Computational Imaging for Phase Retrieval and Biomedical Applications

Citation

Shen, Cheng (2023) Computational Imaging for Phase Retrieval and Biomedical Applications. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/pahb-cx81. https://resolver.caltech.edu/CaltechTHESIS:05272023-062623589

Abstract

In conventional imaging, optimizing hardware is prioritized to enhance image quality directly. Digital signal processing is viewed as supplementary. Computational imaging intentionally distorts images through modulation schemes in illumination or sensing. Then its reconstruction algorithms extract desired object information from raw data afterwards. Co-designing hardware and algorithms reduces demands on hardware and achieves the same or even better image quality. Algorithm design is at the heart of computational imaging, with model-based inverse problem or data-driven deep learning methods as approaches. This thesis presents research work from both perspectives, with a primary focus on the phase retrieval issue in computational microscopy and the application of deep learning techniques to address biomedical imaging challenges.

The first half of the thesis begins with Fourier ptychography, which was employed to overcome chromatic aberration problems in multispectral imaging. Then, we proposed a novel computational coherent imaging modality based on Kramers-Kronig relations, aiming to replace Fourier ptychography as a non-iterative method. While this approach showed promise, it lacks certain essential characteristics of the original Fourier ptychography. To address this limitation, we introduced two additional algorithms to form a whole package scheme. Through comprehensive evaluation, we demonstrated that the combined scheme outperforms Fourier ptychography in achieving high-resolution, large field-of-view, aberration-free coherent imaging.

The second half of the thesis shifts focus to deep-learning-based methods. In one project, we optimized the scanning strategy and image processing pipeline of an epifluorescence microscope to address focus issues. Additionally, we leveraged deep-learning-based object detection models to automate cell analysis tasks. In another project, we predicted the polarity status of mouse embryos from bright field images using adapted deep learning models. These findings highlight the capability of computational imaging to automate labor-intensive processes, and even outperform humans in challenging tasks.

Thesis Files