The tissue-to-image-stack chain in practical detail: fixation and staining chemistry, sectioning, imaging parameters, and an artifact catalog mapped to downstream reconstruction cost.
Want to see this in real data? EM figure library collects every micrograph on this site, and H01, Step by Step carries real electron micrographs of human cortex — rendered straight from the public H01 volume — from whole-sample down to individual synapses, alongside the pipeline that produced them.
This page itself is still text-only on the diagram side. For equipment schematics and further depth, see:
| Time | ~2 h, plus a 90 min lab |
| Prerequisites | Units 01–02 |
| You need | Access to any public EM volume in Neuroglancer (MICrONS, FlyWire, or H01 all work) |
| You finish with | A completed acquisition QA report on a real volume, with artifacts localized and costed |
The governing fact of this unit: acquisition quality sets a ceiling on reconstruction quality that no amount of downstream machine learning or proofreading labor can raise. A fold destroys the tissue. A missing section destroys the continuity. You can annotate around damage, but you cannot recover what was never imaged. This is why acquisition QA is not a formality — it is the highest-leverage QC in the entire pipeline, and it is the one most often deferred.
Every step exists to solve a specific problem, and each one introduces a characteristic failure. Learn the pairs.
What it does. Cross-links proteins to arrest ultrastructure within seconds, before autolysis and osmotic swelling destroy the extracellular space and the fine processes.
Typical protocol. Transcardial perfusion in rodents with a buffered aldehyde mix — commonly around 2–2.5% glutaraldehyde plus 2% paraformaldehyde in 0.1 M cacodylate or phosphate buffer, at physiological pH, often with added calcium. Glutaraldehyde is the workhorse because it is bifunctional and cross-links rapidly; paraformaldehyde penetrates faster and buys time.
What failure looks like in the final image.
Teaching point for annotators. When you see a region of unusually open neuropil, do not treat it as a segmentation opportunity. Flag it. It is a region where your geometric measurements — spine neck diameter, apposition area, extracellular fraction — are not comparable to the rest of the volume.
Biological tissue is nearly transparent to electrons. Contrast comes entirely from heavy metals bound to specific structures. Membranes are what matters for connectomics, so the protocol is optimized for lipid.
The standard sequence:
The combination is usually called rOTO (reduced osmium – thiocarbohydrazide – osmium). Two reasons it dominates volume EM: it produces membrane contrast strong enough to image quickly at low dose, and the metal load makes the block electrically conductive, which is what makes block-face SEM possible at all without catastrophic charging.
What failure looks like:
What it does. Water is replaced by solvent (graded ethanol or acetone), then by epoxy resin (Epon/Araldite, LX-112, Durcupan, Spurr’s), which is polymerized to a solid block that can be cut at tens of nanometers.
The unavoidable cost: dehydration shrinks tissue, typically on the order of 5–20% linearly depending on protocol. This is systematic, not random. Every absolute length, area, and volume measurement in EM connectomics is affected. Report measurements as measured, state the protocol, and prefer ratios and comparisons within a volume over absolute values compared across studies.
What failure looks like: cracks and tears (usually from too-rapid dehydration or incompletely infiltrated resin), and resin that is too soft or too brittle to section cleanly, which shows up at the next step.
Two families, with different artifact profiles.
Serial sectioning (for ssTEM / ssSEM). An ultramicrotome with a diamond knife cuts 30–50 nm sections, which are collected onto grids, tape (ATUM), or a reinforced substrate (GridTape). Sections are then imaged, in some cases by many microscopes in parallel.
Block-face (SBEM). Image the block face, then shave off a slice with a diamond knife inside the chamber, repeat. FIB-SEM substitutes an ion beam that mills a few nanometers at a time, giving isotropic voxels.
The parameters you will actually be asked about:
| Parameter | Typical range | Increase it and… | Decrease it and… |
|---|---|---|---|
| Landing energy (SEM) | 1–2 keV | More depth signal, more charging, more beam damage | Better surface specificity, weaker signal |
| Dwell time per pixel | 0.1–2 µs | Better SNR | Faster acquisition, noisier images |
| Beam current | pA–nA | Better SNR at fixed dwell | Less damage and charging |
| Tile overlap | 5–15% | More robust stitching | Less redundant data, faster |
| Section thickness (z) | 30–50 nm | Fewer sections, faster, cheaper | Better z-continuity, more data |
Dose is a budget. SNR improves roughly with the square root of electron dose, and dose is the product of beam current and dwell time. Doubling SNR costs roughly 4× the acquisition time. This is why “just image it better” is rarely the answer at petascale — the honest tradeoff is usually to accept a noisier image and spend the savings on better segmentation and more proofreading.
Multibeam SEM attacks the throughput term directly: 61 or 91 electron beams scanning in parallel, aggregating on the order of a gigapixel per second. That is the technology that moved 1 mm³ from “impossible” to “an eighteen-month project”.
A volume is 800 µm × 800 µm × 800 µm at 4 × 4 × 40 nm. Your instrument sustains 0.2 gigapixels per second including overheads. How long?
voxels_xy per section = (800,000 / 4)^2 = 200,000^2 = 4.0 x 10^10 px
sections = 800,000 / 40 = 20,000
total pixels = 4.0e10 x 2.0e4 = 8.0 x 10^14 px
time = 8.0e14 / 2.0e8 px/s = 4.0 x 10^6 s ~= 46 days of continuous imaging
Then multiply by your real duty cycle. At 60% uptime this is ~77 days; and this counts only imaging, not sectioning, not QA, not re-imaging failed sections. When someone says a 1 mm³ volume takes “about a year”, this is the arithmetic behind it.
Staining penetration (§1.2). The reagents — most likely osmium, TCH, or lead — did not reach the block interior. The tell is that the gradient follows block geometry, not tissue anatomy or acquisition order.
Fixes, in order of practicality: cut smaller blocks (the standard answer — penetration depth is the constraint, so reduce the distance); extend incubation times; use microwave-assisted processing; check reagent freshness, particularly TCH.
Diagnostic contrast: if the washed-out region followed acquisition order rather than block position, you would suspect beam or detector drift instead. If it followed anatomy (e.g. only white matter), you would suspect a genuine tissue-composition effect. Always ask which coordinate system the defect lives in — that identifies the stage that produced it.
This is the reference table to keep open while doing QA. The right-hand columns are what turn “the data looks bad” into a decision.
| Artifact | How to recognize it | Root cause | Downstream effect | Cost class |
|---|---|---|---|---|
| Lost section | A z-gap; structures discontinuous across one z index, volume-wide | Section lost during collection | Every process crossing that z must be bridged by inference | Data loss — unrecoverable |
| Fold | Dark band with duplicated/compressed tissue, usually linear | Section wrinkled on collection | Tissue in the fold is unusable; segmentation splits along it | Data loss in the folded strip |
| Tear / crack | Sharp-edged gap, often following a vessel | Dehydration or sectioning stress | Local loss; false boundaries at edges | Data loss, localized |
| Knife chatter | Periodic bands, fixed spacing, perpendicular to cutting direction | Knife or block vibration | Adds false boundaries; raises split rate | Labor — proofreadable |
| Compression | Section shorter along cutting axis than expected | Knife compresses the section | Systematic geometric distortion; must be corrected in alignment | Labor, plus measurement bias |
| Charging | Bright streaks or smears trailing the scan direction; local distortion | Non-conductive surface accumulating electrons | Model confidence collapses locally; splits | Labor |
| Curtaining (FIB-SEM) | Vertical stripes parallel to milling direction | Uneven milling rate | Texture noise; degrades boundary detection | Labor |
| Weak membrane contrast | Faint or interrupted membrane outlines | Understaining | Merge errors — the expensive kind | Labor, high |
| Precipitate | Small very dark irregular particles | Staining chemistry | False synapse detections; minor false boundaries | Labor, low |
| Contamination / debris | Foreign objects, often out of focus | Handling, chamber contamination | Local obstruction | Labor, low |
| Beam damage | Bubbling, mass loss, progressive contrast change | Excess dose | Degradation that worsens with re-imaging | Data loss if severe |
| Misalignment / drift | Structures shift between adjacent sections | Stitching or registration failure | False branch points, synapse mislocalization | Labor, correctable by re-alignment |
| Seam visibility | Intensity step at tile boundaries | Stitching / illumination correction failure | Boundary artifacts along a regular grid | Labor, correctable |
The cost distinction matters. A labor artifact means your reconstruction will be correct, eventually, after paying for it in proofreading hours or better algorithms. A data loss artifact means some biological question is unanswerable in that region, permanently. Report them separately. A QA report that gives one quality score conceals exactly the distinction the project needs.
Merge errors are worse than split errors, and understaining causes merges.
Consequence for acquisition: when trading dose against speed, protect membrane contrast. Noise that raises the split rate is recoverable. Faint membranes that raise the merge rate are much less so.
Roughly: (b) > (d) > (c) > (a), though the exact order depends on your endpoint.
(b) 4 consecutive lost sections = a 160 nm gap. Most thin neurites cannot be reliably bridged across that; the volume is effectively cut into two independently reconstructable halves at that z. This is a structural break in the dataset and it must be reported prominently, because any claim about a process crossing that plane is now inference rather than observation.
(d) Weaker contrast throughout raises the merge rate everywhere. Volume-wide, invisible in summaries, expensive. It also cannot be fixed by re-imaging, because the metal simply is not in the tissue.
(c) Charging on 15% of sections is bad but bounded and localized; it mainly raises splits, and split-heavy regions can be prioritized in the proofreading queue. Some charging is also correctable by adjusting imaging conditions for the remaining sections, so catching it early has real value.
(a) 4 scattered lost sections is normal operating loss. Each is a single-section gap that alignment and segmentation routinely bridge for all but the thinnest processes. Log it; do not panic.
The transferable lesson: distribution matters more than count. Four scattered losses and four consecutive losses have the same headline number and completely different consequences. Never report artifact rates without their spatial distribution.
Run a pilot reconstruction before full acquisition. Take a small sub-volume — something on the order of 100 × 100 × 100 µm — through the entire pipeline: align, segment, skeletonize, and have a human proofread a handful of neurons. Measure the error rate.
This costs perhaps 1–2% of the project and it is the only way to find out that your staining protocol produces a merge rate the segmentation cannot handle, while you can still change the staining protocol. Teams that skip this step discover the problem after acquiring a petabyte.
Per section and per tile, not just per volume:
Define these in advance, in writing, with numbers:
| Gate | Example threshold | Action if breached |
|---|---|---|
| Consecutive lost sections | > 2 | Stop; investigate collection before continuing |
| Cumulative lost-section rate | > 1% | Review handling protocol |
| Membrane CNR drop vs baseline | > 20% | Stop; check staining batch and beam conditions |
| Fold area fraction per section | > 5% | Flag section; re-cut if the block allows |
| Alignment residual, 99th percentile | > 1 voxel at native xy | Re-run alignment before ingesting |
| Pilot segmentation merge rate | above the level your proofreading budget can absorb | Do not scale; revisit prep |
The specific numbers are yours to set — they depend on your endpoint and your budget. What is not optional is setting them before you start, because a threshold chosen after seeing the data is not a threshold.
Every derived product must be traceable to the acquisition conditions that produced it. Minimum machine-readable record:
Why the tile-level timestamp matters more than it sounds. When you later find a quality anomaly, the first diagnostic question is always “does this defect follow block position, anatomy, or acquisition time?” Time-correlated defects point to instrument drift or reagent degradation; position-correlated defects point to penetration or geometry. Without per-tile timestamps you cannot ask the question.
These are context slides rather than QA specimens; the artifact catalog in §2 is what you take to a real volume. Use each panel to rehearse the diagnostic question that runs through this unit — which coordinate system does a defect live in: block position, anatomy, or acquisition time?
Module12 L3 S04: High-resolution imaging. Tie it to the dose budget in §1.5: SNR improves only with the square root of dose, so doubling it costs roughly four times the acquisition time. Ask what was traded for image quality here, and hold to the standing rule — protect membrane contrast, because faint membranes cause merges.
Module12 L3 S08: High-throughput sectioning. Section handling is where lost sections, folds, wrinkles, and knife chatter originate (§1.4). Ask which of those the depicted approach is exposed to, then sort each one into the data-loss or the labor column of the artifact table.
Module12 L3 S10: The handoff from imaging to reconstruction. This is the boundary past which acquisition quality becomes a ceiling nothing downstream can raise. Check what metadata crosses it — per-tile timestamps and machine-readable defect masks are what let you diagnose an anomaly months later (§3–§4).
Module13 L2 S08: Manual work set against automated work. Use it to locate the pilot-reconstruction rule in §3: a small sub-volume taken all the way through segmentation and human proofreading is what tells you whether your staining produces a merge rate the proofreading budget can absorb — while you can still change the staining.
Attribution: assets_outreach source decks (historical/context visuals).
Setup. Open any public volume in Neuroglancer — MICrONS, FlyWire/FAFB, or H01. Do not use a curated tutorial view; navigate to arbitrary coordinates.
If you want to see what a whole preparation-and-acquisition chain looks like end to end before you start, H01, Step by Step walks one real dataset through every stage in this unit — fixation, ROTO staining, ATUM sectioning, 61-beam imaging, and the QC that ran during acquisition — with the resulting micrographs at each scale. It also shows how to pull image data out of a public volume yourself, which is a faster route to the sampling step below than clicking through the viewer.
Task. Produce a QA report.
| Not yet | Proficient | Strong | |
|---|---|---|---|
| Sampling | Convenience locations | Stated sampling rule, followed | Rule justified, and coverage bias acknowledged |
| Identification | “Image looks noisy” | Artifacts named from the catalog | Root cause proposed, with the coordinate-system reasoning that supports it |
| Cost classification | Absent | Data-loss vs labor assigned | Distribution considered (scattered vs clustered), not just count |
| Localization | Prose only | Coordinates given | Machine-readable defect list a pipeline could consume |
| Impact | Not attempted | Qualitative | Quantified with stated assumptions |
| Recommendations | Generic | Tied to observations | Tied to observations and costed against the tradeoff triangle |
Instructor note. Run step 2 as a calibration exercise first: have everyone score the same three locations, then compare scores publicly before proceeding. Inter-rater spread on “membrane contrast 1–5” is typically large on the first attempt and shrinks sharply after one round of discussion. That shrinkage is the learning, and it is also a live demonstration of why annotation protocols need calibration sessions (Unit 05).
Deferring QA until acquisition finishes. Recover: pilot reconstruction, always, before scaling.
Reporting a single global quality number. Recover: report per-region and per-section distributions, and always separate data loss from labor.
Chasing SNR instead of contrast. A noisy image with crisp membranes segments better than a clean image with faint ones. Recover: measure membrane CNR specifically, not overall image SNR.
Losing the acquisition log. Recover: treat metadata capture as a pipeline stage with its own tests. If the log is not machine-readable, it does not exist.
Assuming nominal section thickness. Recover: measure it empirically from traced structures and use the measured value in all geometry.
Some of what this unit teaches is technique. Some of it is professional norm — the things experienced people do without being asked, and which nobody states out loud because they assume you already know. Those are worth naming, because they are distributed unequally by background rather than by ability.
From this unit:
Separate data loss from labor when you report quality. One quality score conceals the only distinction the project actually needs. This is the difference between “expensive to fix” and “unanswerable forever”.
Report the distribution, not the count. Four scattered lost sections and four consecutive lost sections have the same headline number and completely different consequences.
Pilot before you scale. Running a small sub-volume through the whole pipeline costs 1–2% of a project. Skipping it is the most expensive habit in the field, and no one is ever told to do it explicitly.
The collected set, and why making these explicit is a fairness intervention rather than etiquette, is in the hidden curriculum.
The alignment and segmentation algorithms that consume this data (Units 04 and 08), and how to read the biology in a well-prepared image (Units 05–07). It also does not cover cryo-EM, correlative light-EM workflows in depth, or non-EM volumetric methods (Unit 02).
This unit is anchored to canonical papers and datasets used in connectomics practice. Use these as required preparation before activities.
Capability target: Diagnose EM preparation and imaging artifacts and apply QA gates before reconstruction.
Artifact Triage Lab - Classify artifacts and decide whether blocks pass acquisition QA. The unit's own lab above is the graded version of this exercise; do that one.
Identify acquisition artifacts and define practical QA gates before reconstruction.