Synthetic neuron shapes improve automated brain mapping

Reconstructing a brain requires following the fine branches of individual neurons through densely packed tissue. A branch may twist, narrow, or pass close to another cell. An automated tracing system must preserve its identity through these changes, because a mistaken connection can alter the circuit that scientists infer from the reconstruction.

The scale of this task is visible in the male fruit fly central nervous system map described by Google Research on September 3. The collaboration led by HHMI Janelia reconstructed more than 166,000 neurons and 125 million synaptic connections, spanning the brain and ventral nerve cord. Human experts verified the automated reconstructions. Extending this approach to larger nervous systems depends partly on reducing the amount of correction each reconstruction needs. 1

An earlier study from Google Research addresses that problem through synthetic training data. MoGen, presented at ICLR 2026, generates three-dimensional neuron fragments that can help a reconstruction system judge whether an assembled shape is plausible. Its contribution is a measured improvement in the reconstruction of real tissue, linking generative modeling to the accuracy of a scientific instrument. 6

Following a cell through a volume

In electron-microscopy connectomics, a reconstruction algorithm assigns parts of a three-dimensional image volume to individual cells. This segmentation must remain consistent along a neuron's extended processes, or neurites. Two errors are especially consequential: a split divides one process into separate objects; a merge assigns parts of different cells to the same object.

Flood-filling networks introduced a way to trace a neuron iteratively. Starting from a seed location, the network predicts which nearby voxels belong to the same object. Its developing segmentation feeds back into subsequent predictions as tracing expands through the volume. In a 2018 zebra finch experiment, Januszewski and colleagues reported a mean error-free neurite path length of 1.1 millimeters, with four mergers over 97 millimeters of test-set path length. The improvement came with substantially greater computational cost. 2

These measurements capture something that an overall pixel-accuracy score can obscure: how long the reconstruction preserves a cell's identity. For circuit analysis, a small region of incorrectly assigned tissue may have a much larger effect than its volume suggests. A false join can bring the connections of two cells into a single reconstructed object.

!Three schematic tracings show a continuous branch, a split that breaks one branch into two objects, and a merge that joins two separate branches.

Figure 1. Reconstruction errors change object identity. Colors denote segmented objects; the highlighted gap and bridge indicate a split and a merge. Original schematic, not microscopy data.

Evaluating the shape of a proposed connection

PATHFINDER extends reconstruction into a sequence of related decisions. The system segments image volumes, proposes ways to assemble neuron fragments, and evaluates the plausibility of the resulting shapes. In their 2025 preprint, Januszewski and collaborators tested it on axons in a mouse cortex volume and reported an order-of-magnitude reduction in reconstruction error relative to the previous state of the art. 3

Shape evaluation adds information beyond the immediate appearance of a boundary. A proposed join has consequences for the larger object: its branches, its continuity, and the arrangement of its processes. Assessing that object gives the system another basis for accepting or rejecting an assembly.

This creates a training problem. A classifier needs examples of the geometries it will encounter, while verified biological reconstructions are expensive to produce. MoGen supplies additional examples at this particular stage. The generated fragments augment the training set of PATHFINDER's shape-plausibility classifier. The final reconstruction continues to be derived from the imaged specimen. 6

Learning to generate a neuronal surface

A neuron fragment can be represented by points sampled from its surface. Together, their three-dimensional coordinates describe the fragment's geometry without requiring a dense grid across all the empty space around it.

MoGen learns to generate these point clouds using flow matching. In the general formulation introduced by Lipman and colleagues, a neural network learns a velocity field that transforms a simple distribution, such as random noise, into a distribution resembling the training data. At generation time, a numerical solver follows the learned field to produce a sample. The training objective fits that field along prescribed probability paths, avoiding the need to simulate a complete generation trajectory for every training example. 4

For neuronal geometry, the changing sample is a collection of spatial coordinates. As generation progresses, an initially unstructured set of points becomes organized into a surface. This trajectory describes the model's sampling procedure; it does not simulate the biological growth of a neuron.

The architecture builds on PointInfinity, whose design separates a variable-sized point representation from a fixed-sized latent representation. Read and write operations exchange information between the two, while much of the computation occurs in the smaller latent representation. This makes it possible to increase the number of output points without increasing every part of the network's internal representation at the same rate. 5

MoGen adapts the latent-transformer approach by incorporating local geometric context. Its evaluation also considers geometric and topological properties tailored to neuronal structures, alongside a human study. These choices align the generator with the shapes that its downstream classifier must assess. 6

Testing whether synthetic shapes improve real reconstruction

The researchers trained the mouse-axon generator using surface samples from 1,795 human-verified axons. They then introduced millions of generated shapes into the reconstruction training pipeline. In the reported experiment, a training mixture containing 10% simulated data reduced reconstruction errors on reserved mouse axons by 4.4%, primarily through fewer merge errors. 7

The evaluation therefore goes beyond whether generated fragments look convincing. It asks whether training with them improves a separate system's decisions on real biological material. That downstream test is essential: visual realism alone cannot establish the usefulness of synthetic training data.

The authors estimate that the improvement could save more than 157 person-years of proofreading at whole-mouse-brain scale. This is a projection, rather than labor saved on a completed mouse connectome. The mouse-axon result also leaves benefits for other cell types and imaging methods to be established. 6

The scientific value of the result lies in where the improvement occurs. Human verification is needed because residual errors can change the inferred circuitry. Preventing some of those errors before review can make expert time available for larger volumes or more detailed analysis. The relationship between an error count and labor nevertheless depends on the difficulty of finding and correcting the errors that remain.

Toward more targeted training data

The next opportunity is to make generation selective. The researchers identify controllable properties such as length, spatial extent, and branching as ways to explore neuronal shape. They propose directing generation toward geometries that are particularly difficult to reconstruct, rather than relying only on a broad assortment of samples. That remains a research direction beyond the reported augmentation result. 7

This suggests a feedback loop between reconstruction and training: identify recurring failures in real tissue, generate related geometries, and test whether the additional examples reduce those failures on held-out specimens. Such a loop would need to preserve uncommon but valid biological structures. A shape classifier that becomes too restrictive could reject the very variation a connectome is meant to reveal.

The released MoGen project provides a demonstration notebook and pretrained checkpoints for fruit fly, mouse, and zebra finch data. These resources make the generator available for further investigation; their availability should be distinguished from validation of a complete reconstruction pipeline in each species. 8

MoGen points to a specific role for generative AI in scientific measurement: learning useful structural variation from verified observations, then testing whether that variation improves the interpretation of new observations. In connectomics, progress on that question can help turn more of the tissue already visible under the microscope into circuits that researchers can study.

References

1. Michał Januszewski and Viren Jain. A connectomics milestone: Mapping the complete male fruit fly brain. Google Research, September 3, 2026. Scale, scope, and verification of the male central nervous system reconstruction.
2. Michał Januszewski et al. High-precision automated reconstruction of neurons with flood-filling networks. Nature Methods 15, 605–610, 2018. Iterative segmentation and path-length evaluation.
3. Michał Januszewski, Thomas Templier, Kenneth J. Hayworth, David Peale, and Harald Hess. Accelerating neuron reconstruction with PATHFINDER. bioRxiv, 2025; doi:10.1101/2025.05.16.654254. Reconstruction through fragment assembly and shape evaluation; preprint abstract hosted by HHMI Janelia.
4. Yaron Lipman et al. Flow Matching for Generative Modeling. ICLR 2023; arXiv:2210.02747. Learning vector fields for continuous generative transformations.
5. Zixuan Huang et al. PointInfinity: Resolution-Invariant Point Diffusion Models. CVPR 2024, 10050–10060. Point-cloud generation using separate data and latent streams.
6. Franz Rieger, Jan-Matthis Lueckmann, Viren Jain, and Michał Januszewski. MoGen: Detailed Neuronal Morphology Generation via Point Cloud Flow Matching. ICLR 2026. Official publication abstract describing the architecture, evaluation, and downstream result; conference record.
7. Michał Januszewski and Franz Rieger. AI-generated synthetic neurons speed up brain mapping. Google Research, April 16, 2026. Training-set details, synthetic-data mixture, reconstruction results, and proposed extensions.
8. Google Research. MoGen project and model release. 2026. Demonstration notebook and species-specific pretrained checkpoints.