MoGen: synthetic neurons improve brain mapping AI, reducing errors by 4.4%

Google Research developed MoGen (Neuronal Morphology Generation), a generative AI model that produces realistic synthetic neuron shapes to improve brain mapping. The model uses PointInfinity point cloud flow matching to transform random 3D points into neuron geometries. When 10% synthetic data from MoGen was added to the training of the PATHFINDER reconstruction model, it reduced overall reconstruction errors by 4.4%, primarily by decreasing merge errors. While seemingly modest, this improvement translates to saving 157 person-years of manual proofreading at the scale of a complete mouse brain. MoGen was trained on 1,795 verified mouse cortex axons and validated by human experts. The researchers also trained versions for zebra finch and fruit fly neurons. The model and trained weights are released as open-source.