Proofreading · Lobeworks/17

Proofreading, in connectomics, is the correction by people of an automatic segmentation, in which proofreaders find the places where the machine joined two neurons into one or cut one into pieces, and repair them; it is where most of the human effort behind a connectome goes.

Proofreading. Proofreading, in connectomics, is the correction by people of an automatic segmentation, in which proofreaders find the places where the machine joined two neurons into one or cut one into pieces, and repair them; it is where most of the human effort behind a connectome goes.

It took about 33 person-years for the FlyWire brain, over 50 for the hemibrain and an estimated 44 for the male nervous system. A proofreader does not inspect voxels. In a viewer that shows the grey sections beside the three-dimensional shape of each segmented cell (Neuroglancer, written by Google, runs in a web browser), they look for anatomy that cannot exist: a neuron with two cell bodies is almost always a merge, a branch that stops dead in the middle of tissue is almost always a split, and a flat flange or a piece that jumps across the brain is suspect. Then they drop into the sections at that spot and cut or join.

Merges matter most. A merge credits one neuron's synapses to another and wires the graph wrong in a way that can look plausible, which is why segmentation methods such as flood-filling networks trade computation to avoid them.

The edits are versioned. CAVE, the system behind FlyWire, keeps the segmentation as a graph of small fragments, so an edit adds or removes an edge, and any analysis can be rerun on the brain as it stood on a given day.

Many hands did it. FlyWire was proofread first by fly laboratories around the world and later by teams in Princeton and Cambridge, with citizen scientists contributing; older projects traced neurons by hand as skeletons in a tool called CATMAID.

Questions: How does a proofreader find a merge in three dimensions? By looking for anatomy that cannot exist. A segmented neuron with two cell bodies is almost always two neurons joined, a branch that stops in the middle of tissue is almost always a split, and a flat flange or a piece that jumps across the brain is suspect. The proofreader then goes to the electron micrographs at that spot, finds where two strands were confused, and cuts or joins. Why does connectome proofreading need version control? Because many people edit the same brain while others analyse it. CAVE stores the segmentation as a graph of small fragments, so a merge or a split is the addition or removal of an edge, and keeps the history so that any analysis can be rerun on the brain as it stood on a given day. Without that, a result could change every time someone repaired a neuron.