ivadomed to nnunetv2 backend migration

[hermancollin]

Checklist

• I've given this PR a concise, self-descriptive, and meaningful title
• I've linked relevant issues in the PR body
• I've applied the relevant labels to this PR
• I've added relevant tests for my contribution
• I've updated the documentation and/or added correct docstrings
• I've assigned a reviewer
• I've consulted ADS's internal developer documentation to ensure my contribution is in line with any relevant design decisions

Description

This PR replaces the ivadomed dependency with nnunetv2. The old models will unfortunately no longer be supported directly, but they can always be used by reverting to an earlier commit. The main changes are

• addition of model cards to list all available models with download links, intended use, training data details
• refactoring download_models.py where URLs to the 3 standard ivadomed models were hardcoded; this script now only downloads one model at a time (generalist model by default)
• refactoring of segment.py where some ivadomed-related options are no longer available. Input resampling is also applied on our side now (it was previously done in ivadomed itself based on the input px size VS training px size)
• refactoring of apply_model.py; here, we now use the nnunet inference function instead of ivadomed.inference.imed_inference()

Linked issues

Resolves #765 (master issue)
Addresses #797 (download_model refactoring)
Addresses #798 (segmentation CLI refactoring)
Addresses #799 (model cards)

[hermancollin]

for testing, check if URLs exist

[hermancollin]

Had to bump python version to 3.11 but this currently causes problems with CI...

[hermancollin]

For some reason imageio 2.26 is downloaded, which makes all test fail. Despite the pinned imageio<2.28.0, on the master branch, 2.34 is downloaded:

Not sure I understand what's going on.

[mathieuboudreau]

For some reason imageio 2.26 is downloaded, which makes all test fail. Despite the pinned imageio<2.28.0, on the master branch, 2.34 is downloaded:

Not sure I understand what's going on.

Likely a loose version condition in one of the pip dependencies that we install in the conda env file - those won't enfoce our own conditions (annoyingly, and we never found a better solution to this). IVADOMED maybe? I'd remove all pip installations in the file, and see if the correct version gets installed, and then re-added them one at a time

[hermancollin]

Likely a loose version condition in one of the pip dependencies that we install in the conda env file - those won't enfoce our own conditions (annoyingly, and we never found a better solution to this). IVADOMED maybe? I'd remove all pip installations in the file, and see if the correct version gets installed, and then re-added them one at a time

Unpinning imageio fixed it. Now it gets 2.33 I think which is good enough.

[hermancollin]

damn this is something I didn't expect... we trained all nnunet models on grayscale images, meaning they have a single channel. While trying axon_segmentation(), I encountered an image that, although looking black and white, was saved in RGB format (meaning 3 channels). This makes nnunet fail directly and unfortunately, from my experience, this is a very common occurence.

We thus have 2 options:

1. before inference, read all images and if they are in RGB format, save a grayscale conversion in PNG format + update input filenames list
2. use nnUNetPredictor.predict_from_list_of_npy_arrays instead of nnUNetPredictor.predict_from_list_of_files. I don't link this option because it adds a lot of complexity to this PR: we will need to manage the RAM on our side and it will be a problem for large number of inputs. By comparison, I currently use predict_from_list_of_files which is memory efficient.