Release Notes (v185)
This is the first public release of this dataset, so this will describe the state of the data here. Subsequent release notes will focus on what has changed since the previous version.
Proofreading: All of the neurons with cell bodies at least partially within the volume have had their axons proofread and extended as far as the consortium feels is possible with confidence. Their dendrites have also been proofread and extended as much as is possible. Efforts have been made to connect orphan spine heads to their dendritic compartments that are associated with synapses between the pyramidal cells in the volume. This means the ‘soma subgraph’ has been proofread thoroughly. Read Dorkenwald et al. 2019 for details. The IDs of these proofread cells can be found in soma_valence table under the ‘e’ and ‘i’ cell types. There are some glial cells that still have merger problems.
In addition, the automated synapses that compose the pyramidal cell subgraph have been manually QCed and false positive synapses identified. The resulting QCed list of 1,961 synapses between 334 excitatory cells are available as a csv and was the basis for all analysis presented in Dorkenwald et al 2019. This also contains the spine head volumes of those synapses.
Proofreading is ongoing on this dataset, but systematic efforts have not progressed to describe fully as of this release. In other words, analysis of objects outside of the soma subgraph likely contain false merges and splits and users should be wary of blind analysis.
Voxel Resolution: The sample was imaged on a TEM microscope whose camera pixels were calibrated to be 3.54 x 3.54 nm. The sections were cut using a Leica UC-7 with a nominal section thickness of 40 nm. Due to the limitations of neuroglancer, we have entered the resolution of voxels in the segmentation and meshes as [4, 4, 40] x, y, z. The synapse table and soma valence tables are also using [4, 4, 40] to convert voxel locations to nm for the x, y, z location expressed in nm. In papers we have been applying post hoc corrections to the mesh and synapse coordinates of [.885, .885, 1] when doing measurements.
We are aware that there are sectioning artifacts which cause compression of the sections. We have fit identified corresponding cell body locations in an in vivo two photon stack and the reconstructed aligned volume. We then fit an affine transformation to those correspondences that transforms points in the EM volume to the in vivo stack. By transforming unit vectors along the cartesian directions of the EM volume we estimate that the effective voxel size in nm is (3.38, 4.22, 36.78]. There are however some uncertainties as to how well calibrated the in vivo stack stages are and how precise the affine registration is, so have generally been reporting distances using the nominal imaged and sectioned distances.
Soma Valence Table: This table was generated by manual identification of cell body locations and calling of cell types. We are aware of a few errors in this table, but it is being provided as is, in order to completely and precisely reproduce the results reported in Dorkenwald et al. . Those errors include, two excitatory neurons [648518346349538733, 648518346349539433] which were misidentified as inhibitory early in proofreading when their dendrites were not complete. There is also one neuron [648518346349527116] marked as ‘uncertain’ which is clearly excitatory. There are also a scattering of neurons at the edges of the volume that were not included. We intend to correct these errors in subsequent releases.
Soma sub graph meshes: During proofreading, disconnected components can be linked together. Objects which are globally connected, but locally disconnected can be identified, such as when two dendritic branch touch one another, or an axon of a neuron comes in contact with its own dendrite. The soma mesh files we provide as a set of hdf5 files are more accurate meshes of the soma neurons, in the sense that they contain “link_edges” which connect disconnected portions of the mesh, making the graph of the mesh more accurately reflect the topology of the neuron. Second, care was taken to not remove duplicate vertices between locations where a neuron contacts itself. The static meshes one can download from the neuroglancer segmentation via cloud-volume were created from the static segmentation and so are prone to both leave disconnected portions of the mesh apart, and improperly merge together locations of self contact. Depending on the analysis you want to do, this may or may not be a concern.
The most relevant keys of the hdf5 file are as follows: faces, link_edges, vertices.
You can refer to MeshParty as a package to simplify loading, analysis and visualization of these meshes in python.