Flow imaging microscopy (FIM) is an increasingly popular technique for collecting light microscopy images of particles larger than 1 μm in samples such as therapeutic protein formulations. While FIM is commonly used to monitor the number of particles in a sample, the particle images returned by FIM contain particle morphology information that may be used to distinguish between different types of particles. For example, this morphology information may be used to differentiate between protein aggregates formed by different mechanisms or identify cells of different species. However, this analysis has previously been prohibited by the difficulty of extracting and analyzing relevant particle morphology information from FIM images.

This thesis describes the development of algorithms that use statistical and machine learning methods to analyze particle morphology information in FIM images and applications of these algorithms in protein formulation development. Several algorithms were developed that use convolutional neural networks (ConvNets) and other approaches to analyze the morphology of both individual particles as well as the particle population in a sample. These algorithms were then used to identify the impact of conditions such as accelerated and real-time stability stresses on protein aggregate morphology. This thesis will also describe generalizations of this approach for analyzing light and fluorescence microscopy images collected from imaging flow cytometry (IFC) and for distinguishing between blood cells and cells of different bacterial species as proof-of-concept for a FIM-based bloodstream infection (BSI) diagnostic test. These results demonstrate the effectiveness and usefulness of FIM-based particle morphology analysis techniques for analyzing particle morphology in therapeutic protein formulations, blood samples, and other particle-containing samples.