packaging nnunet for inference

[naga-karthik]

This PR adds a script for running inference using nnUNet. Currently, the script is tested on joplin with CPUs only. The inference on 1 Zurich subject takes about 30s on average.

[valosekj]

Thanks! 💯 Tested and worked well!

The commands I used:

single image

python run_inference.py --path-images ~/data/sci-zurich/sub-zh01/ses-01/anat/sub-zh01_ses-01_acq-sag_T2w.nii.gz --path-out . --path-model ~/data/nnunetv2/nnUNet_results/Dataset501_tSCIZurichCrop/nnUNetTrainer__nnUNetPlans__3d_fullres

For this specific subject (sub-zh01_ses-01_acq-sag_T2w.nii.gz), inference took 87.33 seconds.

dataset

python run_inference.py --path-dataset ~/extrassd1/janvalosek/data/site_012_lesion_inference --path-out . --path-model ~/extrassd1/janvalosek/data/nnunetv2/nnUNet_results/Dataset525_tSCILesionsZurichGoodSubjects/nnUNetTrainer__nnUNetPlans__3d_fullres

Total time elapsed: 291.84 seconds

[valosekj]

Just an idea - since the script is actually universal, we could consider moving it to ivadomed/data-conversion (but we would need to rename the repo again since data-conversion would not fit).

[naga-karthik]

That's a great idea! Could you please open an issue in the data-conversion repository so that others might have an opinion too?

[valosekj]

That's a great idea! Could you please open an issue in the data-conversion repository so that others might have an opinion too?

👍🏻 I reopened the issue ivadomed/utilities#9 discussing the repo name.

[valosekj]

Okay, the repo has been renamed (ivadomed/utilities#9 (comment)) to ivadomed/utilities. I guess we can push the script there.

[valosekj]

Tested (#40 (comment)) and working fine! I guess we can close this PR and move the script to https://github.com/ivadomed/utilities.

[valosekj]

I guess we can close this PR and move the script to https://github.com/ivadomed/utilities.

Done in ivadomed/utilities@1d8112f.

[jcohenadad]

It would be nice to also have a section in the README to make it easier for people to run the inference, as done here: https://github.com/ivadomed/model_seg_ms_mp2rage#segment-an-image and here: https://github.com/ivadomed/model_seg_mouse-sc_wm-gm_t1#run-inference

[naga-karthik]

Thank you @jcohenadad for the suggestion! It has been added in commit 7fbb42e

[naga-karthik]

Merging as reviewed in a meeting with @valosekj !