al. (2010). Evaluation of volume-based and surface-based brain image registration methods. NeuroImage, 51(1), 214–220. Elsevier B.V. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/20123029

by @binarybottle

[jflournoy]

I want to make sure I'm reading this thread correctly, since both Klein et al (2010) and Ghosh et al (2010) indicate that registration using (e.g., ANTS +) custom templates should improve registration. Is it the case that after another decade of experience, you, @satra, have found that as long as the sample is large enough and participants are older than 4 years, one might as well just register directly MNI152NLin2009cAsym? I wonder if somewhere out there is a paper being written to support this, since we are currently using Ghosh et al 2010 to justify using the custom template approach. Thanks for the discussion here.

[satra]

@jflournoy - the conversation has become even more complicated i would say. anatomical registration is just one part of the story. now you have multimodal registration via MSMAll, Diffusion Map embedding alignment, and epi-T1-based surface registration.

at the end of the day, i would say the decision should be based on what you are doing registration for, and how you may consider validating the registration.

if you are able to tune parameters, i think any registration tool can give you good registration. perhaps one way to validate, not fully adequate is to run freesurfer on your template, and just as we did in klein et al, evaluate how well the template aparc areas overlap with your individuals registered to template. this is not perfect, but may be give you a good idea of registration errors and locations.

sorry there are two many parameters to the puzzle, so an answer is not straightforward.

[oesteban]

Using custom templates is now possible via TemplateFlow. Because this thread is very valuable, I have referenced it from TemplateFlow's documentation