How to actually read an EM image: the organelle catalog with sizes, the minimum criteria for calling a synapse, Gray type I vs II, and a calibrated confidence protocol.
Key community resources for this unit:
| Time | ~2.5 h reading; 75 min studio |
| Prerequisites | Units 01–03. Unit 03’s artifact catalog in particular — you cannot distinguish biology from artifact without it. |
| You need | A public EM volume open in Neuroglancer |
| You finish with | Calibrated compartment and synapse calls with justified confidence tiers, plus a personal cue-reliability ranking |
Everything downstream depends on someone being able to look at a patch of grayscale noise and say correctly what it is. Segmentation networks are trained on those judgments. Proofreading decisions rest on them. Every synapse count in every connectomics paper traces back to a human who decided that a particular smudge was a postsynaptic density.
This unit teaches the actual visual cues, with sizes. Not “use organelle evidence” — which organelles, how big, and what they rule out.
This is the reference table. Sizes are approximate and vary with preparation, but the relative sizes and the presence/absence patterns are what you actually use.
| Structure | Size | Appearance in EM | Found in | Practically absent from |
|---|---|---|---|---|
| Synaptic vesicle, clear round | 35–50 nm | Small circular profiles, clear lumen, clustered | Presynaptic terminals | Dendrites, glia |
| Synaptic vesicle, pleomorphic/flattened | ~35–50 nm | Oval or flattened profiles; shape is partly a fixation artifact but is diagnostically useful | Inhibitory terminals | Excitatory terminals |
| Dense-core vesicle | 80–120 nm | Circular with a dark core | Peptidergic/monoaminergic terminals; also in transit along axons | — |
| Postsynaptic density (PSD) | 30–50 nm thick, 200–800 nm wide | Dark, granular thickening under the postsynaptic membrane | Postsynaptic side | Presynaptic side |
| Synaptic cleft | 20–30 nm (asymmetric); ~15–20 nm (symmetric) | Uniform-width gap with parallel membranes, often with faint cross-bridges | Between synaptic partners | Random appositions have variable-width gaps |
| Microtubule | ~25 nm outer diameter | Tubule in longitudinal section; small ring in cross-section | Dendrites (abundant, in loose parallel arrays); axons (present, more regularly spaced) | Mature spine heads; most glial processes |
| Neurofilament | ~10 nm | Fine filaments, often in bundles | Axons, especially myelinated | Spines |
| Mitochondrion | 0.2–1 µm diameter, variable length | Double membrane with cristae | Everywhere except thin spine necks and the thinnest processes | — |
| Rough ER / polyribosomes | Ribosome ~25 nm | Studded membrane sheets; ribosome rosettes | Soma, proximal dendrites, dendritic shafts | Axons — a workhorse discriminator |
| Golgi apparatus | ~1 µm stack | Stacked flattened cisternae with vesicles | Soma, proximal dendrite | Axons |
| Smooth ER / spine apparatus | Laminae ~30 nm | Tubules; in spines, stacked laminae with dense material between | Dendrites; spine apparatus in a minority of (mostly large) spines | — |
| Multivesicular body | 200–500 nm | Membrane-bound body containing small internal vesicles | Everywhere; enriched in dendrites | — |
| Glycogen granule | 20–30 nm | Very dark small particles, clustered | Astrocytes — near-diagnostic | Neurons |
| Myelin | 10–20 lamellae | Regular concentric dark lamellae | Around myelinated axons | — |
| AIS undercoating | ~20 nm dense layer | Granular density beneath the axolemma, plus fasciculated microtubules | Axon initial segment (~20–60 µm from soma), nodes of Ranvier | Everywhere else |
The two highest-value entries for a beginner are the ones in bold logic: ribosomes rule out axon, and glycogen granules indicate astrocyte. Those two facts alone resolve a large share of early-annotator confusion.
Probable dendrite (or a proximal dendritic branch), medium confidence. The ribosome rosettes are the strongest single cue — axons are effectively free of polyribosomes in standard EM connectomics practice, so their presence argues strongly against axon. The microtubules and mitochondrion are consistent but not discriminating; both compartments have them.
To raise confidence, look for cues that are independent of the ribosome call:
Note the reasoning pattern, which is the transferable skill: do not stack more of the same kind of evidence. Three microtubule observations are one piece of evidence. A ribosome plus a spine plus a taper is three.
A chemical synapse in EM requires all three:
And a fourth practical requirement that experienced annotators treat as non-negotiable:
The single most common beginner error is calling a synapse from dark contrast alone. Dark contrast at a membrane can be: a genuine PSD, a tangentially cut membrane (very common — a membrane sliced obliquely looks thick and dark), staining precipitate, a glial apposition, or a puncta adherens / adherens junction. Criterion 1 is what separates these: no vesicles, no synapse.
Puncta adherentia have symmetric densities on both sides and no vesicle cluster. They look convincing at first glance. The tell is symmetry plus the absence of a vesicle pool.
| Type I (asymmetric) | Type II (symmetric) | |
|---|---|---|
| PSD | Thick, prominent, clearly asymmetric | Thin, roughly equal to the presynaptic density |
| Cleft | Wider, ~20–30 nm | Narrower, ~15–20 nm |
| Vesicles | Round, clear | Pleomorphic / flattened |
| Usual location | Dendritic spines; some shafts | Shafts, soma, AIS |
| Usual inference | Excitatory (glutamatergic) | Inhibitory (GABAergic) |
The inference is a Bin B claim in the Unit 01 sense. It is a well-supported statistical association, not an identity. Known complications: vesicle shape depends on fixation and can be unreliable; some glutamatergic synapses onto interneuron shafts appear less asymmetric; neuromodulatory terminals do not fit the dichotomy at all.
Therefore: write “putatively excitatory (asymmetric)” rather than “excitatory”, and where the claim matters, corroborate with the identity of the presynaptic cell type, which is usually the stronger evidence.
Not a synapse — do not annotate one. Criterion 1 fails (no vesicle cluster) and criterion 4 fails (not persistent).
The most likely explanations are a tangentially sectioned membrane or a non-synaptic apposition. The presence of a mitochondrion is not evidence either way; mitochondria are everywhere.
The correct output is a negative call, and if the region is ambiguous enough to have cost you time, log it as an uncertain patch so it enters the calibration set. Negative calls are data. An annotator who never says “no” is not calibrated.
Most likely the second contact is not a synapse from this bouton — check for a vesicle cluster at that apposition specifically. A single terminal’s vesicle pool can sit near several membranes; only the apposition with an adjacent vesicle cluster and a cleft counts.
If a vesicle cluster genuinely is present at both, you have a multi-synaptic bouton, which is real and common. Note that one terminal making both an asymmetric contact onto a spine and a symmetric contact onto a shaft would be unusual and worth flagging for expert review — it may indicate a merge error that has fused two different axons into one object. This is a good example of ultrastructural reading catching a segmentation error: the biology looks wrong, so suspect the segmentation.
Work in this order. The order matters — cheap, reliable cues first.
Step 1 — Local geometry. Diameter and its variation. Boutons are swellings connected by thin intervaricose segments; dendritic shafts have relatively smooth caliber; spine heads sit on necks.
Step 2 — Organelles present, and just as important, organelles absent. Ribosomes present → not axon. Glycogen granules → astrocyte. Vesicle cluster → presynaptic. Absence is evidence when the structure would be visible if present at this magnification and plane.
Step 3 — Synaptic role. Does the process bear PSDs (receiving) or vesicle clusters (sending)? Many processes do both, but the balance is informative.
Step 4 — Continuity across sections. This is where most single-plane calls get overturned. Scroll. A process that looked like a bouton may be a dendritic varicosity; a “vesicle cluster” may be a tangential slice through something else.
Step 5 — Neighborhood. What is around it? An axon in a myelinated bundle, a process wrapping a capillary, a profile inside a glial sheath — context frequently settles calls that local features cannot.
Step 6 — Assign confidence with a stated evidence chain.
Tiers are useless unless everyone means the same thing. Define them by evidence count and independence, not by feeling:
| Tier | Definition | Example |
|---|---|---|
| High | ≥ 2 independent cues agree, and continuity across ≥ 3 sections confirms | Ribosomes + a spine + taper toward soma → dendrite |
| Medium | 1 strong cue, or ≥ 2 non-independent cues; continuity checked but partially ambiguous | Vesicle cluster present, PSD unclear on the partner |
| Uncertain | Cues conflict, or the decisive cue is not visible in available sections | Process crosses a fold; identity plausible but unverifiable |
“Uncertain” is a valid and valuable output. A dataset in which 8% of calls are flagged uncertain with reasons is more useful than one in which 100% are forced, because the uncertain set is exactly the training and review priority queue. The uncertain rate per region is also the best available proxy for local data difficulty (Unit 03).
Two cues that share a failure mode are one cue.
Genuinely independent cue families:
Rule for high confidence: two cues from different families. This single rule does more for annotation quality than any amount of exhortation to “be careful”, because it is checkable — a reviewer can look at an evidence chain and see whether it draws on one family or two.
Patch: a ~250 nm process in layer 2/3 neuropil, containing a small cluster of round clear vesicles and one mitochondrion, apposed to a bulbous ~600 nm profile.
Family 1 (geometry): the small process swells locally and narrows on either side across z — a bouton on an intervaricose segment. The partner is bulbous with a narrow attachment visible two sections down — a spine head on a neck.
Family 2 (organelles): round clear vesicles clustered at the apposition. No ribosomes in the small process. The spine head contains no microtubules and shows a faint spine apparatus.
Family 3 (synaptic role): thick asymmetric PSD on the spine-head side; cleft of uniform width; visible across four consecutive sections.
Family 4 (context): several other boutons in the neighborhood contact spines similarly — consistent with normal excitatory neuropil, not with an artifact region.
Call: presynaptic axonal bouton making a type I synapse onto a dendritic spine head. Confidence: high — families 1, 2, and 3 agree independently, and continuity is confirmed over four sections.
Inference licensed: putatively excitatory (asymmetric morphology; sign inferred, not observed).
Not licensed: any statement about synaptic strength, or about the identity of the presynaptic cell without tracing the axon to a soma.
Work these panels with the organelle table in §1 open, and name the cue family behind every call you make. They are stills, and single-plane inspection is precisely the habit this unit exists to break — step 4 of the protocol overturns more calls than any other. Treat the panel as a reference for what a cue looks like, and do your actual calling in a volume you can scroll through z.
RIV-ULTRA S04: Neuron structure at the compartment level. Use it to fix vocabulary before you meet anything ambiguous: for each compartment, recall from §1 which organelles you would expect present and, more usefully, which would be absent. Ribosomes ruling out axon is the highest-value entry in that table.
RIV-ULTRA S08: A somatic region — nuclear envelope with heterochromatin above, one long mitochondrion below. Use it to anchor the soma end of the compartment table in §1: rough ER and a nucleus put you in a cell body, and that is the one place where the ribosome cue is unambiguous rather than a judgment call.
RIV-ULTRA S09: The synapse cue set. Hold anything you would call to all three criteria in §2: a vesicle cluster at the apposition itself, a cleft of uniform width, and a density on the receiving side. Dark contrast alone is the commonest beginner error — no vesicles, no synapse.
RIV-ULTRA S10: Vesicles and organelles at annotation scale. Check size against §1 before naming anything — clear synaptic vesicles run 35–50 nm and dense-core vesicles 80–120 nm, so this is a measurement rather than an impression. Remember that vesicle shape is partly a fixation artifact and is not independent of vesicle visibility.
RIV-ULTRA S14: A comparative panel. Use it for the discipline §4 calls the whole game: pick two features that differ between profiles and ask whether they come from different cue families or share a failure mode. Two cues that degrade together under poor staining are one cue.
RIV-ULTRA S20: A two-panel reference — an EM micrograph with a 1 µm scale bar beside a labeled schematic naming presynaptic terminal, presynaptic and postsynaptic membranes, cleft, vesicles, and the postsynaptic dendrite. Read the schematic first, then find each labeled part in the micrograph beside it. That translation — idealized diagram to real noisy tissue — is the step §2's three criteria have to survive.
RIV-ULTRA S24: An advanced case for review. Build a full evidence chain in the form of the §4 worked example — geometry, organelle content, synaptic role, neighborhood — and stop at the point where the chain would need continuity across sections that a single still cannot supply.
Attribution: Pat Rivlin training materials (MICrONS proofreading deck).
Scenario. Your team is preparing a training-ready annotation subset for segmentation QC. It deliberately contains borderline cases.
Outputs: consensus annotation sheet; disagreement log with counts by type; one rubric revision with rationale.
Why step 5 matters. Vocabulary mismatch is usually the largest category on a first run, and it is entirely fixable by better protocol wording. Teams that run this loop two or three times typically see inter-annotator agreement rise substantially without anyone becoming a better microscopist — the gain comes from the protocol, not the eye. That is the scalability lesson of this unit.
| Not yet | Proficient | Strong | |
|---|---|---|---|
| Evidence quality | Single-cue calls presented as definitive | Two cues per call | Two cues from different families, with independence argued |
| Synapse criteria | Calls from contrast alone | Applies all three criteria | Applies all three plus persistence; correctly rejects adherens junctions and tangential membranes |
| Confidence | Missing or inconsistent | Tiers applied consistently | Tier justified against the operational definitions; uncertain rate is reasonable, neither zero nor excessive |
| Inference discipline | “Excitatory synapse” | “Putatively excitatory (asymmetric)” | Names the assumption and proposes corroboration via presynaptic cell type |
| Error analysis | Counts errors | Classifies disagreements by type | Converts the dominant disagreement type into a concrete protocol revision |
Single-slice overconfidence. Recover: make scrolling a mandatory step in the protocol, not a suggestion. Add a checkbox to the annotation sheet.
Stacking dependent cues. Recover: label each cue with its family. Three cues from one family is one cue.
Forcing labels. Recover: define and reward the uncertain tier; track the uncertain rate per annotator as a calibration statistic, not a performance penalty.
Label drift along a long trace. Recover: build in periodic re-checks — every N micrometers of tracing, re-verify the compartment call from scratch rather than carrying the earlier decision forward.
Reading artifact as biology. Recover: keep the Unit 03 artifact catalog open. When something is anomalous, ask whether its shape follows tissue or follows the section/tile/scan geometry.
Some of what this unit teaches is technique. Some of it is professional norm — the things experienced people do without being asked, and which nobody states out loud because they assume you already know. Those are worth naming, because they are distributed unequally by background rather than by ability.
From this unit:
“Uncertain” is a real answer and should be rewarded out loud. An annotator who never says “no” is not calibrated. Learners suppress uncertainty because they read it as failure; a facilitator has to say the opposite explicitly, once, early.
Two cues from different families, or it is not high confidence. Three observations that share a failure mode are one observation. Nobody states this, and it is the single most common reasoning error in annotation.
The collected set, and why making these explicit is a fairness intervention rather than etiquette, is in the hidden curriculum.
Systematic axon-vs-dendrite classification, which gets its own treatment in Unit 06; glial identification in Unit 07; and how these calls feed proofreading triage in Unit 08.
This unit is anchored to canonical papers and datasets used in connectomics practice. Use these as required preparation before activities.
Capability target: Interpret ultrastructural features reproducibly to distinguish compartments and synaptic context.
Ultrastructure Consensus Round - Build consistency in compartment and synapse labeling. The unit's own lab above is the graded version of this exercise; do that one.
Use compartment, organelle, and synaptic cues to make reproducible interpretation decisions.