EM Prep and Imaging

Walk the chain with its failure pairs from unit §1.
Fixation: slow or delayed fixation shows as swollen astrocytic processes and enlarged extracellular space — it can make segmentation easier and still distorts every geometric measurement.
Staining (rOTO: reduced osmium, thiocarbohydrazide, osmium, then uranyl acetate and lead aspartate): weak membrane contrast is the dominant cause of automated merge errors — the most expensive prep failure.
Embedding: dehydration shrinks tissue roughly 5–20% linearly; every absolute length is affected, so prefer within-volume ratios.
Sectioning: lost sections, folds, knife chatter, compression. Block-face: charging; FIB-SEM curtaining.
Imaging: SNR improves with the square root of dose — doubling SNR costs about 4x the acquisition time.

Instructor script: "Everything on this plate is visible with a light microscope and sparse staining. Nothing on it tells you who connects to whom. The prep chain we are about to walk through exists to get from this to membranes 20 nm apart."
Then ask where, in the chain on the previous slide, errors become irreversible. Target answer: fixation and staining — contrast that is not in the tissue cannot be recovered by re-imaging.

Instructor script: "Panel a is automated segmentation, panel b is manual annotation of the same field. The difference between them is the proofreading bill your prep protocol will generate."
Tie to the non-negotiable rule in unit §3: run a pilot reconstruction on something like 100 x 100 x 100 µm before full acquisition — align, segment, skeletonize, and have a human proofread a handful of neurons. It costs perhaps 1–2% of the project and is the only way to discover a bad staining protocol while you can still change it.

Instructor script: "This is one of 5,019 sections. Imaging all of them took 326 days on a 61-beam microscope and produced 1.8 petabytes." Note the H01 case study's warning: 1.8 PB is the raw acquisition and 1.4 PB the aligned volume — the two get conflated constantly, so say which you mean.
Work the unit's acquisition-time arithmetic alongside: 800 µm cube at 4 x 4 x 40 nm is 8 x 10^14 px; at 0.2 gigapixels per second that is about 46 days of continuous imaging, about 77 at 60% uptime — before sectioning, QA or re-imaging.

Instructor script: "Every membrane you see here was made visible by osmium. If the staining had been ten percent weaker, which of these boundaries would a network still find?"
This is the slide to land the asymmetry: noise that raises the split rate is recoverable; faint membranes that raise the merge rate are much less so. When trading dose against speed, protect membrane contrast.

The cost distinction from unit §2 matters more than the taxonomy: a labor artifact means the reconstruction will be correct eventually, after proofreading hours; a data-loss artifact (lost section, fold, tear, severe beam damage) makes some question permanently unanswerable in that region. Report them separately — one quality score conceals exactly the distinction the project needs.
Diagnostic habit: ask which coordinate system the defect lives in. Follows block geometry -> staining penetration. Follows acquisition order -> beam or detector drift. Follows anatomy -> possibly real tissue composition.

Example gates from the unit, stated in advance with numbers: stop if more than 2 consecutive sections are lost; review handling if cumulative loss exceeds 1%; stop if membrane contrast-to-noise drops more than 20% from baseline; flag sections with more than 5% fold area; re-run alignment if the 99th-percentile residual exceeds one voxel at native xy.
Instructor script: "The numbers are yours to set. What is not optional is setting them before you start, because a threshold chosen after seeing the data is not a threshold."

Use the unit's triage check-yourself on a 20,000-section volume: rank (a) 4 scattered lost sections, (b) 4 consecutive lost sections, (c) charging on 15% of sections, (d) 10% weaker membrane contrast throughout. Answer, roughly b > d > c > a. Four consecutive losses is a 160 nm gap that cuts the volume in two; weaker contrast raises merges everywhere and cannot be fixed by re-imaging. The lesson: distribution matters more than count.

On the third: per-tile timestamps are what let you ask whether a defect follows block position, anatomy or acquisition time. Without them you cannot ask the question at all, and they cannot be recovered later.