The application of Holo-UNet on Biomedical Imaging

Abstract

Computational tools in optical bioimaging have powered discovery of new biological processes. The primarily function of the different computational approaches in biological imaging can be broadly placed into 3 categories, denoising, retrieval/reconstruction, segmentation and clustering. In the last few decades, the hologram-based optical microscopy has made its way from specialised optics laboratory to commercial turnkey imaging solutions (LynceeTec, Holotomo, PHI, Nanolive), of which many are deployed in standard cell biology laboratories to study single cell biology. Holographic optical microscopy has improved rapid identification and track living cells colonies and individual cell morphologies without chemical labelling at sub-cellular resolution. The crux of holographic-based optical bioimaging is the computational retrieval of quantitative information from optical holograms. In this thesis, we deal with specifically with off-axis digital holography. Optical holograms are created through the coherent addition of optical fields that are prone to errors arising such as (1) wavefront aberration, (2) non-stochastic noises from sample and environment in real time reconstruction, (3) automated digital refocusing. While the focus of thesis is on the implementation of machine learning techniques into the workflow of a holographic microscope, the eventual outcome is a complete industry-grade optical hologram processing platform to broaden the wider adoption of Machine Learning for holographic live cell imaging. To rigorously test the industry readiness of computational pipeline, the hologram processing platform was tested with over 15 different user groups of various training for over an entire year. Specifically, we develop a ML integrated workflow for the counting of living cells that regularly cultured for over several months, and document an experiment pipeline that can be used by non-optics experts to easily perform a wide range experiments on holographic microscopy systems using the software platform and ML model developed in this thesis. In chapter 1, I first review state of art holographic microscopy (broadly termed as quantitative phase microscope - QPM) techniques and followed by a detail analysis of the respective imaging challenges such as noise. Chapter 2 reviews existing ML approaches in bioimaging with QPM. Chapter 3 details the design of U-NET for QPM holograms, which we term as Holo-UNet and demonstrates its use-case for denoising low-light holograms using fixed L929 fibroblast cells as an example. Chapter 4 details the design and development of QPM software that support Holo-UNet as well as broader QPM imaging applications. Chapter 5 discusses the integration of Holo-UNet and the software platform into a single workflow for cell counting, and investigate the performance of the model on living fibroblast cells in culture, that differs from the fixed fibroblast cells used in Chapter 3. The use of living cell cultures as samples permits us to accurately test Holo-UNet performance in real biological conditions and better understand the role of environment factors (humidity, temperature) that can impact the performance of Holo-UNet. Show more