search

reconstrue / brightfield

Reconstruction of biocytin-stained neurons detected in brightfield microscopy image stacks
http://reconstrue.com
Apache License 2.0
2 stars 1 forks source link

Pull down training data from brain-map.org #67

Open JohnTigue opened 4 years ago

JohnTigue commented 4 years ago

See also:

Currently The Allen Institute images more cells than it reconstructs. Yet as is on brain-map.org there are hundreds of cells with semi-automatically generated skeletons that have passed human QC review. That's tons more labeled training data.

Also, this and #60 are all the more evidence that adding a mechanism to allensdk that pulls down image stacks (via their RMA interface, but wrapped in friendly Python) is a valuable PR to send back upstream.

Bonus, how to get brightfield stacks into NWB:N 2.0, or at least CZI's JSON.

JohnTigue commented 4 years ago

The cells in the challenge dataset are simply pre-released brain-map.org cells. The Allen Institute is constantly adding more cells.

This project has taken the raw challenge dataset and built specimens_manifest.json as a stand-in for the AllenSDK cell cache manifest JSON. But just as easily, data can be pulled down from brain-map.org or a Wasabi-hosted prerelease. Once the cells have been release, it is only natural to access them through brain-map.org.

So, tons of training data for researchers. And a challenge can be issued at any time by simply selecting ten cells from the brain-map.org data that are without public SWCs.