Investigate topologically bound loss #25
Description
Context
(cc @jcohenadad)
WM and lesions mostly have a regular form. A constraint linked to that could lead the model in the right direction.
Investigation direction
Look into :
- existing work
- convexity measurement
- eccentricity measurement, at least for the WM
- soft label for lesions
- soft label especially for the lesions/GM overlap
- for the GM segmentation, the butterfly shape doesn't seem regular enough to create a regularization based on that. The position relative to the WM could be interesting though, should be investigated.
Note : all of this could be harmful in very early training when the model doesn't really predicts shapes yet, this should be investigated too.