Rodent brain?

[oesteban]

How unfeasible would be to transfer learning a model on to a rat brain? What would it be necessary (e.g., number of animals/scans and input types -anatomical image T1w or T2w?, segmentation-?

paging @eilidhmacnicol

[satra]

kwyk wouldn't be directly applicable, since it is a segmentation model. brainy on the other hand may be more applicable if you are simply looking for skull stripping. if you are looking for parcellation, then it will depend on having labeled volumes.

and by rodents do you mean mouse and rats?

T2 support will come before the end of the year for both models.

also pinging @Aakanksha-Rana

[eilidhmacnicol]

Support for both mouse and rats would be excellent but, currently, TemplateFlow only has rat templates so that has been our priority (at least until we can find an appropriately licensed mouse template).

I don't think FreeSurfer-labelled volumes would be possible, but other labelled volumes can be.

[Aakanksha-Rana]

Labelled volumes would be helpful for transfer learning a model for animal scans, but these labels correspond to parcellation? and how many volumes are there?

[eilidhmacnicol]

To be clear, we don't have labelled volumes - yet; we are still assessing what would be needed on our part.

As Freesurfer is largely incompatible with non-human scans, I am unfamiliar with the segmentation outputs. There are several rodent atlases and prior probability images, so segmentation (into discrete or probabilistic volumes of various levels of parcellation) is achievable with other packages.

[oesteban]

To be clear, we don't have labelled volumes - yet; we are still assessing what would be needed on our part.

But we could generate and curate a small-ish dataset that has been labeled through atlas-based segmentation (WXH?). The question is rather "what would be the minimum amount of examples you would need in this transfer learning?"

[Aakanksha-Rana]

To be clear, we don't have labelled volumes - yet; we are still assessing what would be needed on our part.

But we could generate and curate a small-ish dataset that has been labeled through atlas-based segmentation (WXH?). The question is rather "what would be the minimum amount of examples you would need in this transfer learning?"

if it is about skull stripping, It be can be done with number of examples as low as 100 (but, one need to experiment to have a better idea). But if it is about parcellation, like satra said, we would need some good number of volumes, around a thousand maybe.

[satra]

for rats, we could check with fewer samples. their brains are a little more prototypical to each other :) compared to humans. so something like synthseg may work pretty well for that data with small number of samples.

[eilidhmacnicol]

if it is about skull stripping, It be can be done with number of examples as low as 100

We tested artsBrainExtraction (https://github.com/nipreps/nirodents/blob/f6f8b1cbb223ffe31ebd91c28f868b3628c85f07/nirodents/workflows/brainextraction.py) on more than 100 rats, so it sounds like we can easily provide data for skull stripping (we can even add more that have been acquired since, for good measure!).

But we could generate and curate a small-ish dataset that has been labeled through atlas-based segmentation (WXH?).

We can certainly try to optimise atlas-based segmentation for the same set of images. Waxholm atlas is definitely the most detailed, but the parcellations are mostly subcortical (where signal intensities are more ambiguous). The cortex is only defined by a single label...

something like synthseg may work pretty well for that data with small number of samples

What type of input would synthseg require? Discrete segmentations for each subject?