Testing cervical model on hc-leipzig-7t-mp2rage dataset

[KaterinaKrejci231054]

This issue describes the application of the model r20240523 (for dorsal and ventral rootlets) to images from the hc-leipzig-7t-mp2rage dataset.

Related: #45

Processing steps

Testing rootlet segmentation on raw data

For each subject, 3 raw nifti files are provided (labeled as UNIT1, inv-1_part-mag_MP2RAGE, inv-2_part-mag_MP2RAGE). I tried to test the r20240523 model on these data, but no rootlets segmentation was created (see below):

UNIT1:

inv-1_part-mag_MP2RAGE:

inv-2_part-mag_MP2RAGE:

Testing rootlets segmentation on inverse data

Then, I tried to create inverse images (multiplied by -1). The only rootlets segmentation was created on the UNIT1 inverse image (see below):

[KaterinaKrejci231054]

Gifs of 3 subjects (inverted UNIT1 data) with rootlets segmentation (sub-18, sub-22 and sub-30):

sub-18

sub-22

sub-30

[valosekj]

Thanks for testing the model @KaterinaKrejci231054! The predictions for the inverted UNIT1 image (multiplied by -1) look promising! I believe we can leverage them for the model training. I would do the following steps:

1. run the model on inverted UNIT1 images for 5 subjects to get initial rootlets segmentations (done in Testing hc-leipzig-7t-mp2rage dataset #64)
2. correct the predictions; push them to git-annex (done in https://data.neuro.polymtl.ca/datasets/hc-leipzig-7t-mp2rage/pulls/2)
3. train an initial nnUNet model using all contrasts (UNIT1, inv-1_part-mag_MP2RAGE, inv-2_part-mag_MP2RAGE). A big advantage here is that the contrasts are already coregistered, so we can use the same GT for all 3 contrasts.

Note that the recommended nnUNet trainer now is nnU-Net ResEnc L; see here. We can this train two models: one with the default trainer, the second with the nnU-Net ResEnc L trainer.

@jcohenadad, what do you think?

[jcohenadad]

Excellent plan, thank you!

p.s. looks like the top right rootlets are missing for sub-30 but I'm sure you are aware

[valosekj]

Excellent plan, thank you!

Great, thank you for the confirmation! @KaterinaKrejci231054 will work on it.

p.s. looks like the top right rootlets are missing for sub-30 but I'm sure you are aware

Yes, we are aware of this. After running the inference, we will go through the predictions and correct them.

[jcohenadad]

just throwing this out there, because we talk about doing additional GT: #59 (comment)

[KaterinaKrejci231054]

Each contrast shows different shapes of rootlets (see below) - we need to consider which contrast to use for manual corrections of the rootlet segmentation. Maybe we can use all the contrasts and make only one corrected segmentation.

[valosekj]

Each contrast shows different shapes of rootlets (see below)

Interesting! This can be caused by different inversion times for inv-1 and inv-2.

we need to consider which contrast to use for manual corrections of the rootlet segmentation. Maybe we can use all the contrasts and make only one corrected segmentation.

Yeah. This sounds good. Let's try to leverage information from all contrasts to make one good segmentation. Then, this single segmentation can be reused for all contrasts to train a single model segmenting all MP2RAGE contrasts.

[valosekj]

Model training continues as part of #65

--> closing this issue