Instructor script: "'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 match the EM?' is a scale decision that was made badly, earlier."
Separate the three scales practitioners conflate. Acquisition scale: the voxel size your instrument produces. Reconstruction scale: the smallest object you can reliably segment — always coarser; at 4 x 4 x 40 nm you can see a 20 nm cleft but reliably segment neurites down to roughly 50–100 nm. Analysis scale: the unit your conclusions are about. Decision rule: choose the coarsest acquisition scale whose reconstruction scale still resolves every object your analysis depends on. Not the finest you can afford — the coarsest that works.
Instructor script: "Each panel is roughly four times closer than the last, and it is the same piece of human temporal cortex throughout. At each step, ask what is visible here — and what is fundamentally unobservable at this scale." Left to right: somata and vessels resolve at the cell field; nuclear envelope, nucleolus and myelinated axons in cross-section at the single neuron; membranes, mitochondria and vesicle clouds only at the native 4 nm. A spine-neck question needs the last panel everywhere in the volume, which is what makes it a petabyte problem.
Instructor script: "This is the bridge between scales, and it is exactly where scale leakage is manufactured. What claim is being carried across the join?" Ask whether the transfer comes with a stated registration residual and its maximum, not just its mean. Worked instance of the rule: "fraction of inputs onto spines vs shafts" puts the analysis scale at the spine neck (50–200 nm), which forces EM. "Does area A project to area B at all" puts it at the axon bundle, where light-sheet at 1 µm is the correct choice — about five orders of magnitude cheaper.
Anchor step 2 in the modality chart: resolution, volume and throughput form a budget — you may choose two. FIB-SEM buys resolution and gives up volume. Light-sheet buys volume and throughput and gives up resolution. Multibeam ssSEM (61 or 91 beams in parallel) buys resolution and volume by throwing throughput engineering at the problem, which is why 1 mm³ became feasible. Check-yourself case: thousands of L2/3 neurons, does each project to AL and to PM? That is barcoded projection mapping (MAPseq/BARseq), not EM — it needs statistical power over cells, not geometry within a cell.
Conversions are lossy and usually one-way: volume -> segmentation -> skeleton/mesh -> graph can remove the geometry your question needed. Typical sizes for one cortical neuron: volume GB; mesh 10–100 MB; skeleton 0.1–5 MB; graph bytes. Which representation for which question: "how many synapses between A and B" -> graph; "where on the dendrite" -> skeleton plus synapse coordinates; "are these spines larger" -> mesh; "is this a merge error" -> volume, always. The rule that saves projects: decide which representation the endpoint needs before the pipeline runs, and archive the next-richer one. A reviewer who asks whether a motif enrichment is explained by proximity cannot be answered from a graph with no geometry.
Pipeline from unit §4: vasculature is the best EM<->LM anchor in cortex; soma positions second; layer boundaries are weak. Fit the lowest-complexity transform that works — a flexible enough warp will align anything. Hold out anchors. A mean residual of 3 µm can hide a 40 µm error in one corner. Anisotropy warning: on a 4 x 4 x 40 nm stack, an isotropic smoothing kernel or distance metric is silently wrong by a factor of 10 in z. Check every library call for voxel spacing.
For a 1 mm³ ssTEM volume (~1.6 PB raw): aligned pyramid +30–50%; affinity maps ~1x raw; segmentation labels 0.1–0.5x; meshes 1–10 TB; skeletons 10–100 GB; synapse table ~50–200 GB; edit history grows without bound. Instructor script: "A petabyte in cold storage is cheap. A 200 GB table that must answer 50 concurrent interactive queries is the part that needs engineering. Budget for query load, not just capacity." Unit 04 picks this up.
Good seed for the share-out, from the unit: a tractography paper reports a "structural connection" between two regions; an EM study of one region finds no axons from the other. Both can be correct. Tractography streamlines are model output, not observed axons; an EM volume samples a small region at some detection sensitivity. Neither claim, as usually written, states its sampling and inference model clearly enough to be compared.
For the rejected alternatives, hold learners to the unit's standard: "too low resolution" is not a reason. "Cannot resolve a 1 µm AIS segment's synaptic input, and cannot distinguish symmetric from asymmetric synapses" is.