Region-based formulation of the task

[hermancollin]

This comes to mind because of a recent conversation with a collaborator. For context, they have images with "loose" myelin where the myelin masks sometimes contain holes.

Q: [about the morphometrics] How does it handle if a pixel is included in both the axon and myelin masks (i.e. accidentally)?

A: The axon mask is the one used to compute the axon area. For the myelin, we subtract the axon area from the axonmyelin area, as I said earlier. However, the axonmyelin is obtained by summing the axon and myelin masks. Since they are internally represented as binary arrays, this sum only counts the overlapping regions once. To confirm this, I ran a test using the 2 myelin masks attached, while keeping the same axon mask. Both cases give the same myelin area and myelin thickness.

mask A	mask B

Our current models produce the same information twice: the boundary between the axon and its myelin sheath is predicted in both the axon and myelin masks. This does not affect morphometrics too much because of the above discussion, but it does affect the flow of information during training and over-complicates the task.

For the case of nnUNet, this is not directly an issue because all annotations are given in a single flattened mask (i.e. background = 0, axon = 2, myelin = 1). So during training and at inference time, there is only 1 border between the 2 classes. However, we should still try region-based training (see this post) and compare results.

For SAM, the way I'm currently working does suffer from this issue: the axon and myelin masks both contain this boundary. What I'm realizing is that this is overcomplication of the task has repercussions on the final model. I should simply try to output 2 region masks. The axon mask would be unchanged but the "myelin" mask would now be more of a "whole fiber" mask. The myelin mask could simply be postprocessed by subtracting the axon mask from the nerve fiber mask.