Neuron Segmentation

Dense neuron segmentation in EM is an instance segmentation task. The canonical pipeline first predicts an affinity map (the connectivity of each voxel to its neighbors) with an encoder-decoder, then converts the affinity map into a segmentation via watershed or a similar algorithm.

This section covers three benchmarks:

• SNEMI3D — the classic small isotropic-anisotropic benchmark, used for end-to-end affinity training and waterz post-processing. Evaluated with Rand Index and Variation of Information.

• NISB — a larger, anisotropic neuron-segmentation benchmark evaluated with the NERL skeleton metric. Reproduction targets in tutorials/neuron_nisb/ mirror the upstream BANIS pipeline.

• LICONN — the LICONN volume variant of the NISB benchmark; reuses the BANIS-style affinity pipeline and adds an affinity-mask QC step for the LICONN-specific border artifacts.

• SNEMI3D
• NISB
• LICONN