Google Research: AI generates synthetic neurons and saves 157 person-years in brain mapping

Google Research published research on the MoGen (Neuronal Morphology Generation) system that uses AI to create synthetic neuron shapes designed to accelerate brain mapping. The results show that synthetic neurons can dramatically reduce the need for manual labor by neuroscientists.

How does MoGen create synthetic neurons?

MoGen uses the PointInfinity point cloud flow matching model — an architecture that learns 3D shapes from point clouds. The model was trained on just 1,795 manually verified mouse axons, which is a relatively small dataset in the context of neuroscience. Despite this, the generated synthetic neurons are indistinguishable from real ones according to assessments by expert evaluators.

The key innovation is in the approach: instead of classical data augmentation (rotations, reflections), MoGen generates entirely new neuron shapes that follow the real distribution of morphology but do not copy existing examples.

How much time does this save?

The numbers are impressive. Just 10% of synthetic data added to the training set reduces the error rate by 4.4%. Translated to the full scale of mouse brain mapping, this is equivalent to saving 157 person-years of manual work on neuron annotation.

Manual neuron mapping — tracing axons and dendrites through electron microscope images — is one of the biggest bottlenecks in neuroscience. A single cubic millimeter of brain can contain thousands of neurons whose shapes need to be reconstructed, and a human annotator can spend days on a single neuron.

What does this open up for the future?

Google has released MoGen as an open-source project, enabling neuroscientists around the world to use the system in their own research. The current focus is on the mouse brain, but the authors emphasize that the approach opens the path toward mapping mammalian and eventually human brains.

The research is particularly significant because it shows how AI can solve a problem that is not just about computing speed, but about creating quality training data — synthetic neurons help other AI models recognize real neurons better.