Side Quest: Neuroanatomy for Proofreaders

The EM identification training behind competent proofreading: compartment cues, confidence tiers, a unified axon-dendrite-glia decision sequence, and a self-run calibration drill on a real public volume.

Why this exists

The proofreading side quest states its hard prerequisite in one sentence: you can tell an axon from a dendrite from a glial process in EM, with a confidence tier attached. Until now the only route to that sentence was “do Units 05–07 in full” — three units built for a taught track, with tutorials, studio activities, and run-of-shows a self-studier does not need.

This side quest is the direct route. It pulls the identification material out of the units and the content library into one sequence, adds the one thing no existing page carries — a single decision sequence that covers axon, dendrite, and glia in the same pass — and ends in a calibration drill you run yourself on a real public volume.

One warning before the reading list, because it changes how you should spend your hours. Identification is a perceptual skill. The hidden curriculum puts it bluntly: you cannot close a perceptual gap by reading, any more than you can learn to hear an interval by reading about intervals. The reading below exists to give you a cue vocabulary and a decision procedure. The skill comes from the scored judgments in stages 5 and 6, and if you have to cut hours, cut reading, never judgments.

Before you start

Time

15–20 hours, workable in evenings over two or three weeks.

Hard prerequisite

You have looked at EM of brain tissue and know what an image stack is. Unit 03 covers this; you do not need all of it, but you should recognize an artifact when one crosses your screen.

Soft prerequisite

Basic Neuroglancer navigation, for the drill in stage 6. Half an hour with the dataset access guide covers it.

Not required

Python, a lab affiliation, any track, or the proofreading side quest — this quest is the on-ramp to that one, not the other way around.

Ends with

A scored confusion matrix with accuracy reported by confidence tier, and a one-page personal cue card in your own words.

The sequence

1. Learn the vocabulary: organelles and synapses 4 hours

   Read Unit 05 §1–2 — the organelle catalog and the three criteria for calling a synapse — then go one level deeper with Organelle annotation cues and Synapse classification.

   The organelle catalog is the alphabet everything else is written in. Two entries repay special attention because they carry the most identification weight later: ribosomes and rough ER, which the library rates as the single most reliable negative marker for axonal identity, and synaptic vesicles, whose presence in a cluster at a membrane is one of the three minimum criteria for a synapse — the other two being a parallel membrane apposition and a postsynaptic density, with persistence across sections as the check that keeps single-plane wishful thinking honest.

   Skim Soma ultrastructure and Myelin and nodes of Ranvier rather than reading them in full: what you need now is the neuron-vs-glia soma comparison table in the first, and §7 of the second — myelin as a strong axon identifier, and why nodes of Ranvier are where segmentations of myelinated axons break.

   You finish with: the ability to name, for any organelle you can see in a patch, which compartments it argues for and which it argues against — not “there’s a mitochondrion” but “elongated mitochondrion, consistent with dendrite or axon, tells me little on its own.”

   Check yourself before moving on:

   A process contains a tight cluster of ~40 nm round vesicles. What can you conclude, and what two things do you check before calling the compartment an axon terminal?

   You can conclude that a presynaptic identity is plausible — clustered small round vesicles are the strongest single presynaptic cue. Before making the call: check for a membrane apposition with a postsynaptic density on the partner side (vesicles without an active zone can be a passing vesicle pool, not a terminal), and scroll adjacent sections to confirm the cluster and the contact persist. A vesicle cluster in one plane that vanishes in the next is a tangential slice through something else — Unit 05 treats persistence across sections as the criterion that rescues you from single-plane errors.

2. Learn the tier discipline before the cues that need it 2 hours

   Read Unit 05 §3–4: the six-step compartment decision protocol, the confidence tiers defined operationally, and the five independent cue families.

   This is the stage people skip because it looks procedural rather than anatomical, and skipping it is how you end up fast and wrong. Two rules from this material govern everything that follows:

   • High confidence requires two cues from different families. Two cues that share a failure mode are one cue — microtubule count and cytoplasmic density both degrade together under poor staining, so agreeing with each other proves nothing.
   • “Uncertain” is a valid and valuable output. The tier definitions are evidence counts, not feelings: high means at least two independent cues agree and continuity across three or more sections confirms. If you cannot state the evidence chain, you do not have the tier.

   You finish with: the three tiers memorized as operational definitions, and the habit of writing a call as call + tier + evidence chain — “dendrite, high: ribosomes (organelle family) + spine with PSD (synaptic family), continuous across five sections.”

3. Axon versus dendrite, properly 4 hours

   Read Unit 06 — the four cue families, the exceptions that break the polarity rule, and the five-step local classification protocol — then the library’s Axon-dendrite classification for the four-pass version with reliability ratings per cue, and Axon biology and Dendrite biology for the underlying biology, each of which ends in a worked discrimination example.

   Do not memorize twenty cues with equal weight; learn the reliability ordering. Ribosomes rank highest. Caliber ranks low — the misconception tables in both library entries call out “axons are always thinner than dendrites” as false, and the unit’s rule is that caliber alone never supports a high-confidence call. And learn the four exceptions to the polarity rule (dendro-dendritic synapses, axo-axonic contacts onto the AIS, presynaptic dendrites in the retina, unipolar invertebrate neurons) well enough that meeting one produces “exception, check context” rather than panic.

   You finish with: Unit 06’s five-step decision tree reproducible from memory, with a confidence tier at each exit, and the reliability ordering of cues — which cue you trust when two disagree.

   Check yourself before moving on:

   A process in cortex shows scattered ribosomes along its length, and also a vesicle cluster with an active zone facing a spine. What is the leading hypothesis, and why is it not "an exception to the polarity rule"?

   The leading hypothesis is a merge error — one segmentation object stitched from an axon and a dendrite. Ribosomes are the strongest dendritic cue; a presynaptic active zone is the strongest axonal cue; both are high-reliability cues from different families. Unit 06’s rule: when two high-reliability cues from different families contradict each other, the leading hypothesis is not “one cue is wrong” but “this is not one object.” The known exceptions (presynaptic dendrites in the retina, for example) are region- and cell-type-specific, so in generic cortex the segmentation error is the more probable explanation — and this is exactly the reasoning stage 7 turns into a proofreading move.

4. Glia: the confusion that costs the most 3 hours

   Read Unit 07 — why a glia merge is expensive, the three classes with one near-diagnostic feature each, and the six-step identification protocol — then the library’s Glia recognition for the per-class cue detail and the two worked edge cases (astrocyte vs. thin dendrite, OPC vs. small neuron).

   The discrimination that matters most is astrocyte process versus thin neurite, and the unit’s table for it deserves to be learned outright: glycogen granules, absence of synaptic participation, absence of organized microtubules, and — the cue that survives weak staining better than glycogen does — cross-sectional shape, space-filling and angular where neurites are tubes. Learn the error asymmetry too: calling an astrocyte a neurite feeds merge errors that inject false connectivity; calling a neurite an astrocyte produces a visible, locally fixable split. The two mistakes are not equally priced, and the drill in stage 6 scores them separately.

   You finish with: one diagnostic feature per glial class recallable on demand, the astrocyte-vs-thin-neurite table internalized, and the ability to say which direction of glial misidentification costs more and why.

5. Build the unified checklist 1–2 hours

   No existing page carries the three-way discrimination in one pass — Unit 06’s tree covers axon vs. dendrite, Unit 07’s protocol covers glia and hands off to Unit 06 at step 2, and holding two fenced protocols in your head mid-drill does not work. So the merge is this quest’s contribution, and then, deliberately, yours.

   Start from this synthesis of the two protocols and the library’s tables:

   0. ARTIFACT CHECK   Is this tissue? Folds, tears, staining dropout, myelin-
                       mimicking artifacts -> flag the location, not the biology.
   1. NUCLEUS VISIBLE  Use the soma tables: chromatin pattern, nucleolus, Nissl.
                       Neuron / astrocyte / oligodendrocyte / microglia from the
                       nucleus alone gets you most of the way. (Unit 07 §3, soma
                       ultrastructure §9)
   2. NEURITE OR GLIA  Glycogen granules, synaptic participation, organized
                       microtubules, cross-sectional shape: space-filling and
                       angular -> glial; tube -> neurite. (Unit 07 §2)
   3. WHICH GLIA       One diagnostic feature per class: glycogen (astrocyte),
                       darkest nucleus in the field (oligodendrocyte), dense
                       bean-shaped nucleus + lysosomes (microglia). (Unit 07 §2)
   4. AXON OR DENDRITE Ribosomes/RER -> dendrite. Vesicle cluster + active zone
                       on the sending side -> axon. Then geometry, then context.
                       Caliber alone never upgrades confidence. (Unit 06 §1, §3)
   5. CUE CONFLICT     Two high-reliability cues from different families that
                       disagree -> suspect a segmentation error, not exotic
                       biology. Record it; stage 7 is about this. (Unit 06)
   6. ASSIGN TIER      High = 2 independent families + continuity across >= 3
                       sections. Uncertain is an output, not a failure. (Unit 05)
   

   Then rewrite it in your own words, at most one page, ordered the way you actually check. The rewriting is not busywork: the facilitator guide’s observation is that learners who can describe cues and learners who can apply them are different populations, and compressing the protocol into your own words is the cheapest way to find out which you currently are.

   You finish with: a one-page cue card, in your own words, that you will use — and revise — during the drill. Revision marks on the card afterward are evidence of learning, not of a bad first draft.

6. Run the calibration drill on a real volume 3 hours

   The units specify scored drills on curated patch sets, and this site does not host one — the side quest index has said so honestly since the gap was named. What you can run today, alone, is the self-run version: a public volume with a released segmentation, where the segmentation plays the role of the answer key and of the thing you are learning to distrust.

   The protocol:

   1. Open a public EM volume with a released segmentation in Neuroglancer — the dataset access guide routes you to MICrONS and FlyWire. Write down the dataset and segmentation version.
   2. Build a set of 20 locations before you start judging: drop annotation points semi-randomly through neuropil, resisting the pull toward clean-looking spots. Add a few deliberately hard ones — a myelinated bundle, a perivascular region, dense neuropil.
   3. For each location, hide the segmentation layer first. At native EM resolution, make the call from your cue card: compartment, confidence tier, and the evidence chain (which cues, which families). Scroll adjacent sections — the units are emphatic that single-plane calls are a habit to break, not a shortcut to allow.
   4. Reveal the segmentation and adjudicate. Agreement is provisional truth. Disagreement goes to one of two bins: my error (the segmentation, plus scrolling, shows a cue you missed or misweighted) or candidate segmentation error — your anatomy call stands and the released object looks wrong. Do not force either bin; an unresolvable case is recorded as uncertain.
   5. Score into a confusion matrix: axon / dendrite / glia / uncertain against adjudicated truth. Compute overall accuracy, accuracy within your high-confidence tier, and your uncertain rate.

   Reading the score. Unit 06’s standard: high-confidence calls right at least 90% of the time, with a non-trivial share of calls left uncertain. If your high-tier accuracy matches your overall accuracy, your tiers are decoration and the fix is tier discipline, not more anatomy. Interpret each off-diagonal cell the way Unit 07’s drill does — “astrocyte called dendrite” and “dendrite called astrocyte” are different failures with different costs — and let the worst cell pick which stage you reread before the second round.

   Then make it a habit rather than an event: the hidden curriculum prescribes exactly this as a weekly mixed set, scored by tier, for as long as you are doing identification work at all. And if you want the loop closed by someone other than yourself, the community platforms under Resources beyond this site check your calls and gate access on them — the external review this drill cannot provide.

   You finish with: a confusion matrix with accuracy by tier, an uncertain rate, a dataset and segmentation version written down, and — if the volume cooperated — one or two candidate segmentation errors you found by anatomy alone.

7. Turn cues into proofreading calls 2 hours

   Read Error taxonomy §2–3 for the visual signatures of merges and splits, then Scenarios 2 and 5 of the proofreading worked examples, watching specifically for where the resolution turns on anatomy: caliber and organelle-content match before committing a merge operation, smooth-ER continuity as evidence that two fragments are one process, an excitatory-only branch beside an inhibitory-only branch flagging a merge.

   Everything you trained in stages 1–6 lands here as a short mapping, which is the actual point of this quest:

   You observe	Suspect	The proofreading move
   Ribosomes and a presynaptic active zone in the same object	Merge joining an axon and a dendrite	Walk the object to the join point; split there
   A ~500 nm dendrite becoming a ~100 nm axon at a branch point	Merge at a false branch	Examine the branch point across sections
   A process dead-ending with no terminal bouton, no vesicle cluster	Split — the rest of the cell is another fragment	Search past the break, guided by trajectory and caliber
   Glycogen, no synapses, space-filling shape inside a “neuron”	Glia–neuron merge	Split at the boundary; these seed false inputs
   A myelinated axon “disappearing” between sections	Node of Ranvier, not necessarily an error	Check ~20–30% caliber narrowing before editing anything
   Impossible ~180° hairpin branching in 3D	Merge	View the mesh in 3D; find the join

   The deeper habit under the table: cue conflict is a segmentation-error detector. An annotator sees a weird object and doubts their anatomy; a proofreader sees the same object and doubts the segmentation. Knowing which doubt to reach for — and holding it with a stated confidence tier rather than certainty in either direction — is the competence this quest exists to build.

   You finish with: the prerequisite sentence of the proofreading side quest true of you, with evidence — which is your cue to start that quest.

The artifact

Three things, all of which stages 5 and 6 already produced in draft:

1. The cue card. One page, your own words, revised at least once during the drill.
2. The calibration record. The confusion matrix; overall accuracy, accuracy within your high tier, and uncertain rate; the dataset and segmentation version; and a one-line interpretation of your worst off-diagonal cell.
3. Five evidence chains. Written out in full — call, tier, cues, families, continuity — including at least one uncertain call and, if you found one, the candidate segmentation error with your reasoning.

This artifact is smaller than the proofreading quest’s release memo, and that is deliberate: it is the entry evidence a lab would want before letting you near a correction queue. Calibration — a stated confidence that tracks actual accuracy — is worth more to them than raw accuracy, because a calibrated annotator’s uncertain flags are a usable review queue and an overconfident annotator’s errors are invisible.

What “done” looks like

• You can produce the three-way call — axon, dendrite, glia — with a tier and an evidence chain, without consulting the card for routine cases.
• Your high-confidence accuracy is measurably better than your overall accuracy, and you can say by how much.
• You have written “uncertain” on real cases and can defend why that was the right output, not a failure.
• You have looked at a released segmentation and correctly doubted it at least once — or can explain why your volume gave you no occasion to.
• The exceptions to the polarity rule produce “check the context” rather than a forced call.

Common detours

• Reading all six library entries end-to-end before judging anything. The entries are references, not a novel. The over-reading failure mode is named in the hidden curriculum — reading is private and cannot be judged, which is precisely its appeal and its danger. Get to stage 6 with hours still in the budget.
• Single-plane calls. The units repeat it because everyone does it: scroll. Every cue table you have read assumes continuity checking, and most wrong calls in practice are right-looking calls made in one plane.
• Treating caliber as a cue family. It is one weak cue inside geometry, and the misconception every beginner imports from textbook diagrams.
• Forcing every drill disagreement into “my error.” Released segmentations contain errors; that is the entire premise of proofreading. If your evidence chain is sound and the object still looks wrong, “candidate segmentation error” is the honest bin — and the more valuable finding.
• Scoring overall accuracy only. The number that matters is accuracy within your high tier. Overall accuracy improves with caution; calibration only improves with honesty.

What this side quest does not cover

• A hosted, scored drill corpus. The self-run protocol in stage 6 is the workable substitute, but it is not the curated set with planted errors that Units 05–07 specify, and its answer key — a released segmentation — is itself imperfect, which the protocol turns into a feature but cannot fully escape. A hosted corpus with known ground truth remains an open gap, and the side quest index will say so until it exists.
• Cell typing. Telling an axon from a dendrite is not telling a basket cell from a chandelier cell; that is Neuron type identification and Unit 09’s territory.
• Synapse partner assignment at scale. Stage 1 teaches you to call one synapse properly; production synapse proofreading has its own error modes, covered in Unit 08.
• The proofreading operation itself. Budgets, triage, metrics, stopping rules, and the release decision are the proofreading side quest, which this quest exists to make honest for you to start.
• Anything that certifies you. Same honest limit as every self-study path here: the artifact is the thing you show, and one external review of it is worth more than three self-reviews.

Resources beyond this site

Everything above links inward on purpose — the sequence is the point. But the wider field has resources this site cannot replicate, and four of them fit this quest directly. All were checked as live and as described in August 2026.

An atlas to keep open during stages 1–4. SynapseWeb, from Kristen Harris’s lab at UT Austin, is the closest thing the field has to a canonical ultrastructure reference: free tutorials on axons, dendrites, spines, and astrocytes built from real serial-EM reconstructions, plus Josef Spacek’s Atlas of Ultrastructural Neurocytology — annotated micrographs covering the neuron, the glia, blood vessels, and, usefully for stage 0 of your checklist, abnormalities. When a cue table here names a structure you cannot picture, this is where to go picture it. The book-length treatment behind much of this material is Peters, Palay, and Webster, The Fine Structure of the Nervous System: Neurons and Their Supporting Cells (3rd ed., Oxford University Press, 1991) — out of print, so a library copy.

Practice with feedback, which stage 6 cannot give you. The self-run drill’s honest limit is that you grade yourself. Two open community platforms close that loop by checking your work and gating access on demonstrated competence — the closest existing thing to the hosted, scored drill this site lists as a gap:

• EyeWire II — the Seung Lab’s community platform for proofreading ~100,000 neurons of mouse retina. You learn the tools, pass a challenge, and then map real assigned cells whose results the community verifies.
• FlyWire — the fly connectome’s community, still active on new datasets. Signup grants a sandbox copy of the segmentation to practice edits in consequence-free, and production proofreading is gated by an entry test; the self-guided training includes a glossary of fly ultrastructure. One caveat this quest equips you to state: fly neurites are the polarity-rule exception from stage 3, so the vertebrate axon-vs-dendrite cues transfer only partly.
• connectome.quest — the Seung Lab’s hub that routes to both of the above (and more), with step-by-step guides and video tutorials, all on the same Neuroglancer-family viewer the drill uses.

Tooling for the drill. The MICrONS Explorer tutorials walk the cubic-millimeter dataset’s viewers step by step, and WEBKNOSSOS offers a free browser-based annotation environment with published EM datasets — useful if you want to practice volume annotation itself, with no gate, before touching a released segmentation.

The paper behind the community model. Dorkenwald et al., “FlyWire: online community for whole-brain connectomics,” Nature Methods, 2022, 10.1038/s41592-021-01330-0 — how sandbox training, entry tests, and distributed proofreading actually operate at scale. Read it when you want to know what a production pipeline will expect of you, which is also the best possible framing for the quest you do next.

(For anyone teaching this material rather than learning it: FlyWire’s Codex Academy has classroom worksheets and videos for high school and college outreach — adjacent to this quest, not part of it.)

Where this connects

---

All side quests · The core · Tracks