Date of Award
Doctor of Philosophy (PhD)
This dissertation addresses model-based deep learning for computational imaging. The motivation of our work is driven by the increasing interests in the combination of imaging model, which provides data-consistency guarantees to the observed measurements, and deep learning, which provides advanced prior modeling driven by data. Following this idea, we develop multiple algorithms by integrating the classical model-based optimization and modern deep learning to enable efficient and reliable imaging. We demonstrate the performance of our algorithms by validating their performance on various imaging applications and providing rigorous theoretical analysis.The dissertation evaluates and extends three general frameworks, plug-and-play priors (PnP), regularized by denoising (RED) and deep unfolding (DU), all of which integrate model-based optimization and deep learning. PnP and RED adopt deep-learned denoisers as image priors inside model-based iterative algorithms, while DU interprets the iterations of a model-based algorithm as layers of a deep neural network and trains it end-to-end in a supervised fashion. We contribute to these research areas by 1) providing the statistical interpretation of the PnP algorithms through the analysis of the priors implicitly represented by denoisers; 2) proposing an incremental variant of the widely-used PnP-ADMM algorithm to handle problems involving large-scale measurements; 3) extending the family of PnP algorithms to the non-Euclidean setting based on the general Bregman distance; and 4) developing an end-to-end model-based learning framework for the estimation of quantitative maps from under-sampled, noisy and motion-corrupted MRI data.
Ulugbek S. Kamilov
Ulugbek S. Kamilov, Tao Ju, Netanel Raviv, Brendt Wohlberg,