Two labs can section the same block of fixed cortex, image it under comparable microscopes, and publish synapse densities that differ by a third. The gap is not usually sloppiness. It is a procedural choice made long before anyone counts: how thick the knife cuts the slice. This piece breaks down where that bias enters, what the physics of the electron beam allows, and which questions separate a robust density estimate from a provisional one.

Two Labs, Same Brain, Different Counts

Ask two electron microscopists for a synapse density from the same tissue and you may get numbers that barely overlap. The usual reflex is to blame sampling or antibody specificity. In serial-section work on neuropil, the more common culprit sits at the ultramicrotome: a 40-nanometre section and a 70-nanometre section are not the same measurement window, and they do not see the same population of structures.

Thin sections expose more asymmetric synapses, the kind with a thick postsynaptic density that most excitatory counting relies on. Thick sections bury small spines and dim puncta inside overlapping material. The debate over an optimal cutting protocol is genuinely unresolved because the optimum depends on what you want to count, and labs rarely state that trade-off explicitly.

The deeper disagreement is interpretive. One camp treats the extra synapses found in thin sections as real biology that thick sections hide. The other treats some of the thin-section yield as artifact introduced by compression, chatter, or staining gradients at the cut face. Both readings have support, and the evidence that would settle it is not routinely collected.

The Physics of Seeing Small Things

A transmission electron microscope accelerates electrons through the column at potentials that, in standard biological work, commonly fall in the 50 to 125 kilovolt range. Those electrons cannot cross much biological material before scattering removes them from the imaging beam. That is the physical reason ultramicrotomy exists at all: the specimen has to be thin enough for electrons to pass, and heavy staining with osmium and lead salts supplies the contrast.

Thinner slices improve resolution in the plane of the cut and reduce the blur that comes from overlapping structures along the beam path. They also cost signal. Less material means fewer scattered electrons, weaker contrast, and a greater chance that the section tears or compresses under the diamond knife. A 30-nanometre slice can be beautiful and still be unrepresentative if it crumpled.

Higher accelerating voltage buys penetration, which lets thicker sections be imaged at all. It does not buy back the information lost to overlap. A 100-kilovolt beam through a 100-nanometre section still integrates every density along that path into one projection, and the counting problem lives inside that integration.

How Thickness Biases Synapse Detection

The counting bias has a name in stereology: overprojection. A postsynaptic density that extends beyond the section thickness appears in adjacent sections or smears across the image, and a naive count treats each appearance as a separate synapse. Thick sections make this worse because more of each structure falls inside the slab, and automated segmentation tools trained on one thickness can double-count or merge objects when applied to another.

Thin sections change the failure mode rather than removing it. Small spines that a thick section hides become visible, so counts rise. But thin sections also lose the top and bottom of structures that straddle the cut, and the lost fragments are exactly the ambiguous ones a human annotator would have debated. Hedged estimates in the methods literature put the practical difference between common thicknesses somewhere in the 20 to 40 percent range for dense neuropil, though the figure varies with staining and region.

Automated pipelines add a second layer. A segmentation model tuned on one dataset learns that synapses look a certain size and contrast. Feed it images from a different thickness and it will confidently mislabel. The same reproducibility problem shows up in software generally, as this site has noted in a piece on code repositories preserving the software that results depend on.

Why the Trade-Off Is Hard to State

Most methods sections report a single thickness as if it were a settled parameter, when it is really a compromise among competing goals. Thinner sections favor resolving small profiles and reducing overprojection. Thicker sections favor throughput, ribbon stability, and enough contrast to survive automated detection over large volumes. A lab that wants to map a cubic millimetre of tissue cannot afford to cut at 30 nanometres across the whole block; the section count and the imaging time both become prohibitive. A lab that wants to resolve the smallest asymmetric contacts cannot afford to cut at 70.

The compromise is rarely written down. Reviewers ask whether the sampling scheme is unbiased, and the answer can be yes while the thickness choice still tilts which structures enter the sample at all. That is a subtler form of bias than a bad random-number seed, and it does not show up in a power calculation. It shows up when two groups compare notes and find their densities do not reconcile.

There is also a specimen-side cost that methods sections tend to compress into a sentence. Fixation quality, osmium exposure, and the age of the block all interact with thickness. Older blocks cut thin tend to shatter; well-fixed fresh blocks tolerate it. So a protocol that works in one lab can fail in another for reasons that have nothing to do with the counting method, and the failure is easy to misread as a biological difference.

A Case Study in Contested Counts

The pattern is familiar from published exchanges in the connectomics literature, where groups working on comparable cortical tissue have reported densities that differ substantially and then traded methods sections rather than insults. The recurring hinge is section thickness, often 40 nanometres on one side and nearer 70 on the other, combined with different approaches to correcting for overprojection.

Calibration standards are the second hinge. A lab that calibrates its magnification and section thickness against a known grating or a re-sectioned block can defend its numbers. A lab that trusts the ultramicrotome's nominal feed rate cannot, because thermal drift and knife wear move the true thickness away from the dial. The dispute is less about who is right than about which measurements were never made.

Calls for blind multi-lab validation have circulated for years. The obstacle is practical: sharing blocks, agreeing on a staining protocol, and re-sectioning the same tissue at several thicknesses is expensive and unglamorous. A related piece on two-photon microscopes tracking dendrites that one-photon scatter blurs makes a parallel point about instrument choice shaping what a field can see.

What a Careful Reader Should Ask

Start with the thickness. If a paper reports synapse density without stating section thickness and how it was verified, the number is not yet interpretable. Nominal feed rate is a setting, not a measurement, and the two diverge more often than methods sections admit.

Next, ask whether counts were corrected for overprojection, and by what method. The disector approach, which samples in three dimensions rather than projecting, handles the problem cleanly but requires serial sections and careful alignment. Many studies use a simpler profile count and apply a correction factor borrowed from another tissue.

Then look for re-sectioning. A lab that cut the same block at two thicknesses and reported both densities has done the internal control that most studies skip. Finally, check whether raw images or tracings are available. Without them, a reader cannot audit the segmentation, and the density stands on the authors' word alone.

Practical Steps for Assessing Claims

Before trusting a density figure, work through a short checklist against the methods and any supplementary material.

• Find the stated section thickness and the accelerating voltage, and check whether thickness was verified by a method other than the microtome's feed setting.
• Look for calibration against a known standard, such as a diffraction grating replica or a re-sectioned reference block.
• Confirm whether the counting method corrects for overprojection, and note whether the correction was validated in the same tissue.
• Request raw images or tracings from the corresponding author, and treat refusal as a reason to discount the number.
• Compare the density against at least one independent lab working on comparable tissue before citing it as a regional constant.
• Treat any single-lab density as provisional until a second group reproduces it with a different cutting protocol.

None of this makes the published numbers wrong. It makes them conditional, which is what a measurement of a dense, overlapping structure honestly is. The thickness of a slice is a decision about what counts as visible, and every synapse density carries that decision inside it.