Microscopy Image Volume Synthesis and Segmentation Using Machine Learning

<p dir="ltr">Three-dimensional fluorescent microscopy volumes of tissues provide important information for understanding their structure. Segmentation, a critical step in analyzing cells in these tissues, is typically performed on cell nuclei instead of the cell membrane because of a lack of reliable chemical dyes that can be used to chemically stain cell membranes for imaging. Automated machine learning approaches are necessary for segmenting cell nuclei in large, complex microscopy volumes. Machine learning techniques for segmentation require large amounts of training samples of microscopy volumes along with their corresponding ground truth annotations. Data augmentation techniques that generate additional training samples are necessary as manually annotating ground truth is labor intensive, time inefficient, and prone to human variability. In this thesis, a method to generate synthetic microscopy volumes with irregular or non-ellipsoidal shaped nuclei using generative adversarial networks is proposed. Experimental results show that a machine-learning based segmentation method trained with synthetic microscopy volumes with irregularly shaped nuclei outperforms the same method trained with synthetic volumes containing purely ellipsoidal shaped nuclei. This thesis also proposes a transformer-based nuclei segmentation method that minimizes the use of post-processing steps to obtain instance-level segmentation. Finally, this thesis also presents Distributed and Networked Analysis of Volumetric Image Data (DINAVID), an online system for visualization, segmentation, and quantitative analysis of large 3D microscopy volumes that biologists can use without worrying about managing computational resources.</p>