Implement longitudinal processing stream for MAGeT & morpho?

[mtpark89]

Issue: MAGeT Brain (as currently implemented) seems to lack sensitivity for detecting longitudinal changes (work with Tejas).

This raises the need for either:

1. Developing (& validating) a longitudinal processing stream for MAGeT Brain, that ideally preserves the original design (pairwise atlas-template-subject registration scheme) of MAGeT Brain
2. Segmenting baseline images only and applying between-timepoint (within subject) non-linear registrations.

Either way- think something/plan is needed before we process another large longitudinal dataset- in the interest of being frugal with computing resources.

1. Potential longitudinal stream designs See google doc.
2. Validation datasets MIRIAD: http://www.ncbi.nlm.nih.gov/pubmed/26275383 CoRR

[gdevenyi]

Can you provide a word-wise description of your google doc picture?

[mtpark89]

-Assuming that all timepoints are first rigidly registered (to the baseline image)

***Template-subject registrations -For each template- compute the non-linear registrations for timepoint 1 and timepoint 2 images. -Subtract registrations: (template -> timepoint 2) - (template -> timepoint 1) -This estimates non-linear, between-timepoint differences based on one template only. For 20 templates, 20 (subtracted) registrations that estimates the differences between timepoints. Added benefit of operating on existing registrations in MAGeT Brain to average out error, analogous to our original design.

***Within-subject registrations -Compute both forward (timepoint 1 -> timepoint 2) and backward (timepoint 2 -> timepoint 1) non-linear registrations, then average the two reg. fields- similar to those methods used by UCL group.

***Potential final implementations

1. Propagate registrations through all atlas-template-timepoint 1-timepoint 2 pathways- to estimate volume, displacement (shape), surface area for all timepoints. For additional timepoints, atlas-template-timpoint 1-timepoint 2-timepoint 3 ... timepoint n.
2. Modulate the within-subject registrations (forward + backward/2) by all template-subject registrations {(template -> timepoint 2) - (template -> timepoint 1)}--which can be treated as voxel-wise null distributions of non-linear estimations. Fancy-math-that-I-don't-know-yet to compare within-subject registration vectors to template-subject distributions, to correct (forward + backward/2). Apply (forward + backward/2)-corrected to baseline segmentations & surfaces.
3. I don't know.
4. I don't know.

[gdevenyi]

Infrastructure needed:

• structure for adding extra timepoints per subject (basename_2 basename_3 etc containing files?)
• compute rigid registration between timepoints and baseline timepoint
• use rigid registration as "--initial-fixed-transform" for all future timepoints so that all deformation fields end up in common "subject" space.

Questions: Do affine between timepoints instead of rigid? If so, need to extract determinant.