Statistical and Deep Learning Frameworks for High Throughput Neuronal Signal and Image Processing

Abstract

Quantitative analysis of the central nervous system (CNS) - comprised of the brain, the spinal cord, and the eyes - for a deeper intuition into its function often requires in vivo visualization of its microscopic structures. For the brain, calcium imaging using genetically encoded calcium indicators (GECIs) allows targeted, large-scale imaging of neuronal populations with cellular resolution in animals. Combined with closed-loop optogenetic control of single cells, neuroscientists can potentially test population-based models of the underlying neuronal system, adding a significant body of knowledge to the field. Realization of a closed-loop optical neuronal control system currently lacks computational frameworks (e.g., neuron segmentation) that drive the system’s components based on recent data. Current neuron segmentation methods either require the acquisition of the full movie or are unable to reliably identify active neurons. On another front, in vivo visualization of retinal cells has become possible with the incorporation of adaptive optics (AO) into existing retinal imaging systems, such as optical coherence tomography (OCT). A complete morphometric analysis of the living human retina at cellular level could potentially improve diagnosis, treatment planning, and monitoring of retinal diseases. The current standard approach for quantifying ganglion cells (GCs; one of the fundamental cell types for vision) from AO-OCT volumes is manual, making the task highly subjective, time consuming, and thus not feasible for large-scale studies and clinical use.This dissertation describes the development of computational frameworks for accurate analysis of neurons from high-resolution optical images of the brain and the retina. In part 1, a statistical and information theoretic framework was developed for quantifying the resolution limit and the Cramer Rao lower bound (CRB) in detecting closely timed neuronal spikes from two-photon calcium imaging recordings. Monte-Carlo simulations with biologically derived parameters were used to numerically calculate the resolution limit and compare the performance of the optimal estimators with the CRB. Additionally, we applied our detector to distinguish overlapping transients from experimentally obtained calcium imaging data.In part 2, a fast and robust framework was developed to automatically segment active neurons from two-photon calcium imaging recordings. A convolutional neural network (CNN) is at the core of the framework which exploits the spatiotemporal information in the recorded movies. The method is validated using two separate online datasets and its performance is compared against other state-of-the-art techniques. In part 3, the focus is shifted to analyzing AO-OCT images of the human retina. We developed a weakly-supervised deep learning-based method to automatically segment GCs in the AO-OCT volumetric images. We validated the performance of our framework using images from healthy and glaucoma subjects acquired with two different imagers across various retinal locations and compared the performance with expert graders.In conclusion, this dissertation provides a set of statistical and deep learning frameworks for high throughput neuronal signal and image processing. Our modern computational frameworks can be used to rapidly and accurately parse neuronal activity from calcium imaging data and measure neuronal biomarkers for in vivo monitoring of retinal diseases. The presented automatic frameworks have comparable performance to human experts in detecting brain neurons and retinal GCs, which is important in the long-term goal to facilitate the monitoring of microscopic structures of the CNS through optimal quantification tools.