ATUM
Sample prep · Units 03
Automated tape-collecting ultramicrotome: sections are collected continuously onto tape for later SEM imaging.
Typical: 30-50 nm sections onto tape, imaged later by SEM (Unit 03 §1).
Why it matters: ATUM collects sections onto tape rather than destroying them, so a region can be re-imaged. That single property is why it underpins the largest published volumes.
Often confused with: ATUM (collect then image) vs SBEM (image then destroy). The difference decides whether a bad region is recoverable.
Active zone
Morphology · Units 05
The presynaptic membrane specialization where synaptic vesicles dock and release neurotransmitter.
Typical: 200-800 nm across, matching the PSD it faces (Unit 05 §1).
Why it matters: It is the presynaptic half of the synapse call. A vesicle cluster with no membrane specialization opposite it is not evidence of a synapse — the active zone and the PSD are supposed to be co-extensive, and when they are not, you are usually looking at a tangential cut.
Often confused with: Vesicle cluster (can sit anywhere in a bouton) vs active zone (the docking site at the membrane itself).
Affinity map
Reconstruction · Units 04, 08
A per-voxel prediction of whether neighboring voxels belong to the same object.
Typical: A per-voxel edge score in each direction, thresholded to produce supervoxels.
Why it matters: Intermediate product roughly the size of the raw data; usually deleted after supervoxel generation.
Often confused with: Affinity map vs boundary map. Affinities are between voxel pairs and are directional; a boundary map is a single per-voxel score. The distinction matters when you tune the threshold.
Agglomeration
Reconstruction · Units 04, 08
Merging supervoxels into whole objects, using learned or hand-tuned rules about which neighboring fragments belong to the same cell.
Typical: A single tunable threshold that decides the merge/split balance for the whole volume.
Why it matters: The agglomeration threshold is where the merge/split trade-off is actually set. Move it one way and you buy long correct runs and expensive merges; move it the other and you buy safe fragments and a large proofreading queue.
Often confused with: Agglomeration vs segmentation. Segmentation produces supervoxels; agglomeration decides which of them are the same neuron. Most correctable error is introduced here, not in the first step.
Anisotropy
Imaging · Units 02, 03, 08
Unequal resolution across axes, typical of serial-section EM where z resolution is coarser than xy.
Typical: 4 x 4 x 40 nm is 10x anisotropic in z.
Why it matters: Anisotropy is the largest single driver of automated segmentation error, and it silently breaks any isotropic distance, smoothing, or morphology operation.
Often confused with: Anisotropic voxels vs anisotropic tissue. The first is a property of your sampling (4 x 4 x 40 nm is 10x anisotropic in z); the second is a property of the neuropil. Only the first is yours to choose.
Astrocyte
Cell types · Units 07
A glial cell with pale cytoplasm, glycogen granules, GFAP intermediate filaments, sheet-like space-filling processes, and endfeet on blood vessels.
Typical: Soma ~8-10 um; perisynaptic processes under 100 nm; glycogen granules 20-30 nm (Unit 07 §1).
Why it matters: The most frequent partner in glia-neuron merge errors, because astrocytic processes interleave with every neurite in the neuropil.
Often confused with: Astrocytic process vs thin dendrite or spine neck. The pale watery cytoplasm and glycogen settle it; calibre alone does not. This is the confusion Unit 07 exists to fix, because glia-neuron merges manufacture false connectivity.
Axon
Morphology · Units 05, 06
A neuronal process that transmits output; contains vesicle-filled boutons, lacks ribosomes, and maintains roughly constant calibre between boutons.
Typical: Unmyelinated cortical axons 80-300 nm diameter.
Why it matters: The axon/dendrite call is the most frequent judgment in annotation and it drives everything downstream: reverse it and you reverse the direction of every connection the process makes.
Often confused with: A thin dendrite vs an axon. Calibre alone does not settle it — use the organelle evidence (polyribosomes and abundant rough ER mean dendrite; regularly spaced microtubules with vesicle-filled varicosities mean axon).
Axon initial segment (AIS)
Morphology · Units 05, 06
The proximal axonal segment where action potentials initiate, identified in EM by a dense granular undercoating beneath the membrane and by fasciculated microtubules.
Typical: Roughly 20-60 um long, beginning at the soma or a proximal dendrite.
Why it matters: A rare case where an axon is postsynaptic: chandelier cells synapse onto the AIS, so 'PSD implies dendrite' fails here.
Often confused with: AIS vs node of Ranvier. Both carry a dense undercoating and fasciculated microtubules; the AIS is continuous with the soma, a node sits between two myelin sheaths.
Bin A / Bin B / Bin C claim
Practice · Units 01
A triage of connectomics claims: supported by structure alone (A), by structure plus a declared assumption (B), or not supportable by structure (C).
Typical: Qualitative. Bin A is supported by the structure you measured, Bin B needs an assumption you should name, Bin C needs data you do not have.
Why it matters: Any Bin B claim must name its assumption in the same sentence. The most common failure in connectomics writing is a Bin C sentence in the abstract supported by a Bin A measurement in the results.
Often confused with: The bins are about evidence, not confidence. A Bin C claim stated confidently is still Bin C, and saying which bin you are in is what makes a discussion section checkable.
BossDB
Data & infrastructure · Units 04
A community archive and access service for volumetric neuroscience datasets.
Typical: Cloud-hosted, chunked, accessed through intern or CloudVolume.
Why it matters: It is one of the few places large public EM volumes are hosted with a documented API, so it is often how you get the data at all.
Often confused with: BossDB (storage and serving) vs CAVE (annotation and versioning). They solve different halves of the problem and are frequently used together.
Bouton
Morphology · Units 05, 06
An axonal swelling containing a synaptic vesicle cluster and usually one or more presynaptic release sites.
Typical: 0.5-2 um swellings on an axon; en passant boutons are far more common than terminal boutons in cortical neuropil.
Why it matters: Boutons are where you look for synapses, so how you count them sets your edge weights. A bouton with two active zones onto the same target is one connection with weight two, not two connections.
Often confused with: En passant boutons occur along an axon; terminal boutons at its end. Dendritic varicosities are swellings without vesicle clusters.
CATMAID
Data & infrastructure · Units 04
A collaborative web platform for skeleton tracing in large image volumes.
Typical: Skeleton-first: neurons are traced as node trees rather than segmented as volumes.
Why it matters: It is the tool most manually traced invertebrate connectomes were built in, so its data model shows up in the datasets even where the tool is no longer used.
Often confused with: CATMAID (manual skeleton tracing) vs Neuroglancer (viewing dense segmentation). They represent neurons differently, and a skeleton traced in one is not a segment in the other.
CAVE
Data & infrastructure · Units 04
Connectome Annotation Versioning Engine: the versioned proofreading and annotation platform behind MICrONS and FlyWire.
Typical: Backs MICrONS and FlyWire; queried through caveclient with an explicit materialization version.
Why it matters: CAVE is what makes a petascale segmentation editable and citable at the same time: it versions the edits, so a query can be pinned to a materialization and reproduced later.
Often confused with: CAVE (the annotation and versioning system) vs ChunkedGraph (the edit-tracking data structure inside it). CAVE is the service; the ChunkedGraph is one component.
Calibration
Proofreading & QA · Units 05, 06
The correspondence between stated confidence and actual accuracy.
Typical: Unit 06's rubric asks for high-confidence accuracy at or above 90% alongside a non-trivial uncertain rate; a learner whose high-confidence accuracy equals their overall accuracy is uncalibrated.
Why it matters: In production annotation, calibration is worth more than raw accuracy, because it lets the system route review effort to the cases that need it.
Often confused with: Calibration vs accuracy. You can be 70% accurate and perfectly calibrated, or 85% accurate and useless because your confidence carries no information. Only the first can be worked with.
Cell type
Cell types · Units 09
A grouping of neurons defined by shared morphology, connectivity, molecular identity, or physiology.
Typical: In practice a type is assigned from morphology plus connectivity; EM alone cannot recover transcriptomic type.
Why it matters: Types defined by different criteria do not always agree. When connectivity-defined and transcriptomic types disagree, that disagreement is a finding, not an error to be resolved by preference.
Often confused with: Morphological type vs transcriptomic type. They do not map one-to-one, and a paper that says 'cell type' without saying which it means is not making a checkable claim.
Chandelier cell
Cell types · Units 06
A cortical interneuron that forms axo-axonic cartridge synapses onto the axon initial segments of pyramidal cells.
Typical: Recognized by cartridges of boutons running along axon initial segments rather than by soma appearance.
Why it matters: The canonical example of the polarity-rule exception in Unit 06.
Often confused with: Chandelier cartridges vs ordinary en passant bouton chains. The giveaway is the postsynaptic target: cartridges synapse onto the AIS specifically.
Charging
Imaging · Units 03
Accumulation of electrons on a non-conductive surface, producing bright streaks and local distortion.
Typical: Local, transient, and worst on poorly conductive or thickly embedded regions; drives splits rather than merges (Unit 03 §2).
Why it matters: Raises the split rate. Mitigated by conductive staining, low landing energy, and charge compensation.
Often confused with: Charging streaks vs beam damage. Charging trails the scan direction and can be tuned away; beam damage is cumulative, worsens on re-imaging, and is not recoverable.
Chunked array
Data & infrastructure · Units 04
A large array stored as many independently addressable blocks, so consumers fetch only what they need.
Typical: Chunks commonly 64^3 to 512^3 voxels.
Why it matters: It is why you can open a petabyte volume in a browser: only the chunks in view are fetched, so cost scales with what you look at rather than with what exists.
Often confused with: Chunk size is a performance decision with no right answer. Small chunks make random access cheap and metadata expensive; large chunks do the reverse.
ChunkedGraph
Data & infrastructure · Units 04
A hierarchical graph over immutable supervoxels in which an object is a connected component, enabling fast concurrent edits.
Typical: Stores supervoxel groupings as a hierarchy so an edit is a local change, not a re-segmentation.
Why it matters: A merge becomes adding an edge rather than rewriting voxels - the design that makes distributed proofreading feasible.
Often confused with: It is not a segmentation. It is the structure that records which supervoxels are currently grouped, which is why a root ID is only meaningful with a timestamp or materialization.
CloudVolume
Data & infrastructure · Units 04
A Python client for reading and writing chunked precomputed volumes.
Typical: Reads Precomputed, Zarr and N5 from cloud buckets.
Why it matters: It is the practical way to pull image, segmentation, mesh and skeleton data out of public volumes from Python, and it is what most analysis notebooks start with.
Often confused with: CloudVolume (reads the data) vs caveclient (queries the annotations). A synapse table comes from the second, not the first.
Completeness
Core · Units 01, 08
The fraction of a neuron, or of a volume, that has actually been reconstructed.
Typical: Stated as a fraction of the circuit contained, not of the volume imaged: a 1 mm3 cortical volume still truncates most long-range axons.
Why it matters: Differential completeness between two populations masquerades as biological difference. Match completeness before comparing.
Often confused with: Completeness vs recall. Recall is about detection within your volume; completeness is about whether the volume contains the circuit at all. A perfect reconstruction of an incomplete volume is still incomplete.
Confidence tier
Practice · Units 05, 06
An operational label - high, medium, uncertain - defined by the count and independence of supporting cues.
Typical: Unit 06 asks for high-confidence accuracy at or above 90%, with a non-trivial share of calls left uncertain.
Why it matters: 'Uncertain' is a valid output. The uncertain set is the review priority queue and the best proxy for local data difficulty.
Often confused with: A tier is only useful if it predicts accuracy. If your high-confidence calls are no better than your overall rate, the tier is decoration.
Configuration model
Analysis & statistics · Units 09
A null that preserves each node's in- and out-degree while randomizing connections.
Typical: Preserves each node's in- and out-degree exactly while randomizing who connects to whom.
Why it matters: The minimum acceptable null for a connectome. Erdos-Renyi ignores degree heterogeneity and will show apparent enrichment of almost any motif.
Often confused with: It controls for degree and nothing else. In a spatially embedded connectome, distance explains much of what it will flag as structure, which is why Unit 09 works three nulls rather than one.
Connectivity graph
Analysis & statistics · Units 02, 09
A directed graph whose nodes are neurons (or types, or compartments) and whose edges represent connections.
Typical: Nodes are neurons, edges are connections weighted by synapse count; the threshold that turns synapses into edges is a choice, not a given.
Why it matters: Construction involves at least six consequential choices - node definition, edge definition, threshold, weighting, inclusion criteria, boundary handling - each of which changes the answer and must be reported.
Often confused with: The graph is not the connectome. It is one representation of it, produced by decisions about thresholding, cell inclusion and boundary truncation that all change the answer.
Connectome
Core · Units 01
A map of neural connections within a defined nervous system, region, or volume, at a stated resolution and completeness.
Typical: Ranges from 302 fully verified neurons (C. elegans) to a partially proofread cubic millimeter - which is why the word alone says nothing.
Why it matters: The word alone says nothing about scale or reliability. A 'connectome' may mean 302 fully verified neurons or a partially proofread cubic millimeter. Always state region, resolution, and proofreading level alongside the term.
Often confused with: Connectivity matrix (a representation) vs connectome (the mapped object).
Connectome-constrained model
Analysis & statistics · Units 09
A dynamical model whose connectivity is fixed by a measured connectome, with remaining parameters fitted to data.
Typical: A simulation whose connectivity is fixed to a measured wiring diagram, with dynamics fitted separately.
Why it matters: The strongest current NeuroAI result type: the connectome removes free parameters, which is what makes the model falsifiable.
Often confused with: Constraining a model by a connectome does not validate the connectome. If the reconstruction has systematic errors, the model inherits them and can still fit the data.
Connectomics
Core · Units 01
The study and reconstruction of neural connectivity, spanning acquisition, reconstruction, proofreading, and analysis.
Why it matters: The word covers acquisition, reconstruction, proofreading and analysis, and the failure modes at each stage are different. Most disagreements about a connectomics result are really disagreements about which stage the uncertainty came from.
Often confused with: Connectomics (the field and its methods) vs connectome (the object produced). A methods paper and a dataset paper make very different kinds of claim.
Contrast-to-noise ratio (CNR)
Imaging · Units 03
The contrast between a structure and its background, relative to noise.
Typical: Tracked as a drop against a baseline rather than an absolute: Unit 03 §3 stops acquisition at a >20% membrane CNR drop.
Why it matters: Membrane CNR, not overall image SNR, is the quantity that predicts merge errors. Measure it specifically.
Often confused with: CNR vs SNR. SNR asks whether the signal is above noise; CNR asks whether two tissue classes are separable, which is the only thing segmentation cares about.
Cue family
Practice · Units 05, 06
A grouping of annotation evidence by shared failure mode - geometry, organelle content, synaptic role, context, long-range continuity.
Typical: Units 05-07 ask for two independent families before a confident call - organelle content and geometry, say, rather than two geometric cues.
Why it matters: Two cues that fail together are one cue. High confidence requires two cues from different families.
Often confused with: Two cues from the same family are not independent evidence. Calibre and branching pattern both fail together on a tangential cut.
Curtaining
Imaging · Units 03
Vertical striping in FIB-SEM images caused by uneven milling rate.
Typical: Vertical stripes parallel to the milling direction (Unit 03 §2).
Why it matters: It is the FIB-SEM equivalent of knife chatter: a texture artifact that degrades boundary detection and shows up as extra splits rather than as visible data loss.
Often confused with: Curtaining (FIB-SEM, milling-rate variation) vs knife chatter (diamond-knife sectioning, vibration). Both are periodic banding; only the geometry distinguishes them.
Dendrite
Morphology · Units 05, 06
A neuronal process that typically receives synaptic input; contains ribosomes, tapers with distance from the soma, and may bear spines.
Typical: Shaft diameter 0.5-3 um in cortex.
Why it matters: Dendrites are the postsynaptic side of most excitatory connections, so mislabeling one as an axon inverts every edge it carries. They are also the easier call: polyribosomes and rough ER are near-diagnostic and axons lack both (Unit 05 §1).
Often confused with: Thin distal dendrites vs axons: use ribosomes, taper, and spines rather than calibre.
Dendritic spine
Morphology · Units 05, 06, 08
A protrusion from a dendrite bearing an excitatory postsynaptic site, comprising a head and a narrow neck.
Typical: Head 300-800 nm; neck 50-200 nm diameter.
Why it matters: Spine necks are close to the resolution floor and are a leading cause of split errors.
Often confused with: A spine neck vs a thin axon. Both are sub-100 nm processes; follow it — a spine neck terminates in a head with a PSD and connects back to a dendritic shaft, an axon keeps going.
Dense reconstruction
Core · Units 01, 08
Reconstruction of every object within a volume, rather than a selected subset.
Typical: Kasthuri et al. 2015 reconstructed ~1,500 cubic micrometers densely.
Why it matters: Dense reconstruction is what makes negative results possible: if every process in the volume is reconstructed, an absent connection is evidence of absence rather than of not having looked.
Often confused with: Sparse reconstruction traces selected neurons and leaves the rest unsegmented; it is far cheaper and answers different questions.
Distance-dependent null
Analysis & statistics · Units 09
A null preserving degree and the empirical relationship between connection probability and distance.
Typical: Preserves the connection-probability-versus-distance curve while randomizing the rest.
Why it matters: Required for spatially embedded data, which means all volume EM. Many reported motif enrichments do not survive it.
Often confused with: It is the null that usually matters most in cortex and is the one most often skipped. Erdos-Renyi will call almost any spatially embedded network 'structured'.
DotMotif
Analysis & statistics · Units 09
A declarative language and toolset for expressing and searching connectome motifs.
Typical: Queries subgraph patterns against a connectome graph.
Why it matters: It turns motif search into a declarative query, so the definition of the motif you tested is written down and reviewable rather than buried in loop code.
Often confused with: Finding a motif is not showing it is enriched. The count needs a null model before it means anything.
Dwell time
Imaging · Units 03
The time the beam spends on each pixel.
Typical: 0.1-2 microseconds.
Why it matters: SNR grows roughly as the square root of dose, so doubling SNR costs roughly four times the acquisition time.
Often confused with: Longer dwell buys SNR but also dose. It trades against beam current for the same signal, and against total acquisition time linearly — which at petascale is the binding constraint.
Egress
Data & infrastructure · Units 04
The cost of moving data out of a cloud region.
Typical: The dominant cost of working with public petascale data from outside the hosting cloud, and the reason analysis is usually run next to the data.
Why it matters: Moving a petabyte out can cost more than storing it for a year. Bring compute to the data.
Often confused with: Storage cost vs egress cost. Storage is charged for holding the data; egress is charged for moving it, and at petabyte scale egress is what surprises people.
Elastic / non-linear alignment
Reconstruction · Units 04
Registration allowing local deformation, needed because sections deform non-rigidly.
Typical: Unit 03 §3 asks for a 99th-percentile alignment residual under 1 voxel at native xy before ingest.
Why it matters: A flexible enough warp will align anything, including things that do not correspond. Escalate from rigid to affine to elastic only as residuals demand.
Often confused with: Elastic alignment can hide a real problem by deforming tissue to fit. A low residual on badly folded sections means the fit absorbed the fold, not that the fold is gone.
En bloc staining
Sample prep · Units 03
Applying stain to the resin-embedded block rather than to cut sections.
Typical: Heavy-metal staining of the whole block before embedding, rather than post-section — the prerequisite for block-face and tape-based methods (Unit 03 §1).
Why it matters: Required for block-face methods, where sections are never handled separately.
Often confused with: En bloc vs post-section staining. Post-section staining cannot be used for block-face imaging, because there is no free-standing section to stain.
Endpoint metric
Proofreading & QA · Units 08, 09
The scientific quantity a project is actually measuring, as opposed to a generic quality score.
Typical: Whatever number the paper actually reports - reciprocity rate, motif count, mean degree - rather than a segmentation-quality score.
Why it matters: Ship on the endpoint. 'Exhaustive proofreading of a 20-cell sample changed the ratio from 3.1 to 2.8' is a stronger quality statement than any VI value.
Often confused with: Endpoint metric vs quality metric. VI and ERL measure the reconstruction; the endpoint metric measures your claim. Optimizing the first without checking the second is the most common way to spend a proofreading budget badly.
Error sensitivity analysis
Analysis & statistics · Units 09
Simulating measured merge and split rates on the reconstructed graph and reporting the resulting spread on the effect size.
Typical: Inject merge and split errors at the rates the source dataset reports, and re-measure your endpoint.
Why it matters: Cheap to run and one of the strongest things you can put in a supplement. If the band crosses the null, say so.
Often confused with: Error sensitivity is not the same as a confidence interval. A CI covers sampling noise; this covers reconstruction error, which is systematic and usually larger.
Expected Run Length (ERL)
Proofreading & QA · Units 08
The mean error-free path length along skeletons before an error is encountered.
Typical: Reported in um of correctly reconstructed path; sensitive to merge errors, which truncate runs sharply.
Why it matters: Answers 'how far can I trace before hitting a problem?' Relatively insensitive to merges unless merges are explicitly penalized.
Often confused with: ERL rewards long correct runs and punishes merges hard, so it can look excellent on a volume riddled with splits. Pair it with a split-sensitive measure.
F1 score
Proofreading & QA · Units 08
The harmonic mean of precision and recall, used to summarize detection performance - most often synapse detection - in a single number.
Typical: Reported per class; a synapse detector's F1 is not comparable across datasets with different synapse densities.
Why it matters: Weights the two equally, which is rarely right for connectomics, where merge and split errors have very different costs.
Often confused with: F1 weights precision and recall equally, which is almost never what connectomics wants. Merge and split errors have very different costs, and so do false and missed synapses.
FIB-SEM
Imaging · Units 03
Focused ion beam SEM: an ion beam mills a few nanometers at a time, giving isotropic voxels.
Typical: 4-8 nm isotropic; volumes typically 10^5-10^6 cubic micrometers.
Why it matters: The best tracing conditions available, limited by volume and speed.
Often confused with: FIB-SEM vs SBF-SEM. Both are block-face methods, but FIB mills with an ion beam to a few nm and gives isotropic voxels, while SBF cuts with a diamond knife at 25-50 nm. FIB buys tracing quality and pays in volume.
Fixation
Sample prep · Units 03
Chemical cross-linking, typically with glutaraldehyde and paraformaldehyde, that arrests ultrastructure before autolysis.
Typical: Aldehyde fixation followed by osmium; quality is judged by membrane continuity and the absence of extraction, not by a number.
Why it matters: Slow or failed fixation shows as swollen astrocytic processes and enlarged extracellular space, which distorts every geometric measurement.
Often confused with: Fixation quality is often blamed for what is really a staining or imaging problem. Weak membranes can come from any of the three, and the fix differs in each case.
Flood-filling network (FFN)
Reconstruction · Units 08
A segmentation approach that iteratively grows a single object from a seed, conditioned on the mask built so far.
Typical: Iterative, single-object; expensive to run and the reason segmentation compute is quoted in GPU-days.
Why it matters: FFNs made mm-scale reconstruction feasible by tracing one object at a time with a moving field of view, rather than classifying every voxel independently.
Often confused with: FFN (object-at-a-time, recurrent) vs affinity-based pipelines (voxel-at-a-time, then agglomerate). They fail differently, so their error populations are not interchangeable.
Fold
Imaging · Units 03
A wrinkle in a collected section, producing a band of duplicated or compressed tissue.
Typical: Unit 03 §3 flags a section when folds cover more than 5% of its area.
Why it matters: A data-loss artifact: the tissue in the fold cannot be recovered by any downstream effort.
Often confused with: A fold vs a tear. A fold duplicates and compresses tissue into a dark band; a tear removes it. Both are data loss, but a fold can masquerade as dense staining.
Glycogen granule
Morphology · Units 07
A small, very electron-dense particle found in clusters in astrocytes and not in neurons.
Typical: 20-30 nm.
Why it matters: Near-diagnostic for astrocyte when present; absence is weak evidence because granules are sparse and clustered.
Often confused with: Glycogen granules vs ribosomes. Both are small and dark, both cluster. Glycogen is 20-30 nm and effectively astrocyte-only; ribosomes are ~25 nm, sit on ER membranes or in rosettes, and mark dendrites and somata.
Gold standard
Proofreading & QA · Units 08
An exhaustively verified reference reconstruction used to measure error rates.
Typical: A small densely proofread subvolume, usually agreed by more than one annotator.
Why it matters: Every quality number you report is measured against it, so its own error rate becomes a floor on what you can claim. A gold standard built by one annotator carries that annotator's systematic mistakes into every downstream evaluation.
Often confused with: Gold standard vs ground truth. Ground truth is what is actually there; a gold standard is a human's best reconstruction of it, and the gap between them is real.
Graph matching
Analysis & statistics · Units 09
Finding the correspondence between the nodes of two graphs, for example left and right hemisphere connectomes.
Typical: Used to align two connectomes - across individuals, hemispheres, or species - when node identities are not shared.
Why it matters: Yields a measure of wiring stereotypy - a question only connectomics can answer.
Often confused with: A good match score does not imply corresponding cells. With enough nodes, plausible alignments exist between graphs with no biological correspondence at all.
Gray type I / asymmetric synapse
Morphology · Units 05
A synapse with a thick prominent PSD, wide cleft, and round clear vesicles; usually onto spines.
Typical: Cleft 20-30 nm; PSD 30-50 nm thick and 200-800 nm wide; round clear vesicles 35-50 nm (Unit 05 §1).
Why it matters: Usually inferred to be excitatory (glutamatergic). This is an inference from morphology, not an observation - state it as 'putatively excitatory'.
Often confused with: Type I vs type II. The discriminator is the thickness asymmetry between pre- and postsynaptic densities, not the vesicle shape alone — vesicle flattening is partly a fixation artifact.
Gray type II / symmetric synapse
Morphology · Units 05
A synapse with pre- and postsynaptic densities of comparable thickness, narrower cleft, and pleomorphic vesicles; usually onto shafts, soma, or AIS.
Typical: Cleft ~15-20 nm; densities thin and near-equal on both sides; vesicles often pleomorphic (Unit 05 §1).
Why it matters: Usually inferred inhibitory (GABAergic), with the same caveat about inference.
Often confused with: A genuinely symmetric synapse vs an asymmetric one cut tangentially, which can lose the appearance of PSD thickening. Check across sections before calling it.
GridTape
Sample prep · Units 03
A reinforced tape substrate with apertures allowing transmission-mode imaging of tape-collected sections.
Typical: Used for ssTEM at 4 x 4 x 40 nm scale.
Why it matters: It is what makes high-throughput TEM of serial sections practical: sections land on an indexed tape with support films already in place, so imaging can be automated.
Often confused with: GridTape (TEM, film-backed apertures) vs ATUM tape (SEM, solid tape). Both are reel-to-reel collection; the imaging modality differs.
Idempotency
Data & infrastructure · Units 04
The property that re-running a stage on the same inputs produces the same output.
Typical: A property a pipeline stage either has or does not; the ones that lack it are where reruns silently corrupt state.
Why it matters: Broken by GPU non-determinism, unpinned dependencies, and unseeded randomness. Pin them and test the re-run.
Often confused with: Idempotent is not the same as deterministic. A stage can produce identical output every run and still double-write if it is invoked twice.
Inter-annotator agreement
Proofreading & QA · Units 05, 08
The degree to which independent annotators produce the same labels on the same data.
Typical: Reported alongside throughput; Unit 08 treats throughput without agreement as an unusable performance measure.
Why it matters: Report it alongside throughput. Optimizing throughput alone reliably produces fast, inconsistent annotation.
Often confused with: Agreement is not accuracy. Two annotators trained the same way can agree closely and both be wrong in the same direction, which is exactly what happens when a cue is taught badly.
Isotropic
Imaging · Units 02, 03
Equal resolution in all three axes, as produced by FIB-SEM.
Typical: 4-8 nm isotropic.
Why it matters: Isotropic voxels remove the direction-dependence of every tracing decision — a process running in z is as traceable as one running in x. It is the single largest quality difference between FIB-SEM and serial-section methods.
Often confused with: Isotropic acquisition vs isotropic resampling. Interpolating anisotropic data onto a cubic grid produces isotropic voxels without producing isotropic information.
Knife chatter
Imaging · Units 03
Periodic banding perpendicular to the cutting direction, from knife or block vibration.
Typical: Periodic bands at fixed spacing, perpendicular to the cutting direction; classed as labor rather than data loss (Unit 03 §2).
Why it matters: Adds false boundaries and raises the split rate; recoverable by proofreading.
Often confused with: Chatter vs curtaining — see Curtaining. Chatter is also easily mistaken for genuine periodic structure such as myelin lamellae.
Landing energy
Imaging · Units 03
The energy at which electrons strike the sample in SEM.
Typical: 1-2 keV for block-face connectomics.
Why it matters: Higher energy gives more signal from deeper, at the cost of charging and beam damage.
Often confused with: Higher landing energy reaches deeper and gives more signal, but also charges and damages more. It is not a quality dial; it is a trade against surface specificity (Unit 03 §1).
Local shape descriptor
Reconstruction · Units 08
An auxiliary learned representation of local object shape used to improve agglomeration decisions.
Typical: Auxiliary per-voxel targets - local object size, orientation, offset - predicted alongside affinities.
Why it matters: It gives an affinity-based pipeline some of the object-level context an FFN gets for free, at a fraction of the compute.
Often confused with: It is a training signal, not a segmentation. The output is still affinities that need agglomerating.
Lost section
Imaging · Units 03
A section destroyed or not collected, producing a z-gap in the volume.
Typical: Unit 03 §3 stops acquisition at more than 2 consecutive lost sections, or a cumulative rate above 1%.
Why it matters: Distribution matters more than count. Four scattered losses are routine; four consecutive losses cut the dataset in two for thin processes.
Often confused with: A lost section is unrecoverable data loss; a misaligned section is recoverable labor. They look similar in a z-scroll and cost completely different amounts.
Materialization
Data & infrastructure · Units 04, 09
A frozen snapshot in which annotations are joined to the segmentation state at a specific timestamp and written to queryable tables.
Typical: Named by version number and date; a query without one is not reproducible (Unit 04 §1).
Why it matters: Analyses run against a materialization version, and that version belongs in your methods section like a software version or genome build.
Often confused with: Materialization version vs dataset name. 'minnie65_public' names the dataset; only the materialization version names the state of it that your number came from.
Merge error
Proofreading & QA · Units 06, 07, 08, 09
Two or more distinct objects incorrectly joined into one segment - two neurons fused, or a glial process fused to a neuron.
Typical: The expensive class: hard to spot, because the merged object still looks like a plausible neuron (Unit 08 §1).
Why it matters: The expensive error: invisible in summary statistics, it manufactures connectivity, and it biases motif counts superlinearly toward denser motifs.
Often confused with: Splits are visible and locally repairable; merges are neither.
Mesh
Reconstruction · Units 02, 04
A triangulated surface representation of a segmented object.
Typical: 10-100 MB per neuron across levels of detail; 1-10 TB for a whole volume (Unit 02 §3, Unit 04 §2).
Why it matters: Meshes are what you actually look at in Neuroglancer, and what surface-area and spine-shape measurements come from. They are regenerated whenever an edit changes the segmentation.
Often confused with: Mesh vs skeleton. A mesh carries surface geometry and no topology; a skeleton carries topology and radius and no surface. Path length comes from the skeleton, spine head shape from the mesh.
Microglia
Cell types · Units 07
The resident immune cell of the CNS; identified by a dark elongated nucleus with peripheral heterochromatin, long narrow ER cisternae, and lysosomal inclusions.
Typical: Soma ~6-8 um, smaller than most neurons, with dark dense cytoplasm (Unit 07 §2).
Why it matters: Morphology varies with activation state, making microglia the least stereotyped glial class.
Often confused with: Microglia vs oligodendrocytes. Both have dark cytoplasm; oligodendrocytes sit near myelinated axons and have a rounder, denser nucleus, while microglia have irregular processes and often contain inclusions.
Microtubule
Morphology · Units 05, 06
A cytoskeletal tubule, present in both axons and dendrites; fasciculated bundles are characteristic of the AIS.
Typical: ~25 nm outer diameter.
Why it matters: Microtubule arrangement is a primary axon/dendrite cue: loose parallel arrays suggest dendrite, regular fasciculated spacing suggests axon or AIS. It is one of the few cues visible in a single section.
Often confused with: Microtubules (~25 nm, hollow ring in cross-section) vs neurofilaments (~10 nm, solid, usually bundled). The size difference is the reliable discriminator.
Modularity / community detection
Analysis & statistics · Units 09
Partitioning a network into groups with denser connectivity inside groups than between them, and scoring how well the partition separates them.
Typical: Resolution-dependent: the same graph yields different module counts under different resolution parameters.
Why it matters: Returns a partition for any graph, including a random one, and has a resolution limit. Always compare against degree-preserving rewired graphs.
Often confused with: Modularity always returns communities, including on random graphs. A module count means nothing without a null comparison and a stated resolution.
Motif
Analysis & statistics · Units 09
A small subgraph pattern whose frequency is compared against a null model.
Typical: Two- and three-node patterns; reciprocal pairs and the triad census are the usual starting points.
Why it matters: Motif enrichment is one of the main structural claims connectomics makes, and it is entirely dependent on the null chosen. Unit 09 shows the same data supporting '2.9x enrichment' or 'no detectable effect' depending on that choice.
Often confused with: Motif count vs motif enrichment. The count is a fact about your graph; the enrichment is a comparison, and only the comparison is a claim.
Multibeam SEM
Imaging · Units 03
SEM using many electron beams in parallel to multiply throughput.
Typical: 61 or 91 beams, aggregating on the order of a gigapixel per second.
Why it matters: The throughput advance that moved cubic-millimeter volumes from impossible to feasible.
Often confused with: Multibeam raises throughput, not resolution. It is the reason mm-scale ssEM volumes are feasible at all, but each beam still images at the same nm scale as a single beam.
Multiple comparisons
Analysis & statistics · Units 09
The inflation of false positives when many hypotheses are tested.
Typical: A triad census tests 16 patterns at once; a cell-type-pair analysis tests as many pairs as you have types squared.
Why it matters: Report the number of tests run, including the ones you ran and did not report.
Often confused with: Correcting for the tests you report is not enough. The count that matters includes the tests you ran and discarded, which is why pre-specifying matters more than the correction method.
Multiresolution pyramid
Data & infrastructure · Units 02, 04
A series of progressively downsampled copies of a volume, enabling interactive zoom.
Typical: Costs roughly 30-50% additional storage.
Why it matters: It is what lets a viewer show a whole volume and a single synapse with the same interaction cost, and it is why a stored volume is larger than its raw data.
Often confused with: The pyramid adds roughly 30-50% on top of the base resolution (Unit 04 §2), which is a real line in a storage budget rather than a rounding error.
Myelin
Cell types · Units 06, 07
Concentric wraps of oligodendrocyte membrane insulating an axon.
Typical: Compact lamellae with a period of roughly 12 nm; sheath thickness scales with axon calibre.
Why it matters: Presence of myelin identifies a process as axonal with high confidence.
Often confused with: Myelin lamellae vs knife chatter. Both are periodic banding, and at low magnification in a poorly stained block they are genuinely hard to tell apart.
Nanoscale connectomics
Core · Units 01, 02
Connectomics at synapse resolution, requiring electron microscopy or comparable methods, as distinct from mesoscale projection mapping or macroscale tractography.
Typical: Voxels of a few nm in xy; the only scale at which a synapse is directly resolvable (Unit 02 §1).
Why it matters: The qualifier is what commits you to EM-scale cost. Dropping it in a proposal invites reviewers to assume a cheaper method would do.
Often confused with: Nanoscale (EM, synapse-resolution) vs macroscale (MRI tractography, 0.5-2 mm). Both are called connectomics and they support entirely different claims.
Neurite
Morphology · Units 05, 06
A generic neuronal process, used when the axon/dendrite distinction has not yet been made.
Typical: Unmyelinated processes in cortical neuropil are 80-300 nm; dendritic shafts 0.5-3 um; spine necks 50-200 nm (Unit 06 §1).
Why it matters: Useful precisely because it is non-committal. Prefer it to a guess.
Often confused with: Used as a deliberate hedge when the axon/dendrite call has not been made. Saying 'neurite' is an honest 'I have not decided'; using it after you have decided loses information.
Neurofilament
Morphology · Units 05, 06
An intermediate filament abundant in axons, especially myelinated ones.
Typical: ~10 nm diameter.
Why it matters: Dense neurofilament bundles point to a myelinated axon, which is one of the more reliable large-calibre cues in white matter and deep neuropil.
Often confused with: Neurofilaments (~10 nm) vs microtubules (~25 nm) vs astrocytic GFAP intermediate filaments (also ~10 nm). Filament diameter alone cannot separate neurofilament from GFAP — use the surrounding cytoplasm, which is pale and organelle-poor in astrocytes.
Neuroglancer
Data & infrastructure · Units 04
A browser-based viewer for petascale volumes, meshes, and annotations.
Typical: Browser-based; state encoded in the URL, which is what makes a view citable.
Why it matters: It is the shared viewer of the field, and a Neuroglancer link with a pinned segmentation source is the standard way to point someone at exactly what you are looking at.
Often confused with: A Neuroglancer link records the view, not the data version. Two people can open the same link and see different segments if the underlying segmentation has been edited.
Node of Ranvier
Morphology · Units 06, 07
A gap in the myelin sheath along a myelinated axon, where the axolemma shows a dense undercoating.
Typical: ~1 um of exposed axolemma between sheaths; nodal undercoating ~20 nm, as at the AIS (Unit 05 §1).
Why it matters: Nodes anchor the identity of a myelinated axon across a break in its sheath, and they are where automated segmentation most often splits a long-range axon.
Often confused with: Node vs a myelin sheath that has simply ended, and node vs AIS. A node has paranodal loops on both sides; a sheath end has them on one.
Non-claim
Practice · Units 01
An explicitly stated sentence a study will not assert, with the reason.
Typical: One sentence, written before the analysis, naming something the data cannot establish.
Why it matters: The best single predictor of a weak study brief is an empty non-claim. If you cannot name a sentence you refuse to write, you have not found the boundary of your evidence.
Often confused with: A non-claim is not a limitation paragraph. It is a specific sentence you refuse to write, decided in advance - which is why Unit 01's lab asks for it before any result exists.
Null model
Analysis & statistics · Units 09
A randomized reference distribution encoding what would count as an uninteresting explanation.
Typical: Qualitative by nature. What matters is that it is stated, and stated before the test.
Why it matters: Choosing the null is the scientific step; running the test is bookkeeping. Preserve everything you are not asking about.
Often confused with: The null is not a technicality; it is the claim. Choosing it after seeing the data converts a hypothesis test into a description, and the two do not carry the same weight.
OME-Zarr / N5
Data & infrastructure · Units 04
Cloud-oriented chunked array formats used for large bioimaging volumes.
Typical: Chunked, compressed, with multiscale metadata alongside the arrays.
Why it matters: They are the open, chunked formats that let a volume be read by many tools without conversion, which is the difference between an archived dataset and a usable one.
Often confused with: OME-Zarr and N5 are storage layouts, not viewers or databases. Reading one still requires a client that understands the multiscale metadata.
Oligodendrocyte
Cell types · Units 07
A glial cell that myelinates CNS axons; identified by an extremely electron-dense small round nucleus and dense ribosome-rich cytoplasm.
Typical: Each cell myelinates roughly 20-60 axonal segments.
Why it matters: Oligodendrocyte processes connect to myelin sheaths, and a sheath merged into its axon is a common and costly reconstruction error in white matter.
Often confused with: Oligodendrocyte vs microglia — see Microglia. Also oligodendrocyte vs OPC, which requires context rather than cytoplasm.
Oligodendrocyte precursor cell (OPC / NG2 cell)
Cell types · Units 07
A glial progenitor with a paler nucleus than a mature oligodendrocyte; the hardest routine glial identification.
Typical: Soma similar to a small oligodendrocyte, with more branched processes and, unusually for glia, genuine synaptic input.
Why it matters: Do not use 'OPC' as a synonym for 'uncertain'; they are different annotations with different follow-ups.
Often confused with: OPC vs oligodendrocyte. OPCs receive synapses, which is the one cue that separates them structurally — and it also means an OPC process can be mistaken for a dendrite.
Orphan fragment
Proofreading & QA · Units 08
A segmented piece not attributed to any traced object.
Typical: A segment with no soma and no confident continuation; most of a raw segmentation by object count.
Why it matters: Orphans are the visible part of the split problem, and they dominate the count of objects in a raw segmentation while contributing almost nothing to connectivity.
Often confused with: An orphan fragment vs a genuinely truncated process at the volume boundary. The second is a completeness limit, not an error.
Paranode
Morphology · Units 07
The region flanking a node of Ranvier where myelin terminal loops form junctions with the axon.
Typical: A few hundred nm of tapering loops flanking each side of the node.
Why it matters: Paranodal loops are the structural evidence that you are at a node rather than at a gap in a sheath, so they settle a call that otherwise depends on tracing many sections.
Often confused with: Paranode (loops contacting the axolemma) vs juxtaparanode (the region just beyond it, molecularly distinct but not separable by morphology in EM).
Perisynaptic astrocytic process (PAP)
Cell types · Units 07
A very thin astrocytic lamella wrapping a synapse, part of the tripartite synapse.
Typical: Often under 100 nm thick - among the hardest structures in the volume to segment.
Why it matters: PAPs are the hardest structures in the volume to segment: sub-100 nm sheets wrapping the very synapses you are trying to detect. Where they are merged into a neuron, they manufacture connectivity that does not exist.
Often confused with: A PAP vs a spine neck or a thin axon. All three are under 100 nm. The astrocytic cytoplasm is the cue, and at that calibre there is often too little of it to see.
Peters' rule
Analysis & statistics · Units 01, 02
The hypothesis that connectivity can be predicted from the spatial overlap of axonal and dendritic arbors.
Typical: Predicts connection probability from axo-dendritic overlap alone.
Why it matters: A poor predictor of actual synaptic connectivity in cortex, which is precisely why light-microscopy overlap cannot substitute for EM.
Often confused with: Peters' rule is a null hypothesis, not a finding. Its value is as the baseline that specific connectivity has to beat, and it is routinely cited as though it were a description of how brains wire.
Pilot reconstruction
Practice · Units 03
Taking a small sub-volume through the entire pipeline before committing to full acquisition.
Typical: A small subvolume run through the full pipeline to measure the error rate before committing to scale (Unit 03 §3).
Why it matters: Costs 1-2% of a project and is the only way to discover a prep-level problem while the prep can still be changed.
Often confused with: A pilot tests the pipeline, not the biology. Its purpose is to find out whether your proofreading budget can absorb the error rate you are about to generate.
Polyribosome
Morphology · Units 05, 06
A cluster of ribosomes on mRNA, seen as a rosette in EM; present in somata and dendrites, effectively absent from axons.
Typical: Individual ribosomes ~25 nm, seen in rosettes of five or more (Unit 05 §1).
Why it matters: The single most reliable axon/dendrite discriminator when visible. Absence, however, is weak evidence in very thin processes.
Often confused with: Polyribosome rosettes vs glycogen granule clusters. Both are dark and clustered at similar scale, but glycogen means astrocyte and polyribosomes mean dendrite or soma — so confusing them inverts the cell-class call, not just a detail.
Postsynaptic density (PSD)
Morphology · Units 05
The electron-dense protein assembly beneath the postsynaptic membrane.
Typical: 30-50 nm thick, 200-800 nm wide.
Why it matters: Its thickness relative to the presynaptic side is what distinguishes Gray type I from type II.
Often confused with: A PSD vs a tangentially cut membrane, which also appears as a dark thickening. A real PSD has a cleft of uniform width and a vesicle cluster opposite it; a grazing cut has neither.
Precision
Proofreading & QA · Units 08
Of the items called positive, the fraction that are correct.
Typical: Reported with recall, never alone.
Why it matters: In synapse detection, precision is what stops you inflating connectivity: every false positive adds an edge that was never there, and weak one-synapse connections are where false positives concentrate.
Often confused with: Precision vs recall. Precision asks 'of what I called, how much was real'; recall asks 'of what was real, how much did I find'. Reporting one without the other is uninterpretable.
Precomputed format
Data & infrastructure · Units 04
The chunked, multiresolution volume format used by Neuroglancer.
Typical: Chunked image, segmentation, mesh and skeleton sources under one directory tree.
Why it matters: It is Neuroglancer's native layout, so a volume served as Precomputed can be opened by anyone with a URL and no local software.
Often confused with: Precomputed (Neuroglancer's layout) vs OME-Zarr (the community standard). Both are chunked; they are not interchangeable without conversion.
Proofreading
Proofreading & QA · Units 08
Human correction of automated reconstruction errors - splitting merged objects, joining split ones, and adjudicating synapses - under a fixed time budget.
Typical: Always partial. Unit 08 frames it as an allocation problem, not a cleanup phase with an end.
Why it matters: At petascale this is an allocation problem under a fixed budget, not an open-ended cleanup. It is usually the dominant project cost.
Often confused with: Proofreading is not quality assurance. QA measures what you have; proofreading changes it. A volume can be heavily proofread and still unmeasured.
Proofreading level
Proofreading & QA · Units 08
A defined tier of reconstruction completeness for a cell, with written criteria.
Typical: Stated per cell or per class, not per volume: 'somata verified, axons proofread to 10 um, dendrites untouched' is a usable statement.
Why it matters: Every connectomics result should state the proofreading level of the cells it rests on. Without it, a low connection count cannot be distinguished from incompleteness.
Often confused with: There is no field-wide standard for what a level means, so a level is only interpretable if the paper defines its own. A result reported without one cannot be compared to anything.
Provenance
Data & infrastructure · Units 03, 04
The recorded lineage of a data product: inputs, code revision, parameters, model version, and environment.
Typical: At minimum: dataset, materialization version, query code, and date. Unit 04's lab asks for exactly this header on every notebook.
Why it matters: If it is not machine-readable, it does not exist. Per-tile timestamps in particular are what let you ask whether a defect follows block position, anatomy, or acquisition time.
Often confused with: Provenance is not citation. Citing the paper tells a reader which project produced the data; provenance tells them which state of it produced your number.
Recall
Proofreading & QA · Units 08
Of the items that are truly positive, the fraction found.
Typical: Reported with precision, never alone.
Why it matters: Low synapse-detection recall shrinks every edge weight in your graph, and it does so unevenly - thin processes and small synapses are missed most, which biases exactly the connections that are hardest to interpret.
Often confused with: Recall vs completeness. Recall is about detection within the volume you have; completeness is about how much of the circuit that volume contains at all.
Reciprocity
Analysis & statistics · Units 09
The frequency of bidirectional connections between node pairs.
Typical: Reported as an enrichment over a null, never as a raw rate.
Why it matters: The textbook case where the apparent effect collapses as the null gets stronger - degree heterogeneity and spatial proximity together explain much of the naive enrichment.
Often confused with: Reciprocity is unusually sensitive to merge errors: one merge that fuses a pre- and postsynaptic partner manufactures a reciprocal pair that does not exist.
Registration
Reconstruction · Units 02, 04
Aligning images to a common coordinate frame - across sections, across modalities, or to an atlas.
Typical: Applied per section pair, then globally; quality is judged by residual, not by appearance.
Why it matters: Report residuals locally with a maximum, not as a global mean, and validate on held-out anchors.
Often confused with: Registration (aligning sections to each other) vs stitching (assembling tiles within one section). Both are alignment; they fail at different scales and are diagnosed differently.
Root ID
Data & infrastructure · Units 04
The identifier of a segmented object as of a particular moment in time.
Typical: Changes whenever an edit touches the object; stable only within a materialization (Unit 04 §1).
Why it matters: Not stable. A root ID in a notebook, paper, or figure caption is meaningless without a version or timestamp.
Often confused with: A root ID is not a neuron's name. It is a pointer to a grouping at a moment, and reusing one across versions is the most common reproducibility failure in this field.
SBF-SEM / SBEM
Imaging · Units 03
Serial block-face scanning EM: image the block face, shave off a slice inside the chamber, repeat.
Typical: 10-20 nm xy, 25-50 nm z.
Why it matters: No section handling means no lost sections and better z-alignment, at the cost of destroying the sample.
Often confused with: SBEM vs ssTEM/ssSEM. SBEM destroys each section as it images the block face, so there is nothing to re-image; serial-section methods keep the sections, so a lost region can sometimes be recovered.
Saturated reconstruction
Core · Units 01
Dense reconstruction carried to the point where every object in the volume is accounted for, including glia and vasculature.
Typical: Achieved in small volumes; approximated, never reached, at mm scale.
Why it matters: It is the strongest completeness claim available - every object in the volume reconstructed, not just the ones of interest - and it is what a dense volume has to achieve before absence means anything.
Often confused with: Saturated vs dense. Dense means everything was segmented; saturated means everything was segmented and verified. The second is much rarer than the word's casual use suggests.
Segmentation
Reconstruction · Units 04, 08
Assigning each voxel to an object, producing neuron and glia labels from image data.
Typical: 0.2-0.8 PB of labels for a mm3 volume, with label-aware compression (Unit 04 §2).
Why it matters: Everything downstream is a property of the segmentation, not of the tissue. A connectome is a claim about a specific segmentation at a specific version, and saying so is the difference between a reproducible result and an anecdote.
Often confused with: Segmentation vs reconstruction. Segmentation is the automated labeling; reconstruction is the whole pipeline through proofreading. Papers use them interchangeably and the quality claims are not interchangeable.
Serial section EM (ssEM / ssTEM / ssSEM)
Imaging · Units 03
Volume EM in which the block is cut into thin sections that are collected and then imaged.
Typical: 4 x 4 x 40 nm is the working standard for large volumes, up to ~1 mm3 today (Unit 02 §1).
Why it matters: Sections survive imaging, so they can be re-imaged and imaged in parallel across instruments - which is what makes petascale acquisition possible in finite time.
Often confused with: Block-face methods destroy the imaged material and cannot be re-imaged.
Sharding
Data & infrastructure · Units 04
Bundling many small chunks into larger files with an index, to avoid per-object overhead in object stores.
Typical: Groups many small chunks into fewer large files, trading random-access granularity for a manageable object count.
Why it matters: At petascale, per-request charges on billions of tiny objects can exceed storage costs.
Often confused with: Sharding solves a filesystem problem, not a data problem. It exists because object stores handle a billion tiny files badly.
Shrinkage
Sample prep · Units 03
Tissue volume loss during dehydration and embedding.
Typical: Roughly 5-20% linear, depending on protocol.
Why it matters: Systematic, not random. Prefer ratios and within-volume comparisons over absolute lengths compared across studies.
Often confused with: Shrinkage is systematic and affects every measurement you report in nm. It is not noise and it does not average out — a volume measurement taken from EM is a measurement of fixed, dehydrated tissue.
Skeletonization
Reconstruction · Units 02, 04
Reducing a segmented object to a centreline graph with radii.
Typical: 0.1-5 MB per neuron; 10-100 GB for a whole volume, which makes skeletons cheap enough to archive (Unit 02 §3, Unit 04 §2).
Why it matters: The representation you need for path distance from soma; cheap to store, so archive it alongside any connectivity graph.
Often confused with: Skeletonization method changes the answer. Path length and branch counts differ measurably between TEASAR-style and mesh-contraction skeletons, so a morphology statistic is only comparable within one method.
Soma
Morphology · Units 05
The neuronal cell body, containing the nucleus, Nissl substance (rough ER), and Golgi apparatus.
Typical: Cortical neuron soma 10-25 um diameter.
Why it matters: Somata anchor identity: a segment containing a soma can be assigned a cell type and counted once, which is what makes cell-level statistics possible at all. A volume that clips somata cannot report per-cell degree honestly.
Often confused with: Neuronal soma (large nucleus, prominent nucleolus, abundant rough ER) vs astrocytic soma (~8-10 um, pale watery cytoplasm, glycogen) vs microglial soma (~6-8 um, dark dense cytoplasm) — Unit 07 §1-2.
Spectral embedding
Analysis & statistics · Units 09
Representing nodes as points in a low-dimensional space derived from the adjacency structure.
Typical: Eigenvectors of a graph Laplacian or adjacency matrix, truncated to a few dimensions.
Why it matters: It gives a continuous representation of connectivity that can be clustered or compared, which is useful when discrete community detection is too brittle.
Often confused with: Embedding dimension is a free parameter that changes the structure you see. Choosing it by how good the clusters look is circular.
Spine apparatus
Morphology · Units 05
Stacked smooth endoplasmic reticulum laminae found within a minority of dendritic spines, more common in large spines.
Typical: Stacked laminae ~30 nm apart, in a minority of mostly large spines (Unit 05 §1).
Why it matters: Its presence marks a mature, usually large spine, so it is a useful correlate of synaptic strength when you cannot measure PSD area directly.
Often confused with: Spine apparatus vs ordinary smooth ER in the spine neck. The apparatus is a stack with dense material between laminae, not a single tubule.
Split error
Proofreading & QA · Units 08
One true object incorrectly divided into two or more segments - a neuron broken at a thin neurite, an artifact, or a section boundary.
Typical: The cheap class: obvious, and correctable by joining, which is why proofreading effort concentrates on merges (Unit 08 §1).
Why it matters: Bounded and visible. Pipelines deliberately trade merges for splits at every level.
Often confused with: Splits are visible and merges are not, so a segmentation tuned to minimize the errors you can see is tuned to maximize the ones that corrupt connectivity.
Stitching
Reconstruction · Units 03, 04
Aligning image tiles within one section using their overlap regions.
Typical: 5-15% tile overlap is the working range (Unit 03 §1).
Why it matters: Stitching errors appear on a regular grid, which makes them easy to mistake for real periodic structure and easy to miss entirely if you only inspect within tiles.
Often confused with: Stitching (tiles within a section) vs registration (section to section). A seam artifact is stitching; a shift between adjacent z is registration.
Stopping rule
Proofreading & QA · Units 08
A pre-stated, measurable condition determining when proofreading is complete.
Typical: Qualitative by nature: a stated condition such as 'stop when the headline number moves by less than its error band across a proofreading round'.
Why it matters: The strongest form is convergence-based: stop when further independent proofreading of a sample changes the endpoint metric by less than a stated amount.
Often confused with: A stopping rule is not a budget. A budget says when you run out of time; a stopping rule says when more work stops changing the answer, and only the second is defensible in a paper.
Structure-function inference
Core · Units 01, 09
Reasoning from anatomical connectivity to the computations a circuit might perform.
Typical: Constrained by what structure cannot show: sign, strength, plasticity state and neuromodulation are all invisible in EM.
Why it matters: Structure constrains the space of possible dynamics; it does not determine them. Every such inference should name its extra assumptions.
Often confused with: Constraint (defensible) vs simulation (not supported by anatomy alone).
Subgraph isomorphism
Analysis & statistics · Units 09
The problem of finding occurrences of a query pattern within a larger graph.
Typical: NP-complete in general; tractable here because the patterns sought are tiny.
Why it matters: It is the formal problem behind every motif search, and it is why exhaustive motif counting is expensive on large graphs.
Often confused with: Induced vs non-induced subgraphs. Whether absent edges must also be absent changes the count substantially, and papers often do not say which they used.
Supervoxel
Reconstruction · Units 04, 08
A small, deliberately over-segmented fragment that serves as the immutable atom of the segmentation.
Typical: The atomic unit of the ChunkedGraph: edits change which supervoxels group together, never the supervoxels themselves (Unit 04 §1).
Why it matters: Immutability is what makes concurrent proofreading and stable synapse assignment possible.
Often confused with: Supervoxel vs segment vs root ID. A supervoxel is a fragment, a segment is a group of them, and a root ID names that group at a moment in time. Confusing the three is why version-pinning gets skipped.
Synapse
Core · Units 05
A specialized junction for chemical or electrical communication between neurons.
Typical: Cleft 20-30 nm (asymmetric) or ~15-20 nm (symmetric); PSD 30-50 nm thick and 200-800 nm wide (Unit 05 §1).
Why it matters: In EM, calling a chemical synapse requires all three of: a presynaptic vesicle cluster, a cleft of uniform width, and a postsynaptic density - persisting across sections.
Often confused with: Membrane apposition, adherens junctions, and tangentially cut membranes all mimic synapses; none has a vesicle cluster.
Synapse detection
Reconstruction · Units 04
Automated identification of synapses and assignment of pre- and postsynaptic partners.
Typical: ~5 x 10^8 rows for a mm3 synapse table, 50-200 GB - the hottest analytical table in the stack (Unit 04 §2).
Why it matters: Store partners as supervoxel IDs, never as neuron IDs, so assignments survive proofreading.
Often confused with: Detecting a synapse and assigning its partners are separate problems with separate error rates. A correct cleft with the pre- and postsynaptic sides swapped reverses the edge.
Synapse threshold
Analysis & statistics · Units 09
The minimum number of synapses required to call a connection an edge.
Typical: Commonly 1, 2, or 3.
Why it matters: Not innocent. Because the synapses-per-connection distribution is heavy-tailed, moving from 1 to 3 can remove more than half the edges - and does so unevenly across cell types. Report it and re-run at a second value.
Often confused with: The threshold is a claim about what counts as a connection, and it moves results a long way. Unit 09 treats it as a parameter to report and vary, not a setting to pick once.
Synaptic cleft
Morphology · Units 05
The extracellular gap between pre- and postsynaptic membranes at a chemical synapse.
Typical: ~20-30 nm at asymmetric synapses; ~15-20 nm at symmetric ones.
Why it matters: Uniform cleft width is one of the three criteria for calling a synapse at all. Appositions with variable-width gaps are contacts, not synapses, and treating them as synapses is the commonest way to inflate a connectome's edge count.
Often confused with: A cleft vs the ordinary extracellular space between two apposed membranes. The cleft is uniform in width along its length and often shows faint cross-bridges.
Synaptic vesicle
Morphology · Units 05
A small membrane-bound vesicle in the presynaptic terminal containing neurotransmitter.
Typical: Clear round vesicles ~35-50 nm; dense-core vesicles 80-120 nm.
Why it matters: A vesicle cluster at the apposition is the non-negotiable first criterion for calling a synapse.
Often confused with: Clear round vesicles (35-50 nm, excitatory terminals) vs dense-core vesicles (80-120 nm, peptidergic) vs pleomorphic/flattened vesicles (inhibitory, though the flattening is partly a fixation artifact). Size is measurable and reliable; shape is suggestive and is not.
Triad census
Analysis & statistics · Units 09
Counting each of the 16 isomorphism classes of directed three-node subgraphs.
Typical: 16 possible three-node patterns in a directed graph, counted together.
Why it matters: Sixteen tests, and they are strongly dependent on each other. Use permutation inference and report how many tests you ran.
Often confused with: The census is 16 simultaneous tests, so it needs multiple-comparison correction - and it inherits every sensitivity the underlying edges have to reconstruction error.
Triage
Proofreading & QA · Units 07, 08
Ranking candidate corrections by expected effect on the endpoint per unit of annotator time.
Typical: Ranked by impact on the endpoint metric, not by how wrong an error looks.
Why it matters: Rank by impact, not conspicuousness. A small glia merge often outranks a large obvious split.
Often confused with: Triage by conspicuousness vs triage by impact. Fixing whatever looks worst is the default failure mode, and it systematically overspends on splits.
Ultramicrotome
Sample prep · Units 03
The instrument that cuts resin-embedded tissue into sections tens of nanometers thick using a diamond knife.
Typical: 30-50 nm sections for connectomics.
Why it matters: Section thickness is set here, and it is the parameter that decides your z-resolution and therefore how traceable thin processes are.
Often confused with: Cutting thinner improves z-continuity and multiplies the number of sections, the acquisition time, and the chance of losing one. Unit 03 §1 gives the range as 30-50 nm.
Ultrastructure
Morphology · Units 05
Fine structural detail resolvable only at electron-microscopy resolution.
Typical: The features that matter here span roughly 10 nm (neurofilament) to 1 um (Golgi stack) — Unit 05 §1.
Why it matters: Everything the field calls a 'cue' is an ultrastructural feature, and the whole perceptual skill set in Units 05-07 amounts to reading them reliably and knowing which ones fail.
Often confused with: Ultrastructure (what EM resolves) vs morphology (the shape of the whole cell, resolvable by light microscopy). The distinction is what makes EM necessary for synapses.
Variation of Information (VI)
Proofreading & QA · Units 08
An information-theoretic measure of disagreement between two segmentations, decomposable into split and merge components.
Typical: Reported as a split component and a merge component separately; the sum alone hides which error dominates.
Why it matters: Always report the two components separately. Total VI can improve while merges get worse, because splits usually dominate the sum.
Often confused with: A single summed VI can be identical for a segmentation that is merge-heavy and one that is split-heavy, and those two need completely different proofreading strategies.
Voxel
Imaging · Units 02, 03
A three-dimensional sample of image intensity; the unit of volumetric data.
Typical: Connectomics ssTEM commonly 4 x 4 x 40 nm.
Why it matters: Voxel size sets both what you can resolve and what the volume costs, and the two pull in opposite directions. Every capacity estimate in Unit 03 starts from a voxel count.
Often confused with: Voxel size vs resolution. A 4 nm voxel does not guarantee 4 nm of resolvable detail — staining, dose and optics all cap it below the sampling grid.
neuPrint
Data & infrastructure · Units 04, 09
A graph-database service for querying released connectomes such as hemibrain and MANC.
Typical: Neo4j-backed; queried in Cypher, with neurons as nodes and connections as weighted edges.
Why it matters: It is how the hemibrain and related fly datasets are queried, and its graph-first model is a good fit for motif and pathway questions.
Often confused with: neuPrint (graph queries over a finished connectome) vs CAVE (annotations over an editable segmentation). neuPrint serves snapshots; CAVE serves live state.
rOTO
Sample prep · Units 03
Reduced osmium - thiocarbohydrazide - osmium: an amplified heavy-metal staining protocol.
Typical: Reduced osmium-thiocarbohydrazide-osmium; the standard amplification for en bloc staining of large volumes (Unit 03 §1).
Why it matters: Produces membrane contrast strong enough for fast low-dose imaging, and makes the block conductive enough for block-face SEM.
Often confused with: rOTO amplifies membrane contrast, which is what segmentation depends on. Under-amplified blocks produce weak membranes, and weak membranes produce merge errors - the expensive kind.
webKnossos
Data & infrastructure · Units 04
A web platform for viewing, annotating, and proofreading volume EM data.
Typical: Browser-based, with its own chunked backing store.
Why it matters: It is the annotation and proofreading environment behind several large volumes, and one of the few that handles skeleton tracing and volume proofreading in the same tool.
Often confused with: webKnossos (annotate and proofread) vs Neuroglancer (view). Both open large volumes in a browser; only one is built for making edits.
