Try rootlets segmentation on inv2 (MP2RAGE)

[valosekj]

Try rootlets segmentation on the inv2 image from an MP2RAGE acquisition.

Ideas:

• try the contrast translation from @NathanMolinier
• if the contrast translation does not work, fine-tune the model on the inv2 images directly

[NathanMolinier]

I managed to get predictions (using my contrast translation approach) on a T1w image. However results are not optimal due to my current resampling to 1 mm isotropic.

Axial view	Sagittal view

[jcohenadad]

this is very encouraging though! do you think you could re-train a model that resamples to 0.5mm to see if that helps?

[NathanMolinier]

Yes sure ! I will work on that !

[valosekj]

I managed to get predictions (using my contrast translation approach) on a T1w image.

Very cool, thannks!

I compared the original and the "fake" image (i.e. created by your contrast translation), and it seems that after contrast translation, we lose the contrast between the rootlets and the CSF:

Thus, it is not surprising to me that the rootlets segmentation does not work (as we do not see the rootlets after the contrast translation).

However results are not optimal due to my current resampling to 1 mm isotropic.

What do you mean by this? When I check the header of the "fake" image, it's still 0.7 mm. Do you mean that the resolution is actually 1 mm (i.e. the header has not been updated)?

[NathanMolinier]

No the header is right, but I just perform a resampling to the original resolution at the end of my processing.

So what I mean is that:

• I first resample the input image (0.7mm³ iso) to 1mm³ iso
• I perform the contrast translation with the resampled 1mm³ iso image
• I finally resample the output image to its original resolution (here 0.7mm³) which is upsampling in this case.

I think I could technically improve my results for this task if I train the model with a higher resolution. Because I might lose the rootlets when performing my first resampling.

[valosekj]

Another possible approach to try: transfer learning, i.e., pre-train on T2w, finetune on inv2 image from an MP2RAGE

[valosekj]

I believe we can close the issue for now.

For a bit of context:

• we now have not only the inv2 image but also the inv1 and UNIT1 images (details)
• the model r20240523 trained on T2w images worked reasonably well on UNIT1 images multiplied by -1 (details)
• @KaterinaKrejci231054 used the r20240523 T2w model to segment 5 images from hc-leipzig-7t-mp2rage and manually corrected them (details here and here)
• we will train a MP2RAGE model, progress will be tracked in https://github.com/ivadomed/model-spinal-rootlets/issues/65