PhD Thesis Defense

A typical imaging scenario involves illumination, the object to be imaged, an optical system, and a sensor. In conventional imaging, significant efforts have been dedicated to optimizing hardware in order to directly improve the quality of images captured by the sensor. In this paradigm, signal processing algorithms are typically viewed as a secondary method used to enhance image quality after measurement. However, computational imaging takes a different approach by intentionally distorting measured images through illumination modulation, optical modulation, or special sensing schemes. When combined with carefully designed reconstruction algorithms, computational imaging allows us to recover the desired object information. By co-designing both the hardware and algorithm, we can leverage the computational power to alleviate the demands on hardware, enabling us to achieve comparable image quality with more cost-effective systems or even accomplish previously unattainable imaging goals.

The design of algorithms lies at the core of computational imaging, with two primary approaches being model-based inverse problems and data-driven deep learning methods. In this thesis talk, I will present my research work from both perspectives, focusing on the phase retrieval issue in computational microscopy and the application of deep learning techniques to solve biomedical problems. For the model-based method section, I will begin with a project aimed at correcting aberrations in a multispectral microscope using a computational coherent imaging modality called Fourier ptychography (FP), which was developed in our lab. Despite the success of FP, it still has certain limitations. Therefore, I will propose a novel computational coherent imaging modality based on Kramers-Kronig relations and explain how it, along with its extended version, surpasses FP in various aspects. In the latter part of the talk, I will introduce two separate projects that address challenging technical problems in liquid biopsy and developmental biology, respectively, by utilizing deep learning methods. Through these examples, we can observe how deep learning has the potential to replace human labor and tackle tasks that even prove difficult for human experts.

Whether employing a model-based or data-driven approach, computational imaging has transformed our perception of how to design the entire imaging pipeline to effectively capture object information. The field of computational imaging is continuously evolving, and this thesis talk aims to highlight successful case studies while providing insights into the design of such systems.