Abstract

Computational imaging is a pivotal field that synergizes physical measurement principles with advanced algorithms to generate visual information. An important task in this field is solving imaging inverse problems that aim to reconstruct high-quality images from observed measurements. Model-based deep learning (MBDL) has emerged as a particularly powerful tool for tackling these inverse problems by integrating machine learning (ML)-driven priors with knowledge of the imaging physics. This dissertation focuses on the pervasive challenge of informational incompleteness in computational imaging, arising from various practical and physical limitations, that hinder the widespread adoption of ML-driven computational imaging algorithms in practice. This includes: (a) the incompleteness of datasets that lack ideal ground-truth references for training machine learning models (Part II); (b) incomplete knowledge of the imaging system's physical forward model (Part III); and (c) the incompleteness of acquired data in multimodal imaging scenarios (Part IV). To address dataset incompleteness, this dissertation pioneers various self-supervised algorithmic frameworks. These include methods that leverage inter-measurement consistency for dynamic imaging (Phase2Phase, Chapter 3), employ deformation-compensated learning for motion-affected scenarios (DeCoLearn, Chapter 4), and establish provable techniques for training advanced implicit neural networks without reference ground truth (SelfDEQ, Chapter 5). To address incomplete forward model knowledge, various MBDL techniques are introduced for blind inverse problems that aim to estimate both images and forward model unknowns. These feature block-coordinate plug-and-play algorithms with convergence guarantees that integrate deep denoiser priors on all unknowns (BCPnP, Chapter 6), and self-supervised deep unrolling strategies for concurrent system calibration and image reconstruction in parallel MRI (SPICER, Chapter 7). For incompleteness in acquired multimodal data, unified diffusion models are proposed for diverse reconstruction and synthesis tasks from arbitrary combinations of available input modalities (Any2All, Chapter 8). Furthermore, this thesis demonstrates the broad applicability of MBDL through its successful validation in challenging biomedical applications (Part V), including deformable image registration (PIRATE, Chapter 9) and fast ptychographic image reconstruction (PtychoDV, Chapter 10). Collectively, this dissertation significantly enhances the capacity of computational imaging to function reliably, practically, and effectively under diverse conditions of informational incompleteness through innovations in algorithm design, theoretical understanding, and practical application, paving the way for more powerful imaging solutions.

Committee Chair

Tao Ju

Committee Members

Brendt Wohlberg; Hongyu An; Nathan Jacobs; Ulugbek Kamilov

Degree

Doctor of Philosophy (PhD)

Author's Department

Computer Science & Engineering

Author's School

McKelvey School of Engineering

Document Type

Dissertation

Date of Award

8-18-2025

Language

English (en)

Download

Share

COinS