Using machine learning for analysis of neuronal network activity

Abstract

Analyzing neuronal activity in developing neuronal networks can improve our understanding of neuronal dysfunctions underlying conditions such as Rett syndrome. Two-photon calcium imaging is used to capture neuronal network activity over time. This method produces large sets of images that are typically manually analyzed by skilled neuroscientists. Because this process is both time-consuming and subject to error, discovery of therapies that ameliorate network dysfunction may be slowed. We improve an existing, semi-autonomous machine learning pipeline for two-photon calcium imaging sequence analysis. We introduce to the pipeline neuron detection methods using supervised learning models, heuristic filtering of pixels for signal extraction, and event detection using deconvolution. With these methods, we improve neuron detection performance, alter signal-to-noise ratio of extracted calcium signals, and allow for integration of methods that infer action potential firing underlying these signals.