Strategy for creating ground truth

[jcohenadad]

A straightforward approach to set up our training dataset is co-register all contrasts, so that we only need to perform one labeling per patient.

In order to always work in a fixed resolution, we could maybe consider straightening all contrasts, and interpolating them to a 0.5mm iso resolution. This target resolution is a reasonable tradeoff between required resolution (for lesion segmentation) and computation time.

Pros/Cons of labeling in common space:

• We only deal with one label per subject (instead of one per contrast per subject).
• We benefit from having the other contrasts registered (easier to spot artifacts).

Pros/Cons of labeling in native space:

• Pros
  • there is redundancy on the GT (more reliable wrt. intra-rater variability)
  • labeling takes less time because resolution is coarser (e.g. 0.5x0.5x3mm vs. 0.5mm iso)
  • we could combine the GT from each contrast using STAPLE_soft, average, etc.
  • easier to correct a GT starting from the native image (because processing the image might be complicated in some cases, e.g multi-center studies, etc.)
• Cons:
  • If we have many contrasts (eg 5), labeling needs to be done for each of them

3rd option (hybrid):

• perform GT in common space and then apply inverse warp to get it in native space.

[charleygros]

Agree! This is probably the optimal way. Could be a good time gain for our expert.

I only wonder: if a lesion is not visible on one contrast (while being visible on the other) --> it could be confusing for a single channel network. I know we are planning to run multi contrast models.. but .. worth keeping this in mind for the future?

[jcohenadad]

I only wonder: if a lesion is not visible on one contrast (while being visible on the other) --> it could be confusing for a single channel network.

yup! very relevant comment. So far, looking at the three contrasts (t2 sag, t2 ax and t2* ax), i am convinced that each of them contribute to more than noise 😅 .

my only concern is the co-registration, which needs to be perfect

[jcohenadad]

After discussing internally, we decided to do the ground truth segmentations on the native images. Arguments are:

• It takes less time to perform the segmentation on the native resolution (e.g. 0.5x0.5x3) compared to 0.5mm iso.
• Having the segmentation for each contrast (T2, T2*, etc.) would make it possible to then combine each segmentation and thus improve the precision (as opposed to having only one segmentation per subject).
• The co-registration procedure might improve in the future, therefore the segmentations we create now will not be bound to the image registration.

Next steps:

• Set up a procedure for centralizing the manual segmentations

• Set up a procedure for describing the manual segmentation procedure. We could start from the instructions given for the spinal MS lesion mapping project:

  

• Things to add to this procedure:

  • Generation of a JSON file
  • Create a script that loops across subjects, open the image editor and generate the JSON file. Similar to what is done for the spine generic project

[charleygros]

Email from rater # 1 (20/11/2020) says: first time point only and on T2star only.