The intricate morphology of neurons has fascinated scientists since the dawn of neuroscience. Here we use recent techniques of deep learning to build a generative model for 3-d neural structures.This generative model can be used in the simulation of realistic neural structures and in the inference of neuronal structure from imaging techniques.
Neuronal structures can be approximated as trees in 3-d space. Each neuron is uniquely specified by the adjacency matrix of its tree (morphology) and the location of the tree's vertices in 3-d space. We train generative adversarial networks (GANs) to synthesize the the morphology and geometry of realistic neurons, whose joint statistics are obtained from a neuroanatomy database: neuromorpho.org.
Clone the repository.
$ git clone https://github.com/tree-gan/BonsaiNet