Create model for spinal cord centerline detection

[jcohenadad]

Our current approach for MS lesion detection/segmentation is to (i) segment the spina cord, (ii) crop it using the segmentation mask and (iii) use the cropped output for training a model.

However, this is sub-efficient, considering that the spinal cord segmentation itself relies on spinal cord centerline detection.

Idea: Train a robust model for centerline detection, which outputs would be used in lieu of the spinal cord segmentation, for training DL models.

[sandrinebedard]

UPDATE (after meeting 2022-11-24)

We discussed the pros and cons of training a model to detect the centerline.

PROS:

• We would not need to correct the segmentations errors
• More efficient since the spinal cord segmentation is used to get the centerline, we could directly use the centerline for the training of our DL models

CONS:

• segmentation is useful to co-register contrast together
• only 5-10% (for T2w) needed manual correction
• To first generate the centerline, we need the segmentation
• the spinal cord segmentation can be used to validate if a lesion is detected in the cord for example

VERDICT: we willl continue with the spinal cord segmentation (corrected) to get the centerline

[naga-karthik]

Closing this as progress is being followed on another issue: neuropoly/idea-projects#15