UC Berkeley

Computational imaging involves simultaneously designing optical hardware and reconstruction software. Such a co-design framework brings together the best of both worlds for an imaging system. The goal is to develop a high-speed, high-resolution, and large field-of-view microscope that can detect 3D fluorescence signals from single image acquisition. To achieve this goal, I propose a new method called Fourier DiffuserScope, a single-shot 3D fluorescent microscope that uses a phase mask (i.e., a diffuser with random microlenses) in the Fourier plane to encode 3D information, then computationally reconstructs the volume by solving a sparsity-constrained inverse problem.

In this dissertation, I will discuss the design principles of the Fourier DiffuserScope from three perspectives: first-principles optics, compressed sensing theory, and physics-based machine learning. First, in the heuristic design, the phase mask consists of randomly placed microlenses with varying focal lengths; the random positions provide a larger field-of-view compared to a conventional microlens array, and the diverse focal lengths improve the axial depth range. I then build an experimental system that achieves <3 um lateral and 4 um axial resolution over a 1000x1000x280 um^3 volume. Lastly, we use a differentiable forward model of Fourier DiffuserScope in conjunction with a differentiable reconstruction algorithm to jointly optimize both the phase mask surface profile and the reconstruction parameters. We validate our method in 2D and 3D single-shot imaging, where the optimized diffuser demonstrates improved reconstruction quality compared to previous heuristic designs.