[FEATURE] The SRI atlas is not exactly what is used for BraTS data preparation

[sarthakpati]

Describe the bug
The atlas in the repo [ref] and the atlas used for preparing the BraTS data [ref] are different. Here is the actual information from the images:

1. Atlas from this repo:

{'_path': 'C:\\Projects\\preprocessing\\brainles_preprocessing\\registration\\atlas\\t1_brats_space.nii',
 'dimension': 3,
 'direction': (-1.0, 0.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0, 1.0),
 'number_of_components': 1,
 'origin': (239.0, -0.0, 0.0),
 'pixel_id_type': '16-bit signed integer',
 'pixel_type': '16-bit signed integer',
 'size': (240, 240, 155),
 'size_of_pixel_component': 2,
 'spacing': (1.0, 1.0, 1.0)}

2. Atlas used to prepare BraTS data:

{'_path': 'C:\\Projects\\fets_front-end_validation\\data\\sri24\\atlasImage.nii.gz', 'dimension': 3,
 'direction': (1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0),
 'number_of_components': 1,
 'origin': (0.0, -239.0, 0.0),
 'pixel_id_type': '16-bit signed integer',
 'pixel_type': '16-bit signed integer',
 'size': (240, 240, 155),
 'size_of_pixel_component': 2,
 'spacing': (1.0, 1.0, 1.0)}
Finished.

To Reproduce
N.A.

Expected behavior
These should be identical.

Screenshots
N.A.

Environment
N.A.

Additional context:
I don't think this will make any difference WRT computation. But this will cause confusion when someone has generated a clinical image from the original scans and is trying to compare it with one generated from this tool. An extreme edge case, but it might be work looking into.

[neuronflow]

Interesting. We had this issue with the atlas before after Daniel reported that we are not using the official SRI-24 template
#79

We now integrated the official SRI-24 atlas from the link there. Potentially we were using the template BraTS used in the past (or yet another one).

[neuronflow]

side note brain templates are a mess, see also:
#109

[neuronflow]

should we introduce something we call "BraTS atlas" or similar? From my naive perspective BraTS should actually adopt the official SRI-24 atlas.

[neuronflow]

So it seems BraTS is not using the official SRI-24 nor one of the official MNI-152 atlases. I am not sure why that is.

[sarthakpati]

I don't care which atlas is used, as long as the data/pipeline correctly reflect it. Right now, BraTS does not use the official SRI24 atlas. We can change this to use the official SRI24 atlas, but this will be a huge pain to fix the historic data that was prepared from 2018 onwards.

[neuronflow]

Do you know why BraTS uses custom atlases? Luckily, our tool is agnostic, and most algorithms don't seem to care, but maybe we should supply the BraTS atlas, too?

Imho BraTS should transition towards using the official MNI-152.

[sarthakpati]

Do you know why BraTS uses custom atlases?

Nope.

Luckily, our tool is agnostic, and most algorithms don't seem to care, but maybe we should supply the BraTS atlas, too?

The BraTS atlas should be the default.

Imho BraTS should transition towards using the official MNI-152.

We can use the Kofler-Pati atlas, for all I care 🤣. I am fine as long as we provide a solid mapping between the new atlas and the original data.

[neuronflow]

The BraTS atlas should be the default.

There is not one BraTS atlas we can use as the default. The BraTS challenges use different atlases (see issues above).

[sarthakpati]

Regardless, we should provide options for all the atlases that are being used, along with associated mappings to and from the OG atlas (which is the one I mentioned).

[neuronflow]

I don't think we should make one of these atlases of unknown origin the default. So far nobody I spoke to could tell me where these atlases come from. We should not promote their spread? :)

[neuronflow]

Yes, I am wondering whether we should include all the atlases with the package or provide links to download them?

[sarthakpati]

I don't think we should make one of these atlases of unknown origin the default

I can assure you that the one I have linked to is the one used for the BraTS data processing before the cluster of challenges started. I can pinpoint the origin of this to C. Davatzikos' lab at UPenn, where this was used to process all the data. I do not know why this differs from the standard SRI24 atlas, though.

Yes, I am wondering whether we should include all the atlases with the package

I think all atlases should be present as part of the package, but not in the git repo (as it is right now). These (and the example data) should be downloaded at install time [ref]. I am happy to host the data at a publicly-accessible place through MLCommons.