RSR-NF: Neural Field Regularization by Static Restoration Priors for Computed Dynamic Imaging

Berk Iskender, Yoram Bresler, Sushan Nakarmi, Nitin Daphalapurkar, Marc Louis Klasky

Abstract: In dynamic imaging a spatio-temporal object is reconstructed at all times using its undersampled measurements. In particular, in dynamic computed tomography (dCT), only a single projection at one view angle is available at a time, making the inverse problem very challenging. Moreover, because ground-truth dynamic data is usually either unavailable or too scarce, supervised learning techniques are not applicable. To address this, we propose RSR-NF, which uses a neural field (NF) to represent the dynamic object and, using the Regularization-by-Denoising (RED) framework with a learned restoration operator, incorporates an additional readily trained static deep spatial prior. To optimize the variational objective, we use an efficient ADMM-based algorithm. We compare RSR-NF to three algorithms, demonstrating the improvements by combining the NF representation with static restoration priors, and over state-of-the-art dCT techniques.

External IDs:dblp:conf/mlsp/IskenderBNDK25