CellMincer: Self-Supervised Denoising of Functional Imaging

Abstract

All-optical electrophysiology offers accessibility and scalability in observing neuronal activity beyond what can feasibly be achieved with patch clamp techniques. However, imaging platforms like Optopatch suffer from excessive detection noise, photobleaching, and an inability to organically segment and isolate neurons of interest. These drawbacks preclude its use as a true substitute for direct electrophysiological measurement, but recent advances in deep neural network inference may enable computation to recover the difference in data quality. To date, few robust denoising algorithms have been designed and implemented for voltage imaging data, in part because the lack of ground truth imaging complicates the task of training such a model. This thesis introduces CellMincer, a self-supervised deep neural network for denoising functional imaging. By exploiting a combination of spatiotemporally local contexts and precomputed global features, CellMincer outperforms comparable algorithms at denoising several modes of optical electrophysiology on a range of metrics, including measures of biologically relevant features.