Greetings, I’m new to Allen and the forum but have been engaged with neuroscience for a very long time. I’m interested in doing some finite element work and would like to inquire about the availability and condition of astrocyte meshes, preferably human. Found some SWC tracings but would prefer an actual mesh - is there such a thing? If not I can provide a Python tool that will do a guesstimate from a tracing. But a real mesh would be much better! Thanks - Brian
Follow-up: it’s come to my attention there’s a real problem in this space, and it starts with the observation that there’s no universally accepted file format for meshes. SWC files are skeletons (only), and OBJ is one of the few formats that can retain object information (like if you want to put some ER into your neuron). Many of the other formats are controlled by private industry (FBX etc). The meshing in MICrONS is mostly unusable in its present form, it would have to be considerably massaged to make it computationally useful. We need the ability to convert a skeleton into a mesh instantly, and vice versa, without having to jump through hoops. It should be doable in Python without needing to invoke C++ compilers. There is software that can generate perfect meshes from highly broken starting points, especially given a little bit of guidance.
The meshes of segmented electron microscopy objects like you see in MICrONS are optimized for visual inspection and manual proofreading. However, rendering the meshes for analysis may require some additional processing, including making the mesh ‘watertight’ over gaps in the segmentation. The degree of correction, and resolution of the mesh required, varies across scientific application. So, we do have one single solution.
Here are some resources you may find helpful, if you have not seen already:
1. Introduction to mesh representation in MICrONS: https://tutorial.microns-explorer.org/quickstart_notebooks/06-cloudvolume-download-mesh.html
2. Example code for downloading multiple astrocytes (at low resolution): https://tutorial.microns-explorer.org/examples/astrocyte_meshes.html
3. Mesh repair with MeshParty: https://meshparty.readthedocs.io/en/latest/guide/repair.html
However, one of the challenges with astrocytes in particular is that the segmented reconstructions have many self-contacts. The underlying segmentation treats these contacts as continuous, even though there is cell membrane in between. This makes the mesh representation of astrocytes convoluted and filled with apparent holes. They are a difficult structure to represent well in mesh format, and we are excited for other researchers to take on the challenge.
Hi Susan, thank you for your reply. Count me in, I can help. (Probably right away). I’d like to understand the problem domain better. The computer can do collision detection and such (heuristics), that’s not hard. But can you yourself distinguish when you look in the microscope, maybe with tiny focal adjustments, or is that also difficult? Is there internal structure associated with these proximities, or some reason we can’t tell the computer to just pull them apart? The result might be a very tiny dx error, but if they’re so close we really can’t tell we’re talking … what … 10 nm? How do these proximities show up in EM, is there trouble distinguishing them there? And, what fraction of the tree are we talking about, just a few leaflets or the entire outer portion of a tree starting at a certain resolution?
Here, let me rephrase that. Can you link me to a dataset that has this issue? I’ll run it through the tools I already have and let’s see what happens. I’ve been building pretty good meshes from horrible starting points, even rough sketches from Blender. Depending on the scientific need for example, one of the possibilities is to repair a tree by selecting from a large number of possible solutions and optimizing around distance (or whatever measure makes sense, energy or cost). This will yield a computationally useful result that can be constrained to match the microscope image. There’s no guarantee that it’ll exactly represent the actual astrocyte, but it might be the best we can do given the starting point.