Robust cross-modality neonatal atlas registration (T1w template → T2w subject)

1

Hi all,

I am building structural connectomes for preterm neonates (dHCP data, SHARD-preprocessed dMRI) using the MRtrix3 MSMT-CSD → ACT → SIFT2 → `tck2connectome` workflow (following Tahedl, Tournier & Smith 2025). Tractography/connectome construction works well; my problem is the parcellation registration.

To get AAL90 labels into native DWI space I am using:

1. `antsRegistrationSyNQuick.sh -d 3 -t s` from the UNC neonatal T1w template to each subject’s `desc-restore_T2w` image

2. `antsApplyTransforms -n GenericLabel` to warp the UNC AAL90 atlas into subject T2w

3. SHARD’s rigid `from-T2w_to-dwi` transform + `mrtransform -interp nearest` into DWI

The weak link is step 1: a cross-modality (T1w template → T2w subject) SyN registration.

Template->T2w similarity (NCC) sits around 0.65–0.70 and a couple of subjects fail visual QC in mrview (cortical bands drifting off the ribbon, deep GM slightly mislocated). I’ve already ruled out a stale-transform caching bug (forcing a fresh registration), so this looks like a genuine registration-robustness issue rather than a pipeline glitch.

Questions:

  1. Does the level of misregistration I’m seeing look acceptable for neonatal connectome construction, or would you consider these registrations to have failed QC?
  2. Has anyone found a robust approach for T1w-template → T2w-subject registration in neonatal data?
  3. Would you recommend moving to a modality-matched neonatal T2w template rather than registering directly from a T1w template?
  4. Are there advantages to incorporating Draw-EM tissue segmentations (or other anatomical priors) into the registration process?
  5. More generally, how are others handling atlas-to-subject registration when constructing neonatal structural connectomes?

(For completeness: the template is roughly term-equivalent while my subjects are ~35–36 wk PMA, but the QC failures don’t look systematic across subjects, so I don’t think the age gap is the main driver BUT could be totally wrong)

Thank you so much,

Anchal

2

Hi @anchalbhaskar,

Welcome to the forum! I’ll try to go through your questions:

    1. Does the level of misregistration I’m seeing look acceptable for neonatal connectome construction, or would you consider these registrations to have failed QC?
    2. Has anyone found a robust approach for T1w-template → T2w-subject registration in neonatal data?
    3. Would you recommend moving to a modality-matched neonatal T2w template rather than registering directly from a T1w template?

To be fair, I can not judge the quality of your registrations based on the NCC. However, there are a couple of things you could try, if you really want to stick with the original UNC atlas. For example, I would try to avoid cross-modality template-subject registration, that is template T1w to subject T2w. Both (the UNC template and the dHCP) have T1w and T2w, so I would use the same modality for template to subject registration. Also, I think you should use antsRegistrationSyN.sh instead antsRegistrationSyNQuick.sh.

In addition to that, if you don’t matter which atlas to use, you can try other atlases (newer) available that are less “blurry”, like the new UNC atlas or the M-CRIB atlas.

Are there advantages to incorporating Draw-EM tissue segmentations (or other anatomical priors) into the registration process?

Yes, they would improve the overall alignment, you can add the WM and GM segmentations in addition to your contrast image. Note that you’ll need the same data for the template and the subject.

More generally, how are others handling atlas-to-subject registration when constructing neonatal structural connectomes?

You could give a shot to our new 5TT neonatal algorithm, it will give you the 5TT file and the parcellation needed for connectome construction. You just need to align the T2w to you dwi image and then run this algorithm. Let me know if you want any advise of how to use it. I hope it helps!

Best regards,

Manuel

3

I would advise 2 things. Take two steps to go from subject T2 to T1 (affine) and then subject T1 to MNI (Syn) rather than a direct one. But for the latter SyN do not use the quick script use antsRegistrationSyn.sh script. The Quick script is just to get a quick and dirty registration. Without the Quick, it is longer but more accurate usually. We are doing this not for neonatal but stroke where we have sometimes 2D EPI/FLAIR. We go from EPI/FLAIR to T1 and thence to MNI and this seems to work for the most part.