Semantic Segmentation TASK-01
Labeling tissue types and cellular structures at the semantic level within EM volumes.
Task Overview
What is Semantic Segmentation?
Semantic segmentation involves classifying each voxel in an EM volume according to the biological structure it belongs to. Unlike instance segmentation (which identifies individual neurons), semantic segmentation labels broad categories like "cell body," "axon," "dendrite," "mitochondria," "myelin," and "extracellular space."
Why is it Important?
Semantic labels provide critical context for downstream analysis. They help distinguish neural tissue from non-neural structures, identify subcellular compartments, and enable automated quality checks. Accurate semantic segmentation improves the efficiency of proofreading and annotation stages.
Key Objectives
- Identify and label major tissue compartments in EM imagery
- Distinguish neural from non-neural structures
- Mark subcellular organelles when required by protocol
- Flag ambiguous or damaged regions for review
Current Standard Operating Procedure Authoritative
Scope
This SOP applies to all semantic segmentation tasks in the FlyWire and BANC datasets. Follow these steps exactly unless instructed otherwise by a supervisor.
Procedure
- Open the assigned volume in WebKnossos using the provided task link.
- Review the task region by scrolling through all Z-slices to understand the structures present.
-
Select the appropriate label from the semantic class palette:
- Soma — Cell bodies with visible nucleus
- Axon — Thin processes with vesicle-filled terminals
- Dendrite — Larger processes with spines or PSDs
- Glia — Non-neuronal cells (astrocytes, microglia)
- Myelin — Dark-staining membrane wrapping
- Mitochondria — Subcellular organelles with cristae
- Extracellular — Space between cells
- Artifact — Damaged or ambiguous regions
- Paint each structure using the brush tool. Use appropriate brush size for the feature scale.
- Review boundaries by toggling the segmentation overlay. Ensure clean edges without gaps or overlaps.
- Submit for review when complete. Add notes for any ambiguous regions.
Quality Criteria
- No unlabeled voxels within the task region
- Boundaries follow actual membrane contours
- Consistent labeling across Z-slices
- Artifacts properly marked rather than guessed
Source Document
Historical SOPs Archived
Previous versions of the semantic segmentation protocol are preserved here for reference. These may contain outdated procedures—always use the Current SOP above.
Version History
-
v2.0 (Current) — January 2026
Added artifact class, updated WebKnossos workflow -
v1.1 — September 2025
Added myelin and glia classes -
v1.0 — June 2025
Initial protocol with basic neural/non-neural distinction
Archived Documents
Visual Examples
Reference images showing correct labeling for each semantic class. Click to view full-resolution examples in the gallery.
Quick Reference
- Soma — Large, round profile with nucleus visible
- Axon — Thin, uniform diameter, may have myelin
- Dendrite — Thicker, tapered, with spines
- Myelin — Dark concentric rings around axon
- Mitochondria — Oval with internal cristae structure
Common Failure Modes
- Confusing axons with dendrites Axons are thinner and more uniform. Dendrites taper and have spines. Check for vesicle clouds (axon) vs PSDs (dendrite) at synapses.
- Missing myelin wrapping Myelin appears as dark concentric layers. Don't label myelinated regions as pure "axon"—include the myelin label.
- Inconsistent labeling across Z-slices A structure should maintain its label through the volume. Use 3D view to verify continuity.
- Guessing in damaged regions If tissue is torn, folded, or poorly stained, use the "Artifact" label. Don't guess the underlying structure.
- Boundary leakage Labels should stop at membrane boundaries. Use smaller brush sizes near edges and toggle overlay to check.
Tools Used
WebKnossos
Primary annotation platform for semantic painting
Neuroglancer
3D visualization for context and verification
CAVE
Backend database for storing annotations
Keyboard Shortcuts (WebKnossos)
- 1-9 — Select label class
- B — Brush tool
- E — Eraser tool
- Shift+Scroll — Adjust brush size
- . — Toggle segmentation overlay
Training Videos
Additional Videos
- Synapse Identification Walkthrough (Local file)
- Top-down View Example (Local file)
Related Publications
-
Automated Semantic Segmentation of Electron Microscopy Images