jcohenadad
commented
3 years ago I definitely vote for option 1 because:
- we want to maximize the number of contrasts (for better generalization) and of subjects (also for better generalization)
p.s. and thank you for the very nice description of the pros/cons !
charleygros
sandrinebedard
Description
For step 4 Creating softseg of #2 (see comment), we have multiple choices on which contrasts to inlcude in the softseg, we have to consider that registration/image quality may be bad for some contrasts.
Suggestions
✔️: pros ❌ : cons
1.Missing modalities
We QC the registration, run again processing pipeline and add the qc_fail.yml list as an input. Softseg would not include the problematic contrast. Some subjects are already missing some contrasts, we could also include those subjects (we check at first the available contrasts for a subject before starting registration, softseg ...)
❌ With this method, between subjects, the softseg would not be created with the same number of images and the subjects would not have the same modalities. (We would not include in the training the images with failed registration) ✔️We maximize the use of contrasts to create the softseg. ✔️ We maximize the number of images we have.
2. Directlty exclude a subject if registration failed on 1 modality
In this case, we generate softseg with 6 contrasts and we exclude all the images of the subject if registration failed on one of them to only have subjects with 6 images and the same softseg (average from 6 seg)
❌ Maybe a bit extreme, with 6 modalities, we will have less images for training. ✔️ No missing modalities, softseg will be more coherent between subjects (since we want contrast agnostic CSA, I think this point is important) --> may give better results when training
3. Create softseg only with T1w, T2w and T2s
Since some subjects are missing MTS data and dwi, and currently, registration is better on T1w and T2s --> First and last slices of MTS have sometime mediocre data quality), we create the softseg with T1w, T2w and T2s, if the registration on MTS data and dwi is good, than we will use that softseg. We could also include subjects that have missing modalities (as in method 1)
✔️Softseg will be created with 3 images for all subjects. ✔️ We will have more images for training than with method 2. ❌ Softseg only uses 3 contrasts when there are more available.