Recognizing astrocytes, oligodendrocytes, and microglia in EM, why glia-neuron boundary errors corrupt neuronal connectivity, and a drill-based protocol for getting them right.
Key community resources for this unit:
| Time | Self-study ~2.5 h: about 90 min of reading plus the 60 min discrimination drill. Taught: a 90 min session, per the lecture plan. Deck: the unit’s slide deck is scoped to 60 min and does not follow the plan slide for slide. |
| Prerequisites | Units 05–06 |
| You need | A public EM volume; the Unit 05 organelle table |
| You finish with | A glia recognition checklist and a measured discrimination score by cell class |
The claim this unit has to earn: glia are not background, and correcting glia is not tidying up. Glia and their processes occupy roughly 20–40% of cortical tissue volume, and astrocytic processes specifically form a fine meshwork that interleaves with every neurite in the neuropil. That geometry is exactly what makes them the most frequent partner in a merge error.
Take one merge: a fine astrocytic process is fused with a nearby dendrite.
What happens to the neuron:
What happens to the astrocyte: it loses territory, so any astrocyte morphometry — domain volume, synapse coverage, vascular contact — is also wrong.
What happens to your analysis: the false inputs are not random. They are spatially local, which means they preferentially connect the neuron to its neighbors. A motif analysis will see enhanced local clustering; a distance-dependence analysis will see inflated short-range connectivity. Again the bias points toward an interesting result.
Consequence for triage. Glia-neuron merges should be ranked above many neuron-neuron splits in a proofreading queue, even though a split is more visually obvious. The Unit 08 rule — prioritize by effect on the endpoint metric, not by conspicuousness — has its clearest application here.
Everything that counted the false inputs moves toward the less interesting answer.
What does not change is how the neuron looked before the fix. That is the problem: nothing about the merged object announced the error.
Generalizable principle: when an error’s false positives are spatially local, the bias is toward local structure. Ask where the error’s extra material came from before deciding it is noise.
Each class has one near-diagnostic feature. Learn those three first; everything else is corroboration.
| Nucleus | Pale, euchromatic, often irregular in contour |
| Cytoplasm | Pale and “watery”; few organelles; bundles of intermediate filaments (GFAP) ~10 nm |
| Glycogen granules | 20–30 nm, very dark, in clusters. Neurons do not contain them. |
| Soma size | ~8–10 µm, smaller than most neuronal somata |
| Processes | Irregular, sheet-like, fill the gaps between neuronal elements rather than running as cylinders |
| Perisynaptic processes | Extremely thin (< 100 nm), wrapping synapses — the hardest structures in the volume to segment |
| Endfeet | Flattened expansions covering blood vessel surfaces — a strong contextual cue |
| Absent | Synaptic vesicles, PSDs, and (largely) microtubules |
scripts/render_em_figures.py.
| Nucleus | Small, round, extremely electron-dense heterochromatin. Distinctly darker than neuronal or astrocytic nuclei — usually identifiable at a glance and at low magnification |
| Cytoplasm | Dense, abundant rough ER and ribosomes, prominent Golgi, microtubules present |
| Soma size | ~6–8 µm |
| Processes | Connect the soma to myelin sheaths; each cell myelinates on the order of 20–60 axonal segments |
| Myelin relationship | Inner and outer tongues contain oligodendrocyte cytoplasm; paranodal loops at nodes of Ranvier |
The practical difficulty is not the mature oligodendrocyte — it is the oligodendrocyte precursor cell (OPC/NG2 cell), which has a paler nucleus and can resemble a small neuron or an astrocyte. If a cell looks “sort of oligodendrocyte but the nucleus is not dark enough”, OPC is the leading hypothesis and the correct action is usually to flag rather than force.
scripts/render_em_figures.py.
| Nucleus | Dark, often elongated or bean-shaped, with heterochromatin clumped along the nuclear envelope |
| Cytoplasm | Dense, with characteristic long, narrow ER cisternae; lysosomes, phagosomes, and lipofuscin-like inclusions |
| Processes | Fine, highly irregular contours; frequently seen contacting synapses and vessels |
| State-dependence | Morphology changes substantially with activation state, which makes microglia the least stereotyped of the three |
This is where most errors occur, so it gets its own table.
| Feature | Astrocytic process | Thin neurite |
|---|---|---|
| Glycogen granules | Present | Absent |
| Synaptic participation | None (no vesicles, no PSD) | Vesicles (axon) or PSD (dendrite) |
| Microtubules | Absent or very rare | Usually present |
| Cross-sectional shape | Irregular, angular, sheet-like; fills residual space between other profiles | Roughly circular or elliptical; a coherent tube |
| Cytoplasm | Pale, few organelles | Denser, organelles visible |
| Intermediate filaments | Fine bundles (GFAP) | Absent (neurofilaments differ in distribution and context) |
| Trajectory across sections | Wanders, changes shape section to section | Maintains a continuous, traceable trajectory |
The shape cue is underused and it is robust. Astrocytic processes are space filling — their cross-section is whatever shape is left over after the neurites are packed. Neurites are tubes — their cross-section is a shape in its own right. This cue survives weak staining better than glycogen granules do, which makes it valuable exactly in the regions where everything else fails.
Probably, but this is a medium-confidence call at best, and the reasoning matters more than the answer.
Absence of glycogen granules is weak evidence here: granules are clustered and sparse, so a given thin process may contain none in a given section even when it is astrocytic. This is the Unit 06 lesson about over-reading absence.
Better evidence to seek, in order:
If cues 1–4 do not resolve it, the correct output is uncertain, with a note that the deciding cue was not resolvable. That is a useful annotation.
1. Is this a soma (nucleus visible)?
YES -> nucleus first:
very dark, small, round -> oligodendrocyte (check for myelin links)
dark, elongated, peripheral
heterochromatin + lysosomes-> microglia
pale, irregular, glycogen -> astrocyte
pale, round, big nucleolus,
Nissl/rER, dendrites -> neuron
pale but "not quite" -> OPC candidate -> FLAG
NO -> continue to 2
2. Does the process participate in a synapse
(vesicle cluster or PSD on it)?
YES -> neurite. Go to Unit 06 for axon/dendrite.
NO -> continue
3. Glycogen granules present?
YES -> astrocyte (high confidence)
NO -> continue (absence is weak evidence)
4. Cross-sectional character over 5-10 sections:
space-filling, contour changes to fit gaps -> astrocyte (medium-high)
coherent tube, stable contour -> neurite (medium)
5. Context:
wraps a capillary (endfoot) -> astrocyte (high)
inner/outer tongue of a myelin
sheath, or paranodal loop -> oligodendrocyte (high)
irregular contour + lysosome-rich
parent process -> microglia (medium)
6. Unresolved -> UNCERTAIN, with the missing cue named.
Route to the glia review queue.
Patch: a pale profile, a few hundred nanometers thick and about 2 µm long, pressed flat against the wall of a capillary. The parent process leaves the plane. The segmentation currently assigns it to a nearby dendrite — so this is also a candidate merge, and the call decides an edit.
Step 1 — soma? No nucleus in view. Continue.
Step 2 — synaptic participation? Scroll several sections in each direction: no vesicle cluster, no PSD anywhere on the profile. This step is not skippable even when the vascular context is already shouting the answer. A perivascular call made without ruling out synaptic participation is exactly how a dendrite that happens to end near a vessel stays merged into an astrocyte — or the reverse.
Step 3 — glycogen? None visible. Weak evidence: granules are clustered and sparse, so one thin profile showing none decides nothing. Noted, not spent — the same discipline as the §2 check-yourself.
Step 4 — cross-sectional character over 5–10 sections. The contour does not hold a tube shape. It spreads and flattens to track the vessel wall, section after section — space-filling behavior. Astrocyte-leaning, medium confidence on its own.
Step 5 — context. A flattened expansion covering a vessel surface is the endfoot configuration, and the protocol scores a capillary wrap as high-confidence astrocyte. But one alternative has to be dispatched before taking that: microglial processes also contact vessels, and this is where an honest annotator hesitates. The unit’s own recovery rule supplies the test — microglia require a nucleus or a lysosome-rich parent process, never a fragment. Follow the profile away from the vessel: the parent is pale, organelle-poor, with fine intermediate-filament bundles, and none of the dense, inclusion-rich cytoplasm a microglial process trails back to. Microglia is dismissed for cause, not by default.
Call: astrocytic endfoot; high confidence. The dendrite assignment is therefore a glia–neuron merge, and per §1 it goes into the queue above more conspicuous splits. Evidence chain: space-filling sheet geometry across sections plus vessel-covering context — two independent cues — with synaptic participation ruled out and the microglial alternative rejected on cytoplasmic evidence. Steps 2 and 3 contributed only absences, and were weighted as absences.
Transferable principle: context is the strongest cue in the glial protocol, but it is only safe after the exclusion steps have run. Work the cheap rule-outs first even when the context looks decisive — especially then, because a decisive-looking context is what tempts you to skip them. Two further cases — astrocyte versus thin dendrite, and OPC versus small neuron — are worked in Glia recognition.
scripts/render_em_figures.py.
The call may be right; the note shows the label is being used for the wrong thing.
The appearance fits the protocol’s “pale but not quite” branch: an oligodendrocyte precursor has a paler nucleus than a mature oligodendrocyte and can resemble a small neuron or an astrocyte. So OPC candidate is a reasonable hypothesis, and the protocol’s action for it is flag, not force.
What fails is the note. “OPC” and “uncertain” are different annotations with different follow-ups — an OPC candidate goes to someone who can confirm a cell class, an uncertain call goes to the review queue with the missing cue named. A trainee who writes “OPC, not sure” has merged the two, and if that habit spreads, per-class agreement for OPC stops meaning anything — and OPC agreement is usually the worst of the four classes already (§5).
A better record: OPC candidate — nucleus intermediate in density; no myelin link found in view; neuronal features (large nucleolus, Nissl, dendrites) not seen. That names the evidence for and against, which is what lets a reviewer settle it.
Generalizable principle: a label is a claim, and “uncertain” is its own label. Never let a class name stand in for a confidence level.
Work these against the identification protocol in §3, naming the step that decides each case. They are single planes and the decisive glial cue usually is not one: step 4 asks how a cross-section behaves over five to ten sections, and perisynaptic astrocytic sheets under 100 nm thick are the hardest structures in the volume. Use the panel as a reference for what the cues look like, and do the calling in a volume you can scroll through z.
RIV-GLIA S01: Orientation for glia proofreading. Set the stake before you start looking: glia occupy roughly 20–40% of cortical volume, and a glia-neuron merge does not merely add a branch — it drags a neuron’s arbor past synapses it never contacted, so the false-input yield per micrometer of merged path is unusually high.
RIV-GLIA S03: Astrocyte morphology in a synaptic neighborhood. Read cross-sectional shape before anything else: an astrocytic process is space-filling, taking whatever contour is left over after the neurites pack, where a neurite holds a tube shape of its own. That cue survives weak staining better than glycogen granules do, which makes it the one to reach for where everything else fails.
RIV-GLIA S09: Microglia cues. Look for the pairing that carries the call — a dark, often elongated or bean-shaped nucleus with heterochromatin clumped against the envelope, plus lysosomal and phagosomal content in dense cytoplasm. Microglia are the least stereotyped of the three classes because morphology tracks activation state, so weight nuclear evidence above process shape.
Attribution: Pat Rivlin training materials (MICrONS proofreading deck). Two manifest-listed IDs (S02, S07) were not present in extracted thumbnails and are pending recovery.
Recognition improves with spaced, scored repetition on many short examples, not with reading. This drill is built for that.
Build the set. 40 patches: 10 astrocytic processes, 10 thin neurites, 5 astrocyte somata, 5 oligodendrocyte somata, 5 microglia somata, 5 ambiguous or OPC cases. Show each as a short z-stack (5–10 sections), not a single image — single-plane practice teaches a habit you want to break.
Run it.
| Not yet | Proficient | Strong | |
|---|---|---|---|
| Overall accuracy | < 70% | ≥ 80% | ≥ 85% with a non-zero uncertain rate |
| Error asymmetry | More astrocyte→neurite than neurite→astrocyte | Balanced | Errors skew toward the safe direction, deliberately, and you can say why |
| Cue awareness | Cannot name the deciding cue | Names it | Names it and identifies which cue fails first as staining degrades |
| Matrix reading | Reports accuracy only | Produces the matrix | Diagnoses the dominant confusion and targets the re-drill at it |
| Checklist | Restates the tables | Six actionable checks | Checks ordered by cost, cheapest and most reliable first |
The two error directions have very different costs.
Astrocyte called neurite is the merge-generating error. It puts glial membrane inside a neuron, with all the consequences in §1.
Neurite called astrocyte produces a split — the neurite loses a piece. It is visible (the arbor looks truncated), it is locally fixable, and it does not manufacture false connectivity.
So a well-calibrated annotator working under time pressure should be deliberately biased toward calling ambiguous cases astrocyte or uncertain. This is not sloppiness; it is choosing the cheaper error on purpose, which is what asymmetric loss functions mean in practice. The same logic drives the over-segmentation choice in the pipeline (Unit 04 §1, Stage 4) — the field consistently trades splits for merges, at every level, on purpose.
Treating glia as out of scope. Recover: measure the §5 last metric on your own data and put the number in the project’s QC report.
Over-calling microglia from fragments. Recover: require a nucleus or a lysosome-rich parent process before assigning microglia to a process fragment.
Using OPC as a label for uncertainty. Recover: separate the two. “Uncertain” and “OPC candidate” are different annotations with different follow-ups.
Single-plane calls. Recover: build drills from z-stacks, never single images.
Ignoring astrocyte territory corruption. Recover: if the project makes any glial claim at all, glia must be proofread to a stated standard, not opportunistically.
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:
Measure what your correction pass did to the endpoint, once, and reuse the number. “Correcting glia changed mean input count by X%” is the argument that gets glia into a proofreading budget. Without it, the work reads as tidying.
Rank by impact, not by how obvious the error looks. A small glia merge routinely outranks a large conspicuous split. Queues built by eye get this backwards.
The collected set, and why making these explicit is a fairness intervention rather than etiquette, is in the hidden curriculum.
Vasculature and the neurovascular unit beyond astrocytic endfeet, and glial biology beyond what is needed for identification. Ependymal cells and peripheral glia are out of scope for cortical volume EM.