MouseConnects Nanoscale Connectomics Workflow

The tissue is fixed by transcardial perfusion with a buffered aldehyde mix — typically around 2–2.5% glutaraldehyde plus 2% paraformaldehyde — to arrest ultrastructure within seconds, before autolysis and osmotic swelling distort the fine processes you are about to spend years tracing. Then, because biological tissue is nearly transparent to electrons, contrast is created chemically: a sequence of heavy-metal stains (the rOTO protocol — reduced osmium, thiocarbohydrazide, a second osmium — followed by uranyl acetate and lead) deposits metal on membranes, which is what you actually see in an EM image.

The hard part at HI-MC scale is uniformity. Stain has to penetrate the entire block, and a block that is well stained at the edges and pale in the center produces segmentation quality that varies systematically with position — which can masquerade as a biological gradient. Weak membrane contrast is the single most expensive prep failure, because it is the dominant cause of automated merge errors: the network cannot find a boundary that is barely there.

One cost is unavoidable and worth knowing now: dehydration and resin embedding shrink tissue on the order of 5–20% linearly. Every absolute length, area, and volume measurement in EM connectomics inherits that distortion, which is why careful papers report ratios within a volume rather than absolute values compared across studies.

Depth: Unit 03: EM prep and imaging — the full preparation chain, with the artifact catalog that pairs each step with its characteristic failure.

An ultramicrotome with a diamond knife cuts the embedded block into sections 30–50 nm thick — around a thousand sections per 40 µm of depth — collected onto tape (ATUM) or a reinforced substrate. HI-MC uses this serial-section approach, building on the Lichtman lab's decades of experience with it, for a decisive reason: sectioning does not consume the block. Sections can be re-imaged if something goes wrong, and imaging can be parallelized across many microscopes, which is how petascale volumes get acquired in finite time.

The price is section handling, and section handling has a signature artifact list: lost sections, folds, wrinkles, knife chatter, compression along the cutting axis, debris. Each is survivable in isolation; what matters is the downstream bill. A lost section is a gap the alignment stage must interpolate across and the segmentation model was never trained on. A fold makes a region of tissue simply untraceable. At 10 mm³ — hundreds of thousands of sections — even a small per-section failure rate is a large absolute number, so the operational discipline is catching problems while the block still exists to recut.

Depth: Unit 03 §1.4 and §2 — sectioning versus block-face approaches, and the artifact catalog with each defect's downstream cost.

The sections are imaged by scanning electron microscopes at nanometer resolution — voxels on the order of 4 × 4 × 40 nm are typical for volumes like this. The arithmetic is the whole story. An 800 µm cube at that resolution is 8 × 10¹⁴ pixels; at a sustained 0.2 gigapixels per second that is 46 days of continuous imaging, and roughly 77 days at a realistic 60% duty cycle — before sectioning, QA, or re-imaging failed sections. Multibeam SEM attacks that throughput term directly, scanning with 61 or 91 beams in parallel to aggregate on the order of a gigapixel per second; it is the technology that moved 1 mm³ from impossible to an eighteen-month project. HI-MC is a 10 mm³ target, which is why sustained multi-instrument operation over years is built into the plan.

The core tradeoff is dose. Image quality improves roughly with the square root of electron dose, so doubling the signal-to-noise ratio costs about four times the acquisition time. At petascale, "just image it better" is rarely the answer; the honest move is usually to accept a noisier image and spend the savings on better segmentation and more proofreading. Failure at this step — drift, charging, defocus that nobody caught — is uniquely expensive because acquisition is the one stage you cannot rerun from disk: the QA has to happen while the instrument is still pointed at the section.

Depth: Unit 03 §1.5 and §3 — imaging parameters and the acquisition QA gates that decide when to stop the microscope.

The microscope produces tiles; science needs one coherent 3D volume. First the raw tiles land in an immutable, checksummed archive — the only irreplaceable asset in the project, since everything downstream is recomputable from it, expensively. Then stitching places tiles within each section, and alignment registers each section to its neighbors. Alignment is the hard half, because sections deform non-rigidly — knife compression, folds, stretch — and because errors accumulate: a bias of 0.1 voxel per section across 20,000 sections is a 2,000-voxel drift. Modern pipelines use coarse-to-fine elastic registration with a global relaxation step that spreads residual error across the whole stack instead of letting it pile up in one direction.

The infrastructure numbers explain why HI-MC is a cloud project. A single 1 mm³ volume at 4 × 4 × 40 nm is about 1.5 × 10¹⁵ voxels: roughly 1.5 PB of raw archive, another ~2 PB for the aligned image pyramid, and comparable transient volumes for the model predictions that feed segmentation. HI-MC is ten of those. At this scale, moving data is often more expensive than storing it — egress on a petabyte can cost more than a year of storage — so compute goes to the data, not the reverse.

Failure here is subtle rather than dramatic: a misalignment does not destroy data, it quietly severs every neurite that crosses the bad seam, and the segmentation stage will faithfully turn that seam into a wall of split errors. And because every stored annotation coordinate is defined in the aligned space, revising an alignment later means re-mapping everything — which is why alignment revisions are rare and carefully planned.

Depth: Unit 04: Volume reconstruction infrastructure — the eight-stage reference pipeline, storage layout, and the capacity-and-cost arithmetic worked in full.

No human traces 10 mm³. Machine learning does the first pass, in two main families. In the affinity approach, a network predicts for every pair of neighboring voxels whether they belong to the same object; watershed then produces deliberately-too-small supervoxels, and an agglomeration step merges them into neurons. In flood-filling networks (FFNs) — developed at Google Research and the planned workhorse for HI-MC, as for FlyWire before it — a network grows one object at a time from a seed, repeatedly asking whether the next voxel belongs. A separate model detects synapses and assigns their partners.

The compute is significant — on the order of 1,700 GPU-days per cubic millimeter at typical throughputs, and pipelines budget for three to five full inference passes because the first model version is never the last — but the deeper design decision is about error. The whole stack is deliberately tuned to over-segment: it prefers splits (one neuron in pieces) to merges (two neurons fused), because a split leaves visible evidence of itself while a merge produces an object that looks like a neuron and is not. Where segmentation fails is structural and predictable: thin spine necks that appear in only one or two sections, processes crossing sections at shallow angles, tightly apposed membranes with weak staining, and artifact regions the model never saw in training.

Depth: Unit 08 §1–2 for how the methods work and fail; Unit 04 for the supervoxel architecture that makes the output editable at all.

Automated segmentation is good and wrong: wrong in ways that are individually small and collectively decisive. Humans correct it using CAVE — the Connectome Annotation Versioning Engine, built for FlyWire and MICrONS and serving as HI-MC's backend — which represents the segmentation as an editable graph over immutable supervoxels, so thousands of proofreaders can split and merge concurrently with every edit versioned. This is where NeuroTrailblazers connects most directly to the project: proofreading and annotation at this scale is a trained workforce problem, and trained contributors can work on real data.

It is also the pipeline's dominant cost. Compute and storage are line items you can negotiate with a cloud vendor; proofreading is a hiring, training, and quality-management problem, at a few hours of skilled attention per fully proofread neuron. And complete manual proofreading of 10 mm³ is almost certainly infeasible — the case study is explicit about this — so HI-MC will need strategies that combine automated error detection with targeted human review of the circuits that matter most. That turns proofreading into an allocation problem under a fixed budget: which errors, on which cells, checked to what defined standard, with a stopping rule written down in advance.

Failure at this step is the quietest in the pipeline, because its output looks identical to success. An unproofread merge does not crash anything; it just inflates a connectivity count in a paper two years later. This is why quality is reported per release, with metrics whose blind spots are stated alongside them.

Depth: Unit 08 for the production loop, triage, and metrics; the content library's error taxonomy, strategies, and metrics and QA entries for the reference material; and the neuroanatomy entries for the identification skills the work depends on.

The deliverable is not the images — it is the graph, plus the tables that describe it: a synapse table on the order of 5 × 10⁸ rows per cubic millimeter, skeletons, meshes, and cell annotations. Building the graph is itself a sequence of consequential choices (which edges count, what synapse threshold, which release), and every analysis needs a null model before a claim: a motif count means nothing until you say what you are comparing it against. For HI-MC the target questions are the hippocampal classics — whether dentate gyrus wiring supports pattern separation, whether the CA3 recurrent network looks like an auto-associative memory, how the trisynaptic circuit is actually built at synaptic resolution.

Failure at this step is the only kind the earlier pipeline cannot cause: a technically perfect reconstruction analyzed without regard to its error profile. Residual splits deflate degree; residual merges inflate it; both do so non-uniformly. The defensible habit is to run analyses at two proofreading versions and report what moved — which is also why the project's phased public releases, with versioned access for the community, are part of the pipeline rather than an afterthought.

Depth: Unit 09: Connectome analysis and NeuroAI — graph construction, null models, motif analysis, and the error-sensitivity checks that keep a result honest.