Semantic Segmentation (Labeling)

TASK-02 · Task Guide · Painting & labeling structures

What is Semantic Segmentation?

Semantic segmentation is the process of labeling tissue types and cellular structures in electron microscopy data. This encompasses two related activities: voxel painting (hand-painting specific structures in 3D) and classification labeling (identifying and naming structures like axons, dendrites, and somata).

We have a storied history of hand-painting neurons, mitochondria, vesicles, synapses, T-bars, errors, membranes, and more. The tools and software have changed rapidly over the years as technology capabilities evolve, but the core skill of identifying and labeling biological structures in EM imagery remains the same.

Structure Labels

Soma — Cell body
Axon — Output process
Dendrite — Input process
Glia — Support cells
Blood Vessel — Vasculature
Myelin — Insulating sheath (use when inner/outer is unclear)
Myelin (inner tongue) — Innermost wrap, next to the axon
Myelin (outer tongue) — Outermost wrap, glial side
Synapse / T-bar — Connection
Vesicles — Transport bodies
Extracellular — Outside cells
Artifact / Defect — Imaging error

Note that the distinction between painting and labeling can be confusing for newer tracers. Painting means coloring in the actual voxels of a structure. Labeling means assigning a classification name (like "axon" or "dendrite") to an already-segmented object.

Defect Annotation

Defect annotation is a specialized form of semantic segmentation focused on marking imaging artifacts, tissue damage, and data quality issues. Defects include section tears, folds, staining artifacts, missing slices, and other preparation problems. Accurate defect annotation prevents these regions from corrupting downstream analysis. See the Gallery for visual examples of common defects.

Crack — Fracture in the section
Fold — Tissue folded over itself
Tear — Torn region of the section
Fat Globule — Lipid droplet artifact
Membrane Swirl — Swirled membrane whorl

Procedure

  1. Load the assigned volume or task in the annotation tool
  2. Orient yourself — identify major landmarks and the region of interest
  3. Select the appropriate label/brush for the structure you're painting
  4. Paint or label structures systematically, working through Z-slices
  5. Cross-check your labels in orthogonal views (XY, XZ, YZ) for consistency
  6. Submit for review when complete

Common Failure Modes

  • Confusing axons with dendrites in ambiguous cross-sections
  • Missing myelin wrapping around processes
  • Inconsistent labeling across Z-slices (label drift)
  • Guessing in damaged or low-contrast regions instead of flagging
  • Boundary leakage when painting — extending labels past structure edges

Tools Used

WebKnossos Neuroglancer CAVE VAST (legacy) Omni (legacy)

Purpose

Explain semantic segmentation — painting and labeling tissue types and cellular structures in EM data — including voxel painting, classification labeling, and defect annotation.

Scope

Applies to tracers painting or labeling structures across our datasets. The current voxel-painting method has its own step-by-step procedure in SOP-006: Voxel Painting Cell Segmentation; this guide is the broader task overview.

Original documentation

The source material this guide distills — original team docs kept here for reference.

Related