Ivadomed: Exclude some subjects / centers from the MS lesion database?

[charleygros]

I am re-opening this discussion that we already had in the last years: Should we exclude some subjects / centers from the MS lesion database.

Reasons:

• poor image quality of some images (e.g. Milan, some subjects of UCSF)
• ground truth of some images looks questionable --> could misguide the segmentation training

Pros:

• training is based on data we are more confident about

Cons:

• in some extent, we could no longer consider our dataset as "real-life"
• subjective criteria of exclusion?
• time needed to go through all images
• less training data

[olix86]

I'm not familiar with the particularity of each center and I'm all for a real-life dataset, but I think that if the data is considered too poor to be usable by a human expert (which is the case for some subjects/center from what Anne told me), it's probably unrealistic to expect a model to learn from that

[jcohenadad]

i agree we should exclude some data/sites if quality is too questionable. I guess this will be a case-by-case decision. Maybe we should discuss a few cases at our meetings?

[charleygros]

To help the discussion during meeting:

cd duke/sct_testing/large

then:
```
fsleyes sub-lyonOfsep042/anat/sub-lyonOfsep042_T2star.nii.gz derivatives/labels/sub-lyonOfsep042_T2star/anat/sub-lyonOfsep042_T2star_lesion-manual.nii.gz -cm red &

fsleyes sub-milanFilippi004/anat/sub-milanFilippi004_acq-sag_T2w.nii.gz derivatives/labels/sub-milanFilippi004/anat/sub-milanFilippi004_acq-sag_T2w_lesion-manual.nii.gz -cm red &

fsleyes sub-mghCaterina016/anat/sub-mghCaterina016_acq-sup_T2star.nii.gz derivatives/labels/sub-mghCaterina016/anat/sub-mghCaterina016_acq-sup_T2star_lesion-manual.nii.gz -cm red &