Overcoming Spatiotemporal Trade-offs in Calcium Imaging Using Deep Learning-based Dynamics Modeling Público

Abstract

Recent advances in neural interfaces have enabled monitoring of the activity of increasingly large neuronal populations. Among these techniques, two-photon (2p) calcium imaging is a powerful tool to probe how population-level computations relate to biological structure, as it can identify layers and cell types of interest. However, extracting fast patterns of neural activity from 2p data has proven challenging because of limitations on temporal resolution imposed by the spatiotemporal trade-offs inherent to 2p laser scanning. Noises and nonlinearities introduce additional challenges to analyzing 2p data. This dissertation bridges this gap by taking a dynamical systems approach to denoise and improve temporal resolution of 2p data. In Chapter 1, we provide a broad introduction to the challenges with 2p data and methods for modeling neural dynamics, and discuss the benefits of modeling dynamics from 2p data. In Chapter 2, we present a novel neural network training strategy that offers a principled solution to spatiotemporal trade-offs created by bandwidth limits in neural interfaces. This strategy enables inference of latent dynamics with spatiotemporal super resolution and is applicable for a wide range of neural interfaces. In Chapter 3, we detail the results of extending a state-of-the-art deep learning method that models neural dynamics in spiking activity for application to 2p imaging. We demonstrate that our new method outperforms standard methods in recovering high-frequency components in synthetic tests and predicting single-trial behaviors in 2p recordings from sensorimotor areas in mice performing a forelimb reach task. In Chapter 4, we present machine learning innovations that eliminate the need of deconvolution as a preprocessing step for our approach. This opens the door to modeling fast and complex dynamics from 2p data in settings where massive populations of neurons are imaged with extremely slow sampling rates as a trade-off. In sum, our work provides an avenue to overcome the limits of spatiotemporal trade-offs in 2p calcium imaging, enabling accurate inference of population dynamics across a wide range of sampling rates in vast populations with identified neurons.

Table of Contents

Chapter 1: Introduction: 1

Chapter 2: Deep inference of latent dynamics with spatio-temporal super-resolution using selective backpropagation through time: 14

Chapter 3: A deep learning framework for inference of single-trial neural population dynamics from calcium imaging with sub-frame temporal resolution: 34

Chapter 4: Inferring fast structures from calcium imaging at slow sampling rates using deconvolution-free dynamics modeling: 78

Chapter 5: Dissertation summary and future directions: 99

About this Dissertation

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Laney Graduate School
Department	Biological and Biomedical Sciences
Degree	Ph.D.
Submission	Dissertation
Language	English
Research Field	Biology, Neuroscience Engineering, Biomedical Artificial Intelligence
Palavra-chave	Spatiotemporal trade-offs Computational Neuroscience Neural Dynamics Machine Learning Calcium Imaging
Committee Chair / Thesis Advisor	Chethan Pandarinath, Emory University
Committee Members	Eva Dyer, Georgia Tech Annabelle Singer, Emory University Gordon Berman, Emory University Samuel Sober, Emory University

Última modificação

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Overcoming Spatiotemporal Trade-offs in Calcium Imaging Using Deep Learning-based Dynamics Modeling ()	2022-12-12 22:31:35 -0500	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions