Semiautomatic neuron reconstruction (proofreading)

The Knossos-based proofreading tool makes available a live proofreading environment in which edits contributed by different users are immediately available to all users. It interacts with a custom backend server, which manages the agglomeration graph. Knossos can (1) request a list of objects that are part of the same agglomeration graph connected component, (2) add a connection, (3) remove a connection between two objects in the agglomeration graph and (4) request the surface mesh representations for a set of objects. For efficiency, the backend server holds the agglomeration graph in random access memory. Any edit to the agglomeration graph is directly performed in memory, but also written to an on-disk append-only edit log. The edit log is replayed on top of the original automatically generated proposal agglomeration every time the server is reloaded. This design allows highly efficient operations: in our usage so far, edge insertions required 4, 20 and 42 ms (first percentile, median, 99th percentile, respectively, with 13,461 insertions), edge deletions required 1, 18 and 39 ms (1,875 deletions) and obtaining the connected component to which a supervoxel belonged by depth-first search required 0, 12 and 111 ms (40,311 requests).

We performed a stress-test of the backend server by replaying real user activity from log files with increasing parallelization, from a single client machine connected to the backend server by a 1 gigabit ethernet connection (Extended Data Fig. ​Fig.1010).

a. Response time to request of meshes for cells (either complete or currently undergoing proofreading) and (b) to the request of agglomeration graph connected component containing a given supervoxel, depending on the number of concurrently active simulated users. Box plot center lines represent medians, box limits upper and lower quartiles, whiskers 1.5x the interquartile range.

To identify all synaptic partners of SINs in the tectum, we first manually annotated all incoming and outgoing synapses along the SIN neurites. Starting from these synapses, we then traced the partner cells, until the cell body or, for RGCs, the axon in the optic tract were unambiguously identified. In case of tectal cells, presence of an axon leaving the tectum was used as the criterion to distinguish PVPNs from PVINs.