Inferring 3D dendritic voltage from 2D voltage movies

Dendrites perform non-linear processing of synaptic inputs, and drive active processes such as plateau potentials which are critical to computation and plasticity. Therefore, the ability to image voltage across an entire dendritic tree would represent a major advance in systems neuroscience. Recent advances in genetically encoded voltage indicators (GEVIs) have brought this possibility closer to reality. However, due to the fundamental tradeoff between imaging speed and volume, it has been impossible thus far to image an entire dendritic tree using volumetric imaging at high speeds. To get around this tradeoff, we designed a computational method which fuses 3D anatomical information and 2D voltage imaging, yielding an estimate of the voltage at each point along the tree for each time-step. Our method comprises two steps. In step one, we use the anatomical data to build a microscope model which maps from voltages along the tree to observed fluorescence at the imaging plane. By exploiting the spatial smoothness of the voltage signal, we can approximately invert this model to obtain noisy voltage estimates at each point on the tree. Second, we train a self-supervised neural network to perform spatiotemporal denoising of the inferred voltages. Whereas conventional approaches only allow analyzing voltage for in-focus segments, our technique resolves both in-focus and out-of-focus branches, creating the potential to analyze an entire dendritic tree at single-trial resolution. On simulated voltage movies created via multi-compartment models, we find that we can robustly recover voltage for out-of-focus branches. On real voltage imaging data from CA1 pyramidal neurons we recover known dynamical motifs of voltage propagation. This method represents a crucial step towards accurate recovery of voltage throughout an entire dendritic tree, and will enable fitting of detailed biophysical models and improved understanding of dendritic computation.