How to choose a modality, resolution, and representation for a connectomics question, and how to link measurements across scales without over-claiming.
Key community resources for this unit:
| Time | ~90 min, plus a 75 min lab |
| Prerequisites | Unit 01 — specifically the voxel-count arithmetic in §2 |
| You need | A calculator. Optional: a Neuroglancer link to any public volume |
| You finish with | A scale-selection memo defending one modality choice against two alternatives |
Almost every practical connectomics decision is a scale decision in disguise. “Should we use SBEM or ssTEM?” is a scale decision. “Can we answer this with the existing public dataset?” is a scale decision. “Why doesn’t our tractography result match the EM result?” is a scale decision that was made badly, earlier.
Practitioners routinely conflate three different scales. Separating them prevents most downstream confusion.
Acquisition scale — the voxel size and field of view your instrument produces. Set by physics and budget. Fixed once the data exists.
Reconstruction scale — the smallest object you can reliably segment. This is always coarser than acquisition scale, sometimes much coarser. At 4 × 4 × 40 nm you acquire enough signal to see a 20 nm cleft, but the objects you can segment reliably are neurites down to roughly 50–100 nm, and the reliability drops sharply for processes running steeply through z.
Analysis scale — the unit your conclusions are about: a synapse, a cell, a cell type, a layer, an area, a projection.
The decision rule. Choose the coarsest acquisition scale whose reconstruction scale still resolves every object your analysis scale depends on. Not the finest you can afford — the coarsest that works. Every step finer multiplies data volume, alignment difficulty, and proofreading hours.
A worked instance of the rule: if your endpoint is “fraction of inputs onto spines vs shafts”, your analysis scale is the synapse and the spine neck. Spine necks are 50–200 nm. So your reconstruction scale must be ≤ 50 nm, which forces EM. If instead your endpoint is “does area A project to area B at all”, your analysis scale is the axon bundle, and light-sheet imaging of a bulk tracer at 1 µm is not merely adequate — it is the correct choice, because it costs about five orders of magnitude less.
Learn this table. Ranges are typical rather than record-setting.
| Modality | Resolution (typical) | Practical volume | What it uniquely gives you | What it cannot do |
|---|---|---|---|---|
| Diffusion MRI tractography | 0.5–2 mm | Whole human brain, in vivo | Whole-brain, living, repeatable, human | Cannot see individual axons; produces false and missing tracts; no synapses; no direction |
| Light-sheet / whole-brain LM with tracers | 0.5–2 µm (xy) | Whole mouse brain | Long-range projection maps, many animals | Cannot resolve neurites in neuropil; no synapses |
| Confocal / two-photon | 200–300 nm lateral | mm³, in vivo possible | Function (calcium), molecular labels, live | Diffraction-limited; overlap ≠ connection |
| Expansion microscopy | ~25–70 nm effective | Up to ~mm³ with effort | Molecular identity plus near-EM geometry | Expansion distortion; not yet routine for dense reconstruction at scale |
| Array tomography | ~50–100 nm lateral, 70 nm sections | ~10⁵ µm³ | Multiplexed protein labeling with synapse-scale geometry | Section loss; lower z-resolution than EM |
| Barcoded projection mapping (MAPseq/BARseq) | Single-cell identity, no geometry | Whole brain, 10⁴–10⁶ cells | Projection patterns of enormous numbers of individual cells, cheaply | No synapses, no morphology, no local circuit |
| ssTEM / ssSEM (multibeam) | 4 × 4 × 40 nm | Up to ~1 mm³ today | Dense synapse-resolution reconstruction at scale | Anisotropic; section artifacts; enormous cost |
| SBEM | 10–20 × 10–20 × 25–50 nm | ~10⁶–10⁷ µm³ | Automated block-face series, no section handling | Destructive; z-resolution limits thin-process tracing |
| FIB-SEM | 4–8 nm isotropic | ~10⁵–10⁶ µm³ (larger with hot-knife partitioning) | Isotropic — the best tracing conditions available | Slow; limited volume per run |
Resolution, volume, and throughput form a budget you cannot escape: you may choose two. FIB-SEM buys resolution and gives up volume. Light-sheet buys volume and throughput and gives up resolution. Multibeam ssSEM is the current attempt to buy resolution and volume by throwing an unreasonable amount of throughput engineering at the problem — 61 or 91 beams in parallel — and it is why 1 mm³ became feasible.
Barcoded projection mapping (MAPseq/BARseq). The question is about single-cell projection classes across thousands of cells — it needs statistical power over cells, not geometry within a cell. MAPseq delivers exactly that at a cost per cell that is orders of magnitude below EM.
EM is the wrong tool here for two reasons: (1) a 1 mm³ EM volume does not contain both AL and PM plus the V1 somata, so you would need a much larger volume than is currently feasible; (2) even if it did, the answer needs n in the thousands, and EM proofreading of thousands of complete long-range axons is not tractable today.
The honest limitation to state: barcoding gives you projection presence in a target region, not synapses in that region, and it is vulnerable to fibers-of-passage and to barcode-sharing artifacts.
They are measuring different things at scales that do not nest cleanly.
Tractography infers streamlines from voxel-scale water diffusion orientation. A streamline is a model output, not an observed axon. Known failure modes include false continuations at crossing-fiber regions and difficulty entering gyral crowns. A tractography “connection” is a statement about the most probable pathway given a diffusion model — it is not a claim that a specific axon terminates there.
Meanwhile, an EM volume samples a small region. Absence of axons in that volume is evidence about that volume, at whatever detection sensitivity the reconstruction achieved. If the projection terminates in an adjacent layer or a neighboring column, EM would miss it.
The resolution is not “one of them is wrong”. It is that neither claim, as usually written, states its sampling and its inference model clearly enough to be compared. This is scale leakage in both directions.
After reconstruction, the same neuron exists in four representations. Choosing wrongly is a common and expensive mistake because conversions are lossy and usually one-way.
| Representation | What it is | Keeps | Discards | Typical size, one cortical neuron |
|---|---|---|---|---|
| Volume / labeled voxels | The segmentation itself | Everything, including membrane geometry and organelles | Nothing — but unusable for most analysis | GB per neuron at native resolution |
| Mesh | Triangulated surface | Surface geometry, volume, surface area, spine shape | Interior, intensity, organelles | 10–100 MB |
| Skeleton | Centreline graph with radii | Topology, path length, branch structure, radius | Surface detail, spine head shape, membrane apposition | 0.1–5 MB |
| Connectivity graph | Nodes = cells, edges = synapse counts | Who connects to whom, how strongly | All geometry — compartment, distance, position | Bytes per neuron |
The rule that saves projects: decide which representation your endpoint metric requires before the pipeline runs, and keep the next-richer one archived.
Concrete example of getting this wrong: a team exports a connectivity graph, runs motif analysis, and finds an enrichment. A reviewer asks whether the enrichment is explained by spatial proximity — dendrites that are close connect more. The graph has no geometry, so the question cannot be answered without re-running from skeletons. Archiving skeletons alongside the graph costs a few gigabytes and prevents this entirely.
Question as posed: “On layer 2/3 pyramidal cells, do putatively inhibitory (symmetric) synapses sit closer to the soma, in path distance, than putatively excitatory (asymmetric) ones?” Planned scope: ~300 proofread cells from an existing public volume.
Start from the endpoint, not from what the pipeline happens to emit. The measurement is a distribution of path distance from soma, per synapse, per synapse class. Now hold each representation against that requirement.
Graph: discards all geometry, including the one quantity the question is about. Rejected in one line.
Volume: contains everything, at GB per neuron — order 1 TB of hot label data for 300 cells — and computing path distance from voxels means extracting a centerline anyway. You would be building skeletons expensively, on demand, forever.
Mesh: 10–100 MB per neuron, so 3–30 GB for the study. Surfaces are the right tool for spine shape, which this endpoint does not need, and path distance along the arbor is not a natural mesh operation.
Skeleton plus synapse coordinates: 0.1–5 MB per neuron — 30 MB to 1.5 GB for all 300 cells — and path distance from soma is the native query on a centerline graph. This is exactly the pairing the list above prescribes for “where on the dendrite do those synapses land?”
So far the decision looks free. Here is where it is not. Skeletons discard spine geometry, so a synapse on a spine head gets mapped to the nearest shaft node, and its path distance loses the neck. Asymmetric synapses land mostly on spines; symmetric ones mostly on shafts and somata (Unit 05). The shortening therefore falls mostly on one of the two classes being compared — a differential bias, the dangerous kind. How large is it? The skeleton cannot tell you; it discarded precisely that quantity. Probably small against distance bins tens of micrometers wide — but “probably” is not a number you can print.
The resolution is the archiving rule from this section: run the analysis on skeletons, and archive the meshes — 3–30 GB, noise next to the roughly 2 PB aligned pyramid the volume already carries. On a 20-cell subsample, measure the spine-neck offset on the mesh, and either correct the distances or report the bound.
Transferable principle: choose the representation whose native operation is your endpoint metric; reject the coarser candidates by naming the discarded quantity that disqualifies them; and archive the next-richer representation specifically so you can bound the bias the one you chose introduces.
Linking scales — EM to two-photon function, EM to a reference atlas, one animal to another — is registration, and registration is where confident-looking errors are manufactured.
Anisotropy warning. On a 4 × 4 × 40 nm stack, an isotropic Gaussian smoothing kernel, an isotropic distance metric, or an isotropic morphological operation is silently wrong by a factor of 10 in z. Check every library call for whether it takes voxel spacing. This bug is common, quiet, and it biases every distance-based measurement you make.
Not globally, and not yet. The max tells you there is a region where the transform fails. Steps, in order:
Storage for the raw volume is the smallest line item. Plan for the multiplier.
For a 1 mm³ ssTEM volume (~1.6 PB raw), typical derived footprints:
The operational lesson: the raw volume is written once and read rarely. The synapse table and the segmentation are read constantly, by many people, with latency expectations measured in milliseconds. Budget for query load, not just capacity. A petabyte in cold object storage is cheap; a 200 GB table that must answer 50 concurrent interactive queries is the part that needs engineering. Unit 04 covers the architecture that follows from this.
Use this panel to rehearse the three-scale separation from §1. For each slide, say which scale it is actually about — acquisition, reconstruction, or analysis — because pipeline diagrams routinely blur all three, and noticing the blur is the skill this unit is teaching.
Module12 L1 S06: Voxel size and scale. Whatever resolution the slide names, run it through the decision rule in §1: reconstruction scale is always coarser than acquisition scale, so ask what the smallest reliably segmentable object would be rather than what the voxel size implies.
Module12 L2 S04: A macroscale pipeline. Place it on the modality chart in §2 and name which two corners of the resolution/volume/throughput triangle it buys — and therefore which one it gives up. Nothing at macroscale resolves neurites in neuropil, so connectivity read off it is model output, not observation.
Module12 L2 S05: The bridge from macroscale to microscale. This is exactly where scale leakage is manufactured: check what claim is being carried across the join, and whether the transfer comes with a stated registration residual and its maximum, not just its mean (§4).
Module12 L3 S08: High-throughput imaging. Ask what representation the output eventually becomes — volume, mesh, skeleton, or graph (§3) — because that decision is effectively made here, and every conversion downstream is lossy and one-way.
Attribution: assets_outreach source decks (historical/context visuals).
Scenario. Your lab wants to test whether inhibitory interneurons in mouse visual cortex preferentially target the axon initial segment (AIS) of pyramidal neurons that share functional tuning, versus pyramidal neurons that do not.
Write a two-page memo that selects a data strategy. Required sections:
| Not yet | Proficient | Strong | |
|---|---|---|---|
| Decomposition | Question restated | All required objects listed with sizes | Includes the non-anatomical requirement (functional tuning) and notes it forces in-vivo imaging before fixation |
| Resolution defense | Asserts EM | States required scale per object | States what specifically fails one step coarser, per object |
| Rejections | Alternatives listed | Rejected with a reason | Rejected with the specific unresolvable structure named |
| Registration | Mentioned | Anchors and transform named | Residual target, holdout plan, and consequence of failure quantified |
| Budget | Absent | Raw volume computed | Derived products and proofreading labor included; labor identified as dominant |
| Bottleneck | Generic (“funding”) | A real pipeline step | A step with a stated failure probability and a concrete mitigation |
Objects the endpoint depends on:
The point of the exercise: step 4 changes the entire project structure and is the thing most often discovered too late.
Scale leakage. A mechanistic claim resting on data that cannot resolve the mechanism. Recovery: write the sentence, then write underneath it the smallest object the sentence depends on, then check that object against your reconstruction scale.
Trusting a global registration metric. Recovery: always produce a residual map. If your tooling does not produce one, produce it manually from held-out anchors.
One-way representation conversion. Recovery: archive the next-richer representation. Skeletons alongside graphs; meshes alongside skeletons if storage allows.
Isotropic operations on anisotropic data. Recovery: audit every distance, smoothing, and morphology call for voxel-spacing arguments. Add a unit test that computes a known distance along z.
Assuming public data fits your question. Recovery: before proposing new acquisition, check whether MICrONS, FlyWire, H01, or a BossDB volume already contains the objects you need at the scale you need. Frequently one does, and the honest answer to “what modality?” is “none — this is a re-analysis”.
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:
Report registration residuals locally, with the maximum — never as a global mean. A mean residual of 3 µm can hide a 40 µm failure in one corner. A reader who knows the field looks for the distribution first and distrusts a paper that reports only the mean.
Archive the next-richer representation. Keeping skeletons alongside a connectivity graph costs a few gigabytes and answers the reviewer question you cannot otherwise answer without re-running the pipeline.
The collected set, and why making these explicit is a fairness intervention rather than etiquette, is in the hidden curriculum.
The physics of image formation and the artifact catalog (Unit 03), the systems architecture that stores and serves these representations (Unit 04), and the statistics that operate on the resulting graph (Unit 09).
This unit is anchored to canonical papers and datasets used in connectomics practice. Use these as required preparation before activities.
Capability target: Select an imaging and analysis scale that matches the biological question and resource constraints.
Multiscale Design Exercise - Pick the smallest sufficient dataset design for a defined question. The unit's own lab above is the graded version of this exercise; do that one.
Choose imaging and analysis scale that can resolve required features at manageable cost.