Generating Realistic Morphologies of Neurons in Rodent Hippocampus with DGCAN

Abstract

Dendritic size and branching patterns are important features of neural form and function. However, current computational models of neuronal networks use simplistic cylindrical geometries to mimic dendritic arborizations. One reason for this is that current methods to generate dendritic trees have rigid a priori constraints. To address this, a deep convolutional generative adversarial network (DCGAN) trained on images of rodent hippocampal granule and pyramidal dendritic trees. Image features learned by the network were used to generate realistic dendritic morphologies. Results show that DCGANs achieved greater stability* and high generalization, as quantified by kernel maximum mean discrepancy, when exposed to instance noise and/or label smoothing during training. Trained models successfully generated realistic morphologies for both neuron types, with high false positive rate reported by expert reviewers. Collectively, DCGANs offer a unique opportunity to advance the geometry of neural modeling, and, therefore, to propel our understanding of neuronal function.\n\n * A \"stable/stabilized DCGAN\", as mentioned throughout this work, is a DCGAN which was stable throughout training.