EM Proofreading Tutorials & Community Practice
How human annotators fix machine segmentation errors, validate synaptic connections, and transform raw AI predictions into ground-truth connectomes.
📑 Table of Contents & Learning Flow
1. Diagnosing Segmentation Errors
Automated 3D convolutional models (FFNs, 3D U-Nets) produce two fundamental topological errors. Your first job as a proofreader is rapidly distinguishing them.
False Merges (Over-Segmentation)
Occurs when the AI connects two distinct biological neurons into a single object ID. Why it happens: thin membrane boundaries, poor staining contrast, or vesicle clouds that blur cell membranes.
Key Diagnostic Signatures:
- Multiple Somas: A single object contains two cell bodies or two primary neurite stalks.
- Impossible Branch Angles: Axons making acute <60° hairpins across fascicles without cytoskeletal continuity.
- Discordant Myelination: A myelinated axon suddenly jumping into an unmyelinated dendritic shaft.
- Membrane Discontinuity: Follow the slice sequence: at least one cross-section will show a clear lipid bilayer separating the two paths.
False Splits (Under-Segmentation)
Occurs when a single continuous biological neuron is severed into two or more detached pieces. Why it happens: knife chatter, staining folds, missing sections, or ultra-thin spine necks.
Key Diagnostic Signatures:
- Orphan Dendritic Spines: A spine head floating unattached within 100nm of a dendritic shaft with matching postsynaptic density.
- Abrupt Terminal Stubs: An axon terminates abruptly in the middle of a continuous bundle with no growth cone morphology.
- Artifact Crossings: An open terminal points directly across a fold or knife chatter mark at another open terminal with identical cross-sectional diameter and mitochondrial trajectory.
2. Synapse Verification Criteria
Automated synapse detectors find millions of connections, but have false positive rates of 5–15%. Use this 4-point checklist to confirm true chemical synapses.
1. Presynaptic Vesicle Cloud
Cluster of clear, spherical 30–50nm lipid vesicles docked within 100nm of the presynaptic active zone membrane. In flies, look for the electron-dense T-bar ribbon structure.
2. Synaptic Cleft Rigid Spacing
Strictly parallel extracellular space (15–25nm width) maintained across at least 3–4 consecutive EM sections, containing electron-dense adhesion protein matrix.
3. Postsynaptic Density (PSD)
Thick, dark proteinaceous thickening lining the intracellular face of the recipient membrane (pronounced in asymmetric/excitatory synapses; thinner in symmetric/inhibitory synapses).
4. Multi-Section Continuity
A genuine synapse spans a disc of 200–500nm diameter. If an apparent junction appears on only a single 30nm section without neighboring vesicles, flag as artifact.
3. Platform-by-Platform Proofreading Workflows
Master the standard tool stacks used across FlyWire, MICrONS, and the Mouse Connectome Project.
🖥️ CAVE & Neuroglancer Workflow
The primary production backend for FlyWire and MICrONS. Uses a dynamic chunked graph (`PyChunkedGraph`) to record edits without re-segmenting the petascale volume.
Essential Keybindings:
Double-Click: Center view on point in 3DShift + Click: Toggle segment ID selectionCtrl / Cmd + Scroll: Step through Z-sectionsAlt + Click: Place split/merge graph annotationSpace: Toggle 3D mesh visibility
Materialization Rule: Never run analysis on live unpinned IDs. Always pin to an explicit materialization_version to guarantee reproducible results.
🌐 webKnossos & Skeletonization
Optimized for high-speed volumetric skeleton tracing (Max Planck / scalable minds). Features "Flight Mode" allowing annotators to fly through axons at up to 1 mm/hour.
Core Features:
- Flight Mode: Velocity-based Z-stepping for rapid long-range axon tracking
- Node & Edge Graphs: Hierarchical SWC / NML skeleton export
- Task Queues: Distributed consensus tracing with automated inter-annotator agreement scoring
Best For: Fast topological proofreading, cell-type census, and dense volume bounding-box annotations.
4. Canonical Community Proofreading Resources
Jump straight into official academies, interactive simulators, and documentation produced by global connectomics consortia.
🏆 FlyWire Academy & Codex →
Official community portal for whole-brain Drosophila connectomics (Nature 2024). Includes comprehensive video tutorials, proofreading certification exercises, and interactive cell search.
Drosophila • Whole Brain • CAVE🎮 EyeWire Training Camp →
The pioneering gamified citizen-science platform developed by the Seung Lab (Princeton). Superb interactive introductory onboarding to 3D branch tracing and merge spotting.
Retina • Citizen Science • Gamified📖 CAVEclient Python Docs →
Official API documentation for programmatically querying proofread root IDs, synapse tables, cell taxonomies, and ID edit lineages in Python.
Python • CAVE API • Provenance🔬 SynapseWeb EM Neurocytology Atlas →
The Kristen Harris lab's canonical visual EM atlas (UT Austin). The standard reference for spine morphologies, active zones, PSD variations, and organelle ultrastructure.
Ultrastructure • Anatomy • Spines🌐 webKnossos User Guide →
Step-by-step guides for setting up volume layers, sharing collaborative annotations, flight-mode tracing, and skeleton mesh exports.
Skeletons • Flight Mode • Tooling🖌️ Harvard Lichtman Lab VAST Guide →
The Volume Annotation and Segmentation Tool created by Daniel Berger. Ideal for high-precision manual voxel painting and proofreading dense mammalian neuropil.
Manual Painting • Mammalian • Harvard