Stylized vector art: pipeline stages running above a chunk grid.

Synapse Detection

A connectome is a graph, and the edges of that graph come from a synapse detector. Everything downstream — connection weights, motif counts, cell-type fingerprints, connectome-constrained models — inherits whatever that detector got wrong. The reconstruction pipeline page describes synapse detection in five bullets as one stage of Layer 4. This page is the argument behind those bullets: what the task is, what has been tried, what the published numbers are, and what you have to measure before you are entitled to use somebody else’s synapse table.

Start here: this is a solved problem with three residuals

Read the rest of this page in that order, because the framing matters for how much of your project’s budget belongs here.

Within a volume you have ground truth for, cleft detection is solved. SynEM reports 97% precision and recall on the connectome — the thing you actually publish — even while sitting at 88% per individual synapse, because errors on single contacts wash out once contacts are aggregated into connections. H01’s pipeline detected 149.9 million synapses at a 3.2% / 2.7% false-discovery rate. No modern project fails because it cannot find clefts.

Three residuals are real, and they are what this page is about:

  1. Recall is not symmetric. H01’s detector missed 11% of excitatory and 35% of inhibitory synapses. Aggregation rescues precision, not recall, and an inhibitory deficit distorts every E/I ratio computed from the table.
  2. Partner assignment is weaker than localisation, and much weaker where synapses are polyadic. Buhmann et al. report F1 from 0.59 to 0.73 across four areas of the same fly brain.
  3. Transfer is the operational cost. A detector trained on one volume does not move to another without new ground truth — which is why the site’s open problem is titled “Synapse detection that generalizes” rather than “synapse detection”. As of 2025 the cross-dataset benchmarks say this is still open.

So the honest budget line is not “we must solve synapse detection”. It is: run the established detector, then spend your effort establishing what its recall and partner accuracy are in your region, and report the table’s provenance. Section 7 is that checklist.

For what a synapse looks like and how to recognise one by eye, see Synapse classification.

1. Three problems, not one

“Synapse detection” names at least three tasks that succeed and fail independently.

1. Localisation. Find the synapse. Depending on the method this means segmenting the synaptic cleft as a set of voxels, or predicting a point annotation for the presynaptic site, or classifying the interface between two already-segmented processes as synaptic or not.

2. Partner assignment. Given a synapse, say which segment is presynaptic and which is postsynaptic. This is a separate learning problem: Buhmann et al. (2021) and Turner et al. (2020) both treat it as its own network, and Huang et al. (2018) use a U-Net for presynaptic sites and a separate multilayer perceptron, conditioned on the local segmentation, for postsynaptic partners.

3. Sign or type. Say whether the connection is excitatory or inhibitory. In mammalian cortex this is read from postsynaptic-density morphology; in Drosophila it is read from transmitter identity predicted from ultrastructure (Eckstein et al., 2024). Section 4 is about why this is the hardest of the three.

Two structural facts change the shape of the problem depending on your tissue. First, insect synapses are polyadic: one presynaptic site contacts several postsynaptic partners, so “one cleft” is not “one edge”. Huang et al. (2018) describe theirs as “a complete solution for polyadic synapse detection”, and the WASPSYN annotations record pre- and post-synaptic coordinates “together with their one-to-many connectivity information” (Li et al., 2024). Second, some methods need a segmentation first and some do not: SynEM classifies borders between already-segmented processes (Staffler et al., 2017) while Heinrich et al. (2018) segment clefts directly out of the raw image. That decides whether segmentation errors propagate into your synapse table, and it means the two methods’ F-scores are not measuring the same object.

2. What has been tried, with numbers

Method Year What it predicts Reported performance Tissue and preparation
ilastik synapse detection (Kreshuk et al.) 2011 Asymmetric (presumed excitatory) synapse segmentation from voxel features 0.92 recall at 0.89 precision on 111 validation synapses, “comparable to that of the experts” (three independent annotators) Adult mammalian cortex, FIB/SEM, near-isotropic
Context-cue synapse segmentation (Becker et al.) 2013 Synapse segmentation plus synaptic orientation Evaluated on three datasets; recovers orientation as a by-product EM stacks, mixed
SyConn (Dorkenwald et al.) 2017 Synapses and synapse types alongside mitochondria, compartments and cell types Framework paper; used to compute songbird basal-ganglia wiring SBEM: zebrafish, mouse, zebra finch
SynEM (Staffler et al.) 2017 Neurite-interface classification: synaptic vs non-synaptic 88% precision, 88% recall per synapse; 94% precision, 89% recall for spine synapses; 97% precision and recall at the level of binary cortical connectomes Mouse cortex, conventional en-bloc staining, SBEM
3D U-Net cleft segmentation (Heinrich et al.) 2018 Synaptic cleft voxels, by regression on a signed distance transform Reported as a significant improvement over the prior state of the art on CREMI Adult Drosophila, anisotropic ssTEM; applied to ~50 teravoxels of the whole fly brain
Fully-automatic synapse prediction (Huang et al.) 2018 Presynaptic sites (U-Net) plus postsynaptic partners (MLP on segmentation-conditioned features) Introduced connectome-scale evaluation metrics; reports that complete automatic prediction characterises most connectivity correctly Drosophila, polyadic
Attentional voxel association networks (Turner et al.) 2020 Presynaptic and postsynaptic masks, generated from a cleft mask used as an attention gate Evaluated as part of a combined cleft-plus-partner system Mouse somatosensory cortex
Synful (Buhmann et al.) 2021 Synaptic partners directly, whole-brain F1 of 0.73, 0.68, 0.66 and 0.59 in four different brain areas; 244 million putative synaptic partners extracted from FAFB; 92–96% of edges correctly sorted into weakly (<5 synapses) and strongly (≥5) connected Adult Drosophila, FAFB ssTEM
Cerebellum-specific contact classifier (Park et al.) 2022 Synaptic vs non-synaptic contacts, plus pre/post side and excitatory/inhibitory type F1 = 0.955 on a test volume containing 508 synapses Mouse cerebellar molecular layer
H01 detector (Shapson-Coe et al.) 2024 Three-class U-Net (background / presynaptic / postsynaptic) plus a ResNet-50 excitatory-vs-inhibitory classifier See §4 — the numbers differ by more than threefold between excitatory and inhibitory recall Human temporal cortex, ssEM at 4 × 4 nm, 33 nm sections
SimpSyn (Mohinta et al.) 2025 Dual-channel spherical masks around pre- and post-synaptic sites, single-stage residual U-Net Outperforms Synful in F1 on all volumes in a four-dataset invertebrate benchmark; the authors report that generalisation across datasets remains limited Adult and larval Drosophila, Megaphragma viggianii

Three readings of that table matter more than the individual rows.

The numbers are not on a common scale. Kreshuk’s 0.92/0.89 is per-synapse on near-isotropic FIB/SEM; SynEM’s 88/88 is per-interface on anisotropic SBEM; Buhmann’s 0.59–0.73 is per-partner-pair on ssTEM. A method that looks worse here may simply have been measured on a harder unit in harder tissue. Never rank two detectors by numbers taken from their own papers.

Aggregation is doing enormous work. SynEM classifies interfaces at 88% precision and recall, and yields 97% precision and recall in binary cortical connectomes. Nothing about the classifier changed; what changed is that a neuron pair is usually connected by more than one synapse, so independent per-synapse errors partly cancel when you only ask “are these two neurons connected?”. Buhmann et al. show the same effect from the other side: per-connection F1 of 0.59–0.73, but 92–96% of edges correctly assigned to the weak/strong classes most analyses actually use. The number you need is the one measured at the level of your claim — for most connectomics, the edge, not the synapse.

The field’s accounting has moved. Huang et al. (2018) observed that as segmentation improved, synapse annotation came to consume “upwards of 50% of total effort”. Synapse detection stopped being the pipeline’s afterthought at roughly that point, which is why the later entries above are about partners and generalisation rather than about finding clefts.

3. Benchmarks: what a CREMI score does and does not tell you

CREMI (Circuit Reconstruction from Electron Microscopy Images) is the reference benchmark. It provides three datasets, each consisting of two (5 µm)³ volumes — training and testing, 1250 × 1250 × 125 px — of serial-section transmission EM of the adult Drosophila brain at [4, 4, 40] nm voxel resolution. Each training volume ships neuron ids, synaptic-cleft segmentation, and (pre, post) partner pairs. Three tasks are scored:

CREMI task How it is scored
Neuron segmentation Variation of Information, Adapted Rand error, and Tolerant Edit Distance — the metrics explained in Metrics and QA
Synapse detection F-measure over false positives and false negatives, where a predicted cleft voxel beyond a threshold distance from any ground-truth cleft counts as a false positive, and vice versa for false negatives
Connectivity (partner identification) F-measure over matched (pre, post) pairs, matched by solving an assignment problem that minimises Euclidean distance within a threshold

Note what the scoring does not do: it does not require a predicted cleft to overlap the ground-truth cleft voxel-for-voxel, only to fall within a distance tolerance. Cleft detection is a localisation problem with slack, which is why methods that regress a distance transform (Heinrich et al., 2018) rather than classifying voxels do well on it.

What a good CREMI score predicts. That your method is competitive on anisotropic fly ssTEM, at CREMI’s resolution, under CREMI’s convention for what counts as a cleft. That is worth knowing, and it is why CREMI training data underlies whole-brain fly cleft prediction: Heinrich et al. trained on CREMI’s annotated crops and then predicted clefts across roughly 50 teravoxels of the complete Drosophila brain.

What it does not predict.

WASPSYN (Li et al., 2024) is the benchmark built to measure what CREMI cannot: domain adaptation. Its authors annotated 14 image volumes from a biologically diverse set of Megaphragma viggianii brain regions drawn from three different whole-brain datasets, and ran it as an ISBI 2023 challenge. Their framing carries the number that defines the problem: manual annotation is so expensive that labelled training data is “often smaller than 0.001% of the large-scale image volumes in application”. That ratio, not any single F1, is why cross-dataset generalisation is the live question.

4. Excitatory versus inhibitory, and why it is harder

Cleft detection asks “is there a synapse here?” Sign classification asks “what kind?” — and the evidence for the answer is thinner, more protocol-dependent, and more asymmetric between classes.

The morphological basis is a statistical mapping, not a rule. Synapses sort into Gray’s type I and type II, corresponding to Colonnier’s asymmetric (AS) and symmetric (SS): AS have a thick postsynaptic density, SS a thin one. In cerebral cortex most AS are excitatory (glutamatergic) and most SS inhibitory (GABAergic) — “most”, not all. Cano-Astorga et al. (2024) validate the morphological call against molecular markers, using immunocytochemistry for the vesicular GABA transporter to confirm that symmetric synapses seen under their protocol are GABAergic.

The discriminating feature is a few voxels thick, and staining moves it. The same paper shows that potassium ferrocyanide, used in modern volume-EM protocols to enhance membrane contrast, makes postsynaptic densities thinner as its concentration rises, so symmetric synapses become progressively harder to identify; they recommend 0.1%. The consequence for a detector is stark: the feature your inhibitory classifier depends on is partly a property of the sample-preparation recipe, not only of the tissue. A classifier trained on one lab’s staining is being asked to transfer across chemistry, not just across brains.

The published numbers show the asymmetry directly. In H01, Shapson-Coe et al. (2024) report, from manual proofreading of a selection of axons across all cortical layers:

  Excitatory Inhibitory
False negatives (missed synapses) 11% 35%
False discovery rate 3.2% 2.7%
Correctly classified as this type 86.89% 84.98%

Precision is excellent and near-identical for both classes. Recall is not: the detector misses roughly one inhibitory synapse in three, and one excitatory synapse in nine. That single asymmetry is the most consequential number on this page, and §5 works out what it does to a result.

In insects, sign is a different problem entirely. Fly synapses do not give up their sign through AS/SS morphology, so the question becomes transmitter identity. Eckstein et al. (2024) trained networks on EM images at synaptic sites to predict six transmitters (acetylcholine, glutamate, GABA, serotonin, dopamine, octopamine) across a whole Drosophila brain, reaching 87% accuracy for individual synapses, 94% for neurons, and 91% for known cell types. The 87 → 94 step is the aggregation effect again: a neuron’s synapses vote. It also means fly connectome signs are properties of neurons in practice, so treating per-synapse predictions as independent evidence double-counts.

5. A worked judgement: recounting H01’s excitatory/inhibitory balance

Take the H01 synapse table as released and ask a routine question: what fraction of synapses in human temporal cortex are excitatory?

The straight count. Shapson-Coe et al. report 149,871,669 synapses automatically detected in the volume, of which 111,272,315 were classified as excitatory and 38,599,354 as inhibitory. That is 74.2% excitatory, 25.8% inhibitory. If you pull the released table and compute a ratio, this is your answer.

The corrected estimate. The authors do not report that number as their result. Adjusting for the measured false-discovery, false-negative and misclassification rates in the table above, they estimate the volume actually contains 102.5 million (67.1%) excitatory and 50.3 million (32.9%) inhibitory synapses.

What moved. Almost all of it is the 35% inhibitory false-negative rate. Precision was fine for both classes, so few of the detected synapses are spurious; the problem is the ones that were never detected, and they are disproportionately inhibitory. Running the arithmetic the other way: the raw table contains 38.6 M of an estimated 50.3 M inhibitory synapses — it is short by about 23% of the inhibitory population — while its excitatory count of 111.3 M sits about 9% above the 102.5 M estimate. The naive ratio is wrong in both directions at once, which is why the error on the ratio (74.2% versus 67.1%, or 7.1 percentage points) is larger than the error on either count alone.

How confident should you be in the correction? Less than in the raw count. It rests on proofreading a selection of axons rather than a dense re-annotation, so the rates carry their own sampling error, and they were measured across cortical layers that differ in inhibitory density. The corrected figure is an estimate from a sample; the raw figure is an exact count of a biased measurement. Neither is “the number of synapses in the tissue”, and a paper quoting either without saying which has not told you what it did.

What follows, well beyond H01.

  1. Any excitation/inhibition ratio computed straight from a released synapse table is biased toward excitation, by an amount set by the detector’s class-specific recall. Report the recall figures alongside the ratio, or do not report the ratio.
  2. Comparisons are safer than absolutes. For two regions processed by the same detector and staining, the difference in E/I ratio survives a shared bias that the absolute value does not.
  3. The bias is not random, so more data will not fix it. Ten times the volume buys ten times the confidence in the wrong number.

6. Why detectors do not transfer

Every one of the following differs between two EM volumes, and every one of them can move a detector’s output.

What changes Concrete example
Voxel size and anisotropy CREMI/FAFB at 4 × 4 × 40 nm ssTEM; H01 at 4 × 4 nm in-plane with 33 nm sections; FIB/SEM volumes near-isotropic. Heinrich et al. built a 3D U-Net specifically “to optimally represent isotropic fields of view in non-isotropic data” — the architecture itself encodes an assumption about the sampling grid
Staining chemistry Potassium ferrocyanide concentration changes apparent PSD thickness (Cano-Astorga et al., 2024) — directly attacking the inhibitory-synapse feature
Species ultrastructure Polyadic insect synapses versus predominantly monadic mammalian ones; the output structure differs, not just the appearance
Annotation convention What counts as a cleft, and where a “site” point is placed relative to it. Two ground-truth sets can disagree systematically while both being correct by their own rules
Region within one sample Buhmann et al.: F1 0.59 to 0.73 across four areas of the same fly brain

The honest summary is narrower than “models do not transfer”.

Within a volume and preparation, spatial generalisation is often fine. Heinrich et al. report that their model, trained on CREMI’s small annotated crops, “generalizes well to areas far away from where training data was available” — across a whole fly brain, including lamina.

Across preparations, it is not. SimpSyn’s authors, benchmarking on four invertebrate datasets, conclude that generalisation across datasets remains limited even for the model that wins within each dataset. SynapseNet’s authors built explicit domain-adaptation functionality into their tool rather than relying on a large training set. And WASPSYN exists because, in the organisers’ words, methods that “utilize in-domain labeled data and generalize to out-of-domain unlabeled data are in urgent need”.

That distinction is operational. If you are analysing one released volume, its detector was probably trained on that volume and the within-sample evidence applies to you. If you are bringing a published detector to new tissue, you are in the regime where the evidence says it will degrade, by an amount nobody can tell you in advance.

7. Before you trust a synapse table

You will usually not be running a detector. You will be downloading a table someone else’s detector produced. This is the checklist.

Ask Why it changes what you can claim
Which detector, which version, which materialization? Synapse tables are regenerated. A result that does not name the version is not reproducible — see Provenance and versioning
What was the ground truth: how many synapses, annotated by whom, with what agreement? Kreshuk et al. validated on 111 synapses with three independent experts; SynEM’s test set was 235 synapses among 20,319 non-synaptic interfaces. Numbers from a test set of a few hundred carry visible sampling error
Is the reported figure per synapse or per connection? SynEM: 88% versus 97%. Both are true; they answer different questions
Are precision and recall reported separately for excitatory and inhibitory? H01: 3.2% versus 2.7% FDR, but 11% versus 35% false negatives. A single combined F1 hides exactly the error that will bite you
Is performance stratified by region? Buhmann et al.: 0.59 to 0.73 within one brain. A whole-volume average tells you nothing about your region
What is the polyadicity convention: one row per synaptic site, or one per (pre, post) pair? Miscounting this inflates or deflates every insect connection weight
What distance tolerance was used in the evaluation? CREMI-style scoring matches within a threshold. A generous threshold flatters everything
Was the evaluation on the same preparation as the volume you are using? If not, §6 applies and the published number is an upper bound at best

If none of that is available, you can still bound your own risk in an afternoon. Take a random subvolume, annotate every synapse in it by hand, and compare against the table: missed table entries give you recall, spurious ones give you precision. Two calibration points set expectations. Kreshuk et al. found their algorithm’s error rate “comparable to that of the experts”, and SynEM reports its expert annotators at 93.6–94.6% and 97.9–98.9% precision/recall — so agreement in the mid-90s is roughly what two competent humans achieve. A detector matching your annotations more closely than that should make you suspicious of your annotations, not confident in the detector. For the manual pass itself, SynAnno (Lauenburg et al., 2025) provides guided, neuron-centric synapse proofreading with model-assisted error detection.

Self-check

  1. SynEM reports 88% precision and recall per synapse and 97% for binary connectomes. What changed, and when may you quote the 97%?
  2. A team beats the published CREMI cleft F1. What have they demonstrated, and what have they not?
  3. You compute an excitatory:inhibitory ratio of 74:26 from H01’s released table. Name the measurement that makes this wrong, and the direction.

Answers.

  1. Nothing about the classifier changed; the unit of the question did. Neuron pairs are usually connected by several synapses, so independent per-synapse errors partly cancel when you only ask whether an edge exists. Quote 97% when your claim is about the existence of connections in comparable tissue; not when it is about synapse counts, connection weights, or anything per-synapse.
  2. Competitiveness on anisotropic adult-fly ssTEM at 4 × 4 × 40 nm under CREMI’s annotation convention and distance-tolerant scoring. Not: mammalian tissue, other regions of a fly brain (Buhmann et al. span 0.59–0.73 within one brain), sign classification, or the accuracy of the weighted graph a user would build.
  3. The 35% false-negative rate for inhibitory synapses against 11% for excitatory. The released table is short by roughly 23% of the estimated inhibitory population, so the ratio is biased toward excitation; the authors’ corrected estimate is 67:33, about 7 percentage points away.

What this page does not cover

Go deeper

References