Poor data quality: STIR

[valosekj]

This issue summarizes STIR images with poor data quality.

• sub-cal149_ses-M0_STIR.nii.gz - FOV does not cover the whole SC in sagittal plane:

- sub-mon006 - has FOV covering the whole brain. The SC has only C1-C6 --> moved to PSIR issue

[valosekj]

TODO: add those subjects to etc/exclude.yml - done in 00bd34c

[Nilser3]

@valosekj
I have found some MS studies that have a lesion-manual.nii.gz file, however these are empty (could be cases of a non-lesion patient). Would it be good to identify them so that they do not participate in training?

[valosekj]

I have found some MS studies that have a lesion-manual.nii.gz file, however these are empty (could be cases of a non-lesion patient).

Good point! IIRC, there are indeed some MS patients with no lesions. To keep track of this, we created an empty lesion-manual.nii.gz mask (i.e., only zeros) for such subjects.

Would it be good to identify them so that they do not participate in training?

Good question! In model_seg_sci project, we use all subjects (i.e., even those with empty lesion masks) to teach the model even such cases.

[jcohenadad]

Would it be good to identify them so that they do not participate in training?

As @valosekj mentioned, it is important for the model to learn true negatives. However, it would be good to double check how the loss is computed in this case during training (ie: if there is no segmentation, there is no Dice score 😉, unless if the background is taken into account, but in the latter case that would provoke strong class imbalance).

[plbenveniste]

@valosekj
I think there is a mistake here: sub-mon006 has only a PSIR file.

• [Referencing this in issue 53 which deals with data quality for PSIR contrast.
• Modifying exclude.yml file under right contrast category

[valosekj]

@valosekj I think there is a mistake here: sub-mon006 has only a PSIR file.

• [Referencing this in issue 53 which deals with data quality for PSIR contrast.
• Modifying exclude.yml file under right contrast category

Correct. Good catch! Thanks!