MONAI processing, benchmarking and tests

[ltetrel]

Adds:

• MONAI pre/post processing on NVIDIA GPUs (MPS should also be available in the near future)
• Benchmarking/Profiling on a separate thread using python cprofile and custom profiling for CPU/RAM/cuda (MPS should also be available in the near future)
• Tests including non-regression tests (reference dataset still needs to be generated for each model)

For more information about changes, check associated PRs in original Moose repo.

Tested on gpu with linear interpolation, not b-spline because of VRAM limitations.

Currently, processing is done with simpleITK. Code can be heavily re-factored/simplified if limiting support to only MONAI processing (which also uses itk as backend).

[vercel[bot]]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
moose	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	May 23, 2024 11:08am

[ltetrel]

So current state is that it is working after re-installing everythng in a fresh python venv on ubuntu using those instructions. There are deadlocks on system that have a lot number cores, I will investigate further with a new machine. In the meantime we can disable multi-threading in pre-processing. There is alos the question of high VRAM usage and solutions are either:

1. Enforce high GPU requirement (at least 50GB VRAM)
2. Use linear resampling => needs to be easilly configured by the user (in a yaml file instead of current python file)
3. Increase number of patches
4. Dynamically fall back to CPU if detected RAM usage is too high depending on hardware

[ltetrel]

@Keyn34 The base branch is changed, I will let you merge it,

[LalithShiyam]

@all-contributors please add @ltetrel for adding code and docs

[allcontributors[bot]]

@LalithShiyam

I've put up a pull request to add @ltetrel! :tada:

[LalithShiyam]

@all-contributors please add @Kitware for adding code and docs

[allcontributors[bot]]

@LalithShiyam

I've put up a pull request to add @Kitware! :tada: