Novel time resolved optical and machine learning methods for label free biomedical imaging

Abstract

Recent advances in multiphoton microscopy are paving the way for better visualization and understanding of intracellular organisms, leading to powerful non-invasive methods for cell and tissue diagnosis. Mitochondrial disorder is a type of disorder that affects the cell's mitochondria and impacts the cell's functionality as a whole. Roughly, these mitochondrial disorders affect between 1 in 6000 and 1 in 8000 live births. These disorders are hard to diagnose as each individual is affected differently by these disorders. Current tools for diagnosis of mitochondrial disorders are either invasive or time consuming in nature. The focus of this dissertation is to develop novel nonlinear optical imaging techniques and algorithms for studying mitochondrial heteroplasmy, one of the markers of mitochondrial disorder where the cells or tissues contain a mixture of healthy and diseased mitochondria. The defects in Electron Transport Chain (ETC) are often the causes behind various mitochondrial disorders. This dissertation aims to develop non-invasive optical methods for studying redox states of Cytochromes in living human cells such as Fibroblasts. Cytochromes are a particular type of Hemeprotein that are comprised of iron porphyrin and are primarily responsible for electron transport within the ETC of Mitochondria. We utilize a two-color visible wavelength pump-probe technique that can roughly detect up to 10-6 absorption changes of the probe beam arising from the approximately 10-15 μm thick live human Fibroblast samples. We detail the technical challenges faced in getting the optical instrument sensitivity to the required level so that Fibroblasts can be imaged with sufficient Signal-to-Noise Ratio (SNR). The major technical challenges that had to be overcome to obtain reasonable pump-probe signal levels from fibroblasts are: (1) minimizing the noise floor of the entire optical instrument so that we are performing shot-noise limited detection using a photodiode; (2) correcting for the focal plane mismatch arising due to chromatic aberration in the axial direction. Additionally, we explore machine learning based algorithms for better data interpretability and visualization of Transient Absorption data. We demonstrate that Hyperspectral Autoencoders, a specific type of Convolutional Neural Networks (CNNs), are able to unmix spectral signatures arising from different molecules in Transient Absorption images of reduced and oxidized muscle fibers. We also devise a novel loss function that takes into account the correlation between different channels of the input image to train the neural network so that the inherent randomness of gradient descent algorithms does not impact the network predictions each time. Furthermore, we explore image-to-image translation neural network algorithms to investigate the possibility of translating Reflectance Confocal Microscopy (RCM) images to Second Harmonic Generation (SHG) images. For this study, we collect co-registered RCM and SHG images from canine oral mucosa muscle tissue with the same Field of View (FOV) and excitation objective. We explore various state-of-the-art deep learning algorithms and evaluate their performance in transforming the RCM modality to SHG modality.