Computer Vision Methods for Fast and Automated Processing of C. elegans Whole-Brain Videos

How our brain generates complex behaviors, while adapting, learning and reconfiguring its activity in response to environment and internal states, is a fundamental question in neuroscience. Uncovering circuit mechanisms governing behavior requires simultaneous recording of neuron activities from large and spatially separated brain areas, thus motivating the development of whole-brain fluorescent imaging of neuron activities. C. elegans, an optically transparent nematode, with compact nervous system of 302 neurons, and many genetic tools available, is suitable for whole-brain imaging. Several recent works have established the microscopic setups and platforms for single cell resolution whole-brain imaging. However, there are still several challenges, both for data acquisition and data processing, for wide adaptation of these methods. On the acquisition side, tradeoffs between signal-to-noise ratio (SNR) in images and imaging conditions must be made. For instance, low exposure times and low excitation intensity are critical to achieve fast volumetric imaging rates, and to prevent photobleaching of fluorophores, but this produces low SNR. In contrast, optimizing for higher SNR prevents long term recordings necessary to study long timescale behaviors. On the data processing side, there are several challenging tasks that must be performed to extract high quality, annotated neuron activity traces from videos: cell detection, tracking, calcium trace extraction, and cell identification. All these tasks are challenging due to low SNR in images, dense packing of cell nuclei in head ganglion, and non-rigid deformations of the head. Further, errors generated in cell detection task such as missed detections or false detections significantly affect the accuracy of tracking and identification. Due to these limitations, the rate at which whole-brain imaging data can be collected has surpassed the rate at which meaningful neuron activity traces can be extracted from the datasets. In this thesis, I present methods to address the technical challenges in both whole-brain imaging data acquisition and data processing. Chapter 2 presents an easy to train and fast deep learning based denoising framework that enables to overcome tradeoffs between imaging parameters and signal to noise ratio in images, thus making high quality data acquisition easier with commonly available microscopy setups. Chapters 3 and 4 present methods to overcome two critical challenges in processing whole-brain video stacks, namely cell detection and cell tracking. I present an accurate, robust, and easy to train deep learning framework for 3D segmentation of cell nuclei in images using only 2D annotated data. I also demonstrate an objective screening approach for identification of optimal tracking strategy of nuclei in whole-brain videos. Chapter 5 presents a graphical model framework to solve the final challenge in processing whole-brain videos, i.e. cell identity annotation. I demonstrate higher accuracy and robustness of our methods compared to previous methods. I also show cell identity annotation in several applications such as gene expression pattern analysis, multi-cell imaging and whole-brain imaging. The methods introduced in this work establish an end-to-end workflow for whole-brain imaging data acquisition and processing. These methods will enable generation of high quality and identity annotated neuron activity in data in a high throughput manner. Increased number of datasets will enable researchers to draw statistically powerful conclusions. Further large sets of data will encourage development of newer computational and machine learning methods to discover hidden trends and properties in data.

Date

2022-08-05

Resource Type

Text

Resource Subtype

Dissertation

Full item page