I am struggling to find a meaningful name for those 3D regions in an annotated nrrd volume (Mouse CCFv2 or v3) which are labeled by an identifier which is not a leaf of the hierarchy tree described in 1.json.
Here is an example. In the file annotation_25.nrrd, the Thalamus contains 58834 voxels with id 549, that is, the identifier of the Thalamus in 1.json. This represents approximately 4.5% of the volume of the Thalamus which is not annotated by a name more precise than “Thalamus”.
I would like to understand the circumstances under which such voxels are labeled by a non-leaf identifier (“leaf” refers to the hierarchy tree of 1.json, this means that the “children” list is empty).
Here are my questions:
Ideally, should these voxels with non-leaf identifiers be re-distributed in regions with leaf identifiers?
Ideally, should these voxels be included in new child regions with new names and new identifiers?
Or is it a bit of both?
Visual inspection suggests that it is a bit of both, but I would be very happy to get the opinion of those who created the annotated volumes.
To clarify there are two different levels of annotation strategy.
The first level is at the 12 major structures (Isocortex, Olfactory Areas, Hippocampal formation, Cortical Subplate, Striatum, Pallidum, Thalamus, Hypothalamus, Midbrain, Pons, Medulla, Cerebellum) spanning all the voxels.
The second level is at mid-ontology level (~300 gray matter structures spanning the brain, eg nucleus). At this level, annotation is based on evidence from reference data. Ideally we would like to assigned every voxels to this level however, this is not always practical for example where there is not sufficient evidence.
Your proposal of assigning each disconnected “space” unique identifier looks interesting, not something we thought of at the time. We will take that into consideration in further development.
What kind of analysis do you plan to do? If you are processing voxelized data, you might be able to implement something like you suggest on your own since all the data is fundamentally addressed at a voxel level.
Ideally we would like to assigned every voxels to this level however, this is not always practical for example where there is not sufficient evidence.
This seems to answer my question but I have a slight doubt though. Please correct me if I am wrong: you are assuming that the ontology described by 1.json is complete, so that no new region names nor identifiers should be created when evidence is missing. Ideally, every voxel should be assigned to one of the finest regions named in 1.json. (This holds at least when “finest” is replaced by “mid-ontology level”).
Your proposal of assigning each disconnected “space” unique identifier looks interesting, not something we thought of at the time. We will take that into consideration in further development.
This wasn’t a proposal. I was only considering different options, trying to see which option is the closest to your annotation strategy. My goal is to understand how voxels, say voxels with current identifiers 549, would be labeled in the ideal case, when evidence is sufficient.
What kind of analysis do you plan to do?
We aim at assigning average cell densities for various cell types in the finest level (leaf regions) of 1.json. But we also have to take into account those voxels which are not labeled by a leaf identifiers. At the moment, we only look for a proper name to describe such voxels.
you might be able to implement something like you suggest on your own
If my understanding is correct, assigning voxels with non-leaf identifiers to the closest 3D region with a leaf identifier would be in good agreement with your annotation strategy. Indeed, every voxel “should ideally be assigned to one of the finest regions named in 1.json”. If my latter claim is wrong, I would be grateful to you if you could correct it.