Haejun Han

BioE PhD Defense Presentation

Time and Date: 2:00 pm, Wednesday, April 30th, 2025.

Location: EBB CHOA

Zoom Link: https://gatech.zoom.us/j/7702113243?pwd=ZUVpVUFJS1RXYzdiMVg2dktTbFhiZz09

Meeting ID: 770 211 3243

Passcode: eJjAV9

Advisor: Hang Lu, Ph.D. (School of Chemical and Biomolecular Engineering, Georgia Tech)

Committee Members:

Oliver Hobert, Ph.D. (Department of Biological Sciences, Columbia University)

Patrick McGrath, Ph.D. (School of Biological Sciences, Georgia Tech)

Eva Dyer, Ph.D. (Wallace H. Coulter Department of Biomedical Engineering, Georgia Tech)

Ghassan AlRegib, Ph.D. (School of Electrical and Computer Engineering, Georgia Tech)

Novel Computational Methods for Scalable C. elegans Neuroimaging Analysis

      Deciphering the relationship between neural circuit structure, function, and behavior requires powerful methods for observing and analyzing neural dynamics at scale, posing a fundamental challenge in neuroscience. The nematode Caenorhabditis elegans, with its fully mapped nervous system and powerful genetic tools, provides an exceptional model for dissecting these mechanisms, especially when combined with fluorescence microscopy to visualize neural structure and function in vivo. However, realizing the full potential of this approach is often limited by critical bottlenecks in data analysis throughput, inherent trade-offs in imaging parameters like speed and resolution, and the prohibitive labor cost associated with manually annotating large, dynamic datasets.

      This thesis focuses on developing novel computational tools, leveraging computer vision and machine learning, to overcome these specific limitations in C. elegans fluorescence neuroimaging. First, to address the challenge of quantifying synaptic structures at scale, I developed WormPsyQi, an automated pipeline using machine learning for robust, high-throughput synapse segmentation and analysis across diverse reporter types (Aim 1). Second, to mitigate imaging trade-offs that compromise functional studies, I created Deep Video DeBinning (DVDB), a semi-supervised video super-resolution method that computationally restores spatial detail lost to camera binning, enabling high-fidelity activity analysis from high-speed, high-SNR recordings for applications like studying rapid locomotion dynamics or performing volumetric whole-brain imaging (Aim 2). Third, confronting the annotation bottleneck in analyzing large volumetric neural activity datasets, I designed ASCENT, a self-supervised deep learning framework utilizing a Neuron Embedding Transformer (NETr) to achieve state-of-the-art 3D neuron tracking accuracy without requiring manual annotations (Aim 3).

      Together, these computational tools provide validated, scalable solutions that enhance our ability to perform quantitative analysis of neural circuit structure and function in C. elegans. By automating key analysis steps and computationally overcoming imaging limitations, this work lays the groundwork for deeper investigations into the structure, function, and dynamics of neural circuits in this powerful model system.