SEGConnectomics

Brain segmentation pipeline

Assign neuron identities throughout a three-dimensional image of brain tissue. Measure how long a reconstruction stays correct and how quickly the pipeline processes the volume.

The task

How far can a reconstruction follow one neuron?

Input
A three-dimensional microscopy volume with a defined region and voxel size.
Output
An aligned volume of instance labels, with a consistent identity for each reconstructed neuron across sections.
Tissue cross-sectionTask schematic
11 / 21
Move through the schematic volume. A color follows the same instance across slices. Switch to contours to inspect the boundaries.

End-to-end throughput

Divide the number of input voxels by elapsed wall time, from reading the volume to writing the final labels. Include preprocessing, inference, stitching and postprocessing.

Expected run length

ERL measures the expected length of an error-free path along a proofread reference skeleton. Splits interrupt paths and merges invalidate affected runs. Report ERL in micrometres.

Hardware and reference

Compare runs on the same volume and hardware profile. Record accelerator type and count, memory and storage conditions. Report throughput and ERL together so a faster run can be assessed for continuity.

Resources

Tissue volumes

Image volumes, voxel geometry and reference annotations.

Evaluation code

End-to-end timing and reconstruction continuity scoring.