Connectomics in 2026: Mapping the Brain Wire by Wire with AI

A cubic millimetre of brain — a chunk the size of a grain of sand — holds roughly 50,000 cells and hundreds of millions of synapses. To trace every one of those connections, you have to slice the tissue into sheets thinner than a virus, photograph each sheet under an electron beam, and then follow every tangled wire through a stack of thousands of images. That is connectomics explained in one breath: the industrial-scale project of building a complete wiring diagram of a nervous system, neuron by neuron and synapse by synapse. It has moved from a decades-long hand-tracing fantasy to a working pipeline because AI now does the tracing. In 2024 the field mapped an entire adult fruit-fly brain; in 2025 it mapped a cubic millimetre of mouse cortex with half a billion synapses; and Nature Methods named the whole approach its Method of the Year for 2025.

What this covers: why a connectome matters, the full pipeline from tissue to graph, how AI segmentation and proofreading actually work, the landmark 2024–2025 datasets, and the hard physical limits standing between us and a whole mammalian brain.

Context and Background

The word “connectome” was coined in 2005 by analogy with “genome”: a complete map of the connections in a brain, the way a genome is a complete map of an organism’s DNA. The idea is old — Santiago Ramón y Cajal drew neurons and their contacts by hand in the 1890s — but the ambition of a dense reconstruction, where every cell and every synapse in a volume is accounted for, is recent. The only complete connectome that existed for decades was the nematode worm C. elegans, with 302 neurons, painstakingly traced by hand from electron micrographs and published by Sydney Brenner’s group in 1986. That single result took more than a dozen years of human labour.

There is a reason the worm stood alone for so long. Scaling up is not linear in difficulty. Doubling the tissue volume roughly doubles the imaging time, the storage, and the number of neurons to trace, but it more than doubles the chance that a single tracing error somewhere corrupts a long-range connection you care about. Errors compound with the length of a neuron’s arbor, and mammalian neurons have arbors that thread millimetres through the tissue. So each order-of-magnitude jump in volume demanded a corresponding jump in automated accuracy, not merely more microscope hours or more storage. That coupling — between raw scale and the accuracy you need to survive it — is the throughline of the entire field’s history, and it explains why progress came in sudden bursts tied to algorithmic breakthroughs rather than as a smooth climb.

Everything since has been a race between two curves: how fast you can image tissue at synaptic resolution, and how fast you can interpret the images into a graph of cells and connections. For thirty years, interpretation was the bottleneck. A human tracer following one neuron through a stack of electron-microscopy sections could take hours per cell; a mammalian cortical column has tens of thousands. The field only became tractable when deep-learning segmentation, borrowed and heavily adapted from computer vision, began to do the tracing at machine speed. Today, the throughput bottleneck is shifting back toward imaging and storage. For an adjacent example of AI dissolving a biology bottleneck, see how cryo-EM plus AI collapsed the structure-determination timeline for proteins. The connectomics story rhymes with it: a slow, artisanal measurement turned into a data-and-compute problem. For the state of the art on volume EM as a method, Nature Methods’ 2025 Method of the Year editorial (Nature Methods, 2025) is the canonical reference.

The Connectomics Pipeline: From Tissue to Wiring Diagram

Building a connectome is a linear pipeline with a human-in-the-loop stage bolted onto the end. Tissue is chemically fixed and stained with heavy metals so that membranes scatter electrons; it is embedded in resin and cut into thousands of ultrathin serial sections; each section is imaged by a volume electron microscope; the resulting petabyte-scale image stack is aligned, then segmented by AI into individual neurons; synapses are detected; humans proofread the automated result; and finally the corrected objects are reduced to skeletons and assembled into a graph you can query.

Figure 1: The connectomics pipeline, left to right. Each stage is a distinct engineering discipline — wet-lab sample prep, precision ultramicrotomy, high-throughput electron microscopy, petascale image processing, deep-learning segmentation, and graph analysis. A failure at any early stage (a torn section, a staining artefact, a misalignment) propagates downstream and shows up as a tracing error a proofreader must fix by hand.

Tissue preparation and staining

Electron microscopy sees contrast, and biological membranes are nearly transparent to electrons on their own. So the first job is to make the interesting structures — cell membranes, synaptic vesicles, mitochondria — scatter electrons strongly. That means soaking the fixed tissue in heavy-metal stains, principally osmium tetroxide, along with uranyl acetate and lead, in protocols descended from a lineage called rOTO (reduced osmium–thiocarbohydrazide–osmium). The staining has to be uniform across a block that may be a full cubic millimetre, because any region that takes up too little metal becomes low-contrast and un-traceable. Getting large blocks to stain evenly all the way through is one of the quiet, unglamorous problems that gates the whole field. The tissue is then dehydrated and embedded in a hard epoxy resin so it can be cut cleanly.

Ultrathin serial sectioning

To resolve a synaptic cleft — the gap between two neurons is on the order of 20 nanometres — you need to see structures a few nanometres across in the plane, and you need the third dimension too. That third dimension comes from cutting the resin block into serial sections, each typically 30 to 40 nanometres thick, using a diamond knife on an ultramicrotome. For a cubic-millimetre volume that is on the order of 25,000 to 33,000 sections, each of which must survive intact, be collected without loss or reordering, and be imaged. Two dominant strategies exist: collect the sections onto a tape or grid and image them later with a scanning electron microscope, or ablate the surface with an ion beam and image the freshly exposed face repeatedly. Both trade off differently between resolution, volume, and the risk of losing a section — and a single lost section can sever every neuron passing through it.

Volume electron microscopy

The imaging step is where the money and the throughput live. Several techniques compete, each with a different sweet spot:

• ssTEM (serial-section transmission EM) shoots electrons through thin sections onto a camera. It is fast and high-resolution in-plane, and modern variants such as GridTape automate section handling to feed multiple microscopes in parallel. This is the workhorse behind the fly and mouse connectomes.
• FIB-SEM (focused ion beam SEM) alternately mills away a few nanometres of the block face with an ion beam and images the new surface. It gives near-isotropic resolution — the same fineness in depth as in-plane — which makes segmentation dramatically easier, but the volume it can process is smaller.
• MultiSEM / multibeam SEM uses an array of 61 or 91 electron beams scanning in parallel, reaching imaging speeds reported up to roughly 1,820 megapixels per second — close to a hundred times a conventional single-beam SEM (ZEISS multibeam array tomography). Parallelism is the only way petascale volumes become feasible on a human timescale.

The output is a stack of aligned images at roughly 4-nanometre pixel resolution in-plane. At that sampling, a cubic millimetre of tissue is not a modest file — it is on the order of a petabyte or more, which is why storage and I/O now sit near the centre of the problem.

There is a hidden step between imaging and segmentation that quietly consumes enormous effort: alignment, also called stitching and registration. Each section is imaged as a mosaic of thousands of overlapping tiles, and each section sat on the microscope stage in a slightly different position and rotation than its neighbours. Before any AI can trace a neuron across the depth of the stack, the tiles within a section must be stitched into a seamless plane, and every plane must be warped into register with the ones above and below it so that a wire crossing a section boundary lines up on both sides. Sections can be torn, folded, wrinkled, or partially missing, and the alignment software must model those distortions with non-rigid transforms rather than simple shifts. A misalignment of even a few pixels at a section boundary looks, to a segmentation network, exactly like a neuron changing direction — which is why alignment errors are a leading upstream cause of the merge-and-split mistakes proofreaders later have to fix.

How AI Reconstructs Neurons: Segmentation, Synapses, and Proofreading

Direct answer: AI reconstructs a connectome by running a 3D convolutional neural network over the aligned image volume to assign every voxel to a neuron, detecting synapses as a separate classification task, and then handing the automated segmentation to human proofreaders who fix the merge-and-split errors the network makes. Two network families dominate the segmentation step — affinity-graph U-Nets and flood-filling networks.

Figure 2: Two families of 3D segmentation. One path predicts affinities between neighbouring voxels and then agglomerates them; the other floods outward from a seed to grow one object at a time. Both start from a 3D CNN reading the raw electron-microscopy voxels, and both feed a downstream merge-and-split error graph that proofreaders resolve. The choice shapes the entire error profile a human team will spend months cleaning up.

Affinity graphs and 3D CNNs

The classic approach trains a 3D U-Net to predict, for each pair of neighbouring voxels, an affinity: the probability that the two voxels belong to the same neuron. That gives you a dense graph over the volume where edge weights are affinities. You then run a watershed to produce over-segmented fragments called supervoxels, and agglomerate them greedily until the merge scores say stop. The strength of this method is speed and parallelism — affinities are a per-voxel prediction you can compute in tiles across a cluster. The weakness is that a single wrong high-affinity edge can merge two neurons into one blob, and undoing such merges downstream is expensive.

Flood-filling networks

Google and the Max Planck Institute took a different route with flood-filling networks (FFNs). An FFN is a CNN with a recurrent pathway: it starts from a seed voxel inside one neuron and iteratively predicts, extends, and refines the mask of that single object, flooding outward until it hits a membrane on all sides. Because it reconstructs one neuron at a time with a memory of the shape it is building, it makes far fewer topological errors. The published result improved reconstruction accuracy by roughly an order of magnitude over prior deep-learning methods (Google Research), measured as the expected run length of a neuron traced before an error. The cost is compute: flooding each object is more expensive than a single feed-forward affinity pass, so production pipelines often combine both — affinities for a fast draft, FFNs for accuracy where it matters.

Synapse detection and skeletonisation

Segmentation gives you the wires; you still need the connections. Synapse detection is a separate learned task: a network classifies voxels or local regions as pre-synaptic (vesicle-filled boutons) or post-synaptic (density on the receiving side), and pairs them to produce directed synaptic contacts with a polarity. Only when segmentation and synapse detection are both in hand can you assign each synapse to the two neurons it links. Finally, the volumetric neuron masks — huge, memory-hungry objects — are reduced to thin skeletons, a centre-line representation that captures branching structure at a tiny fraction of the storage. The skeletons plus the synapse table are the connectome: a directed graph where nodes are neurons and weighted edges are counts of synapses.

Proofreading and the human loop

No automated pipeline is error-free at connectome scale, and the errors that remain are exactly the ones that corrupt the graph: a merge fuses two neurons into one, and a split breaks one neuron into pieces. Both distort connectivity. So the last stage is human proofreading, and it is enormous. FlyWire — the platform that finished the fly brain — turned proofreading into a crowdsourced, gamified effort where hundreds of scientists and volunteers corrected the automated segmentation in a shared 3D environment, guided by tools that flag likely errors for review.

Figure 3: The proofreading feedback loop. Automated segmentation produces a draft graph, error-detection heuristics flag suspicious merges and splits, humans confirm or correct them, and the wiring graph is rebuilt. This loop is what converts a “mostly right” AI reconstruction into a scientifically usable connectome, and it is where most of the human effort — and calendar time — is spent.

Landmark Connectomes: Fly, Human Fragment, and Mouse

The abstractions above only mean something against real datasets. Four results define where the field actually is in 2026, and each pushed a different axis — completeness, human tissue, mammalian scale, and function.

C. elegans — the reference point

The worm remains the only whole nervous system mapped for a complex-behaviour organism, and it is small: 302 neurons and roughly 7,000 chemical synapses. Its value now is as a benchmark and a sanity check — a system small enough that the entire graph fits in your head, where researchers test what a connectome can and cannot predict about behaviour. It also frames the scaling problem starkly: going from 302 neurons to a fly’s ~140,000 is nearly a 500-fold jump, and from the fly to a mouse another ~500-fold.

The fly brain connectome — FlyWire, 2024

The landmark of 2024 was the complete fly brain connectome of an adult Drosophila melanogaster, published in Nature by the FlyWire consortium. The proofread reconstruction contains 139,255 neurons and about 54.5 million synapses, annotated with cell types, nerves, hemilineages, and predicted neurotransmitter identities (Nature, 2024). It is the first complete wiring diagram of an entire adult brain of any animal that can walk, fly, court, and learn. Within months of release, researchers used it to simulate the fly’s visual and motor circuits and reproduce known behaviours in software — a striking demonstration that a static wiring diagram carries real functional information.

H01 — a fragment of human cortex

In parallel, the Lichtman lab at Harvard and Google’s connectomics team released H01, a reconstruction of roughly one cubic millimetre of human temporal cortex — about half a grain of rice. The imaging alone is 1.4 petabytes. The volume contains on the order of 57,000 cells (roughly 16,000 neurons, 32,000 glia, and 8,000 blood-vessel cells) and about 150 million synapses (Google Research). H01 is not a connectome of a human brain — it is a tiny biopsy — but it is the largest synaptic-resolution reconstruction of human tissue to date, and it surfaced genuinely new biology, including rare but strong axonal connections where two neurons form dozens of synapses onto each other.

MICrONS — a functional cubic millimetre of mouse

The 2025 landmark was MICrONS (Machine Intelligence from Cortical Networks), an IARPA-funded effort that reconstructed a cubic millimetre of mouse visual cortex: more than 200,000 cells and about 500 million synapses (Nature, 2025). What sets MICrONS apart is that the same tissue was first imaged functionally — calcium imaging recorded the activity of roughly 75,000 neurons in an awake mouse watching videos — and then imaged structurally by EM and co-registered. For the first time at this scale, you can ask what a neuron does and see exactly what it is wired to. That structure-plus-function pairing is the whole point, and it is why 2025 felt like an inflection.

Figure 4: The scaling ladder of connectomes. From 302 neurons in the worm to ~140,000 in the fly, to cubic-millimetre human and mouse fragments with hundreds of millions of synapses, to the projected whole-mouse and whole-human targets at exabyte and zettabyte scale. Each rung is roughly two to three orders of magnitude harder in data volume than the last, which is why the ladder gets steep fast.

Dataset	Organism	Volume / scope	Neurons (approx.)	Synapses (approx.)	Raw data
C. elegans	Nematode	Whole nervous system	302	~7,000	small
FlyWire	Fruit fly	Whole adult brain	~139,000	~54.5 million	multi-TB
H01	Human	1 mm³ cortex fragment	~16,000	~150 million	1.4 PB
MICrONS	Mouse	1 mm³ visual cortex	>200,000 cells	~500 million	~PB

The neuron and synapse counts above are from the primary publications; treat the round numbers as the field’s reported figures, not exact constants, since proofreading continues to refine them.

Trade-offs, Gotchas, and What Goes Wrong

A connectome is a beautiful object that is also, in three important ways, a lie of omission — and the gap between what it captures and what a brain does is where most of the honest caveats live.

The first and largest problem is scale versus data. The fly brain took a whole-brain volume in the terabytes and a decade of consortium effort. A mouse brain is roughly 500 times larger by volume, which projects to about an exabyte of raw EM data; a human brain would be on the order of a zettabyte (10²¹ bytes) (whole-brain imaging prospects, 2025). Storing, moving, and segmenting an exabyte is not a bigger version of today’s problem; it is a different problem, gated by storage cost, network bandwidth, and the electricity to run segmentation across it. The US BRAIN CONNECTS program funded whole-mouse-brain tooling in 2023, but funding turbulence in 2025 showed how fragile the largest efforts are.

The second problem is that a connectome is static and structural. It tells you which neurons connect and how many synapses join them, but not the strength of those synapses moment to moment, not their sign in every case, and — critically — nothing about neuromodulation. Chemicals like dopamine and serotonin reconfigure how a circuit computes without changing a single wire; a wiring diagram is blind to them. Two animals with near-identical connectomes can behave differently, and the same connectome can support different computations under different neuromodulatory states.

The third is the single-specimen problem. Each dense connectome is one individual, frozen at one instant. It cannot show plasticity, learning, or the natural variation between brains. And residual proofreading errors never fully vanish: even a low per-neuron error rate, multiplied across hundreds of thousands of cells, leaves thousands of uncertain connections that downstream analyses must treat with care. A connectome is a map, not the territory — and specifically it is a map with no traffic on it.

There is also an economics gotcha that rarely makes the headlines. The dominant cost of a large connectome is not the microscope time but the human proofreading and the compute-plus-storage to hold and process a petascale volume. Proofreading a cubic-millimetre mammalian dataset is measured in tens of person-years even with AI doing the first pass, which is why crowdsourcing and error-flagging automation are not niceties but necessities. As volumes grow, the proofreading burden grows with the number of neurons and the length of their arbors, so a whole mouse brain does not need 500 times the proofreaders of a cubic millimetre — it needs a qualitatively better automated segmenter, one accurate enough that humans only spot-check. That is the real research frontier: not just imaging faster, but pushing automated accuracy high enough that human correction stops being the rate-limiting step. Until then, every jump up the scaling ladder is bought with human hours, and that cost curve, not the physics of electron optics, is what most constrains how fast the field can move.

Practical Recommendations

If you are a researcher, engineer, or student trying to engage with connectomics rather than just read about it, the field is unusually open — the landmark datasets are browsable and downloadable, which means you can do real analysis without running a microscope.

Start by treating the connectome as a graph-analysis problem, because that is what it is once the imaging is done. Pull a published dataset, load the neuron-and-synapse graph, and ask concrete questions before worrying about biology: degree distributions, motifs, path lengths, and community structure. Then layer the biology back on — cell types, neurotransmitters, and, where available, function.

A short checklist for getting oriented:

• Pick the right dataset for your question. Whole-circuit logic → FlyWire; mammalian structure-function → MICrONS; human tissue detail → H01.
• Use the hosted explorers first. MICrONS Explorer, the FlyWire browser, and the H01 viewer let you inspect neurons in 3D before you download a byte.
• Respect the caveats. Note proofreading status and treat single-specimen findings as hypotheses, not laws.
• Separate structure from function. A connectome constrains what a circuit can do; it does not tell you what it is doing without activity data.
• Watch the storage math. Even a subvolume can be terabytes — plan compute and I/O before downloading.
• Compare to artificial nets carefully. Biological connectivity motifs are suggestive for artificial network design, but the analogy is loose.

Frequently Asked Questions

What is connectomics in simple terms?

Connectomics is the science of building a complete wiring diagram of a nervous system — a map showing every neuron and every synaptic connection between them. In practice it means slicing brain tissue into ultrathin sections, imaging each with an electron microscope, and using AI to trace every wire through the resulting image stack. The end product is a graph: neurons are nodes, synapses are the edges. It is to the brain’s wiring what a genome is to an organism’s DNA.

Why is AI essential for connectomics?

Because the tracing is otherwise impossible at scale. A cubic millimetre of cortex contains hundreds of millions of synapses spread across tens of thousands of sections, and a human tracing one neuron by hand can take hours. Deep-learning segmentation — 3D convolutional networks predicting affinities, or flood-filling networks growing one neuron at a time — does that tracing at machine speed. Without AI segmentation, the fly and mouse connectomes would still be decades away. Humans then proofread the AI’s output rather than tracing from scratch.

Has anyone mapped a whole brain connectome?

Yes, for small animals. The nematode C. elegans (302 neurons) was mapped in 1986, and the adult fruit fly’s whole brain (~140,000 neurons, ~54.5 million synapses) was completed by the FlyWire consortium and published in Nature in 2024. For mammals, only fragments