Normalizing flow regularization for photoacoustic tomography

Abstract

Proper regularization is crucial in inverse problems to achieve high-quality reconstruction, even with an ill-conditioned measurement system. This is particularly true for three-dimensional photoacoustic tomography, which is computationally demanding and requires rapid scanning, often leading to incomplete measurements. Deep neural networks, known for their efficiency in handling big data, are anticipated to be adept at extracting underlying information from images sharing certain characteristics, such as specific types of natural or medical images. We introduce a Normalizing Flow Regularization (NFR) method designed to reconstruct images from incomplete and noisy measurements. The method involves training a normalizing flow network to understand the statistical distribution of sample images by mapping them to Gaussian distributions. This well-trained network then acts as a regularization tool within a Bayesian inversion framework. Additionally, we explore the concept of adaptive regularization selection, providing theoretical proof of its admissibility. A significant challenge in three-dimensional image training is the extensive memory and computation requirements. We address this by training the normalizing flow model using only small-size images and applying a patch-based model for reconstructing larger images. Our approach is model-independent, allowing the reuse of a well-trained network as regularization for various imaging systems. Moreover, as a data-driven prior, NFR effectively leverages the available dataset information, outperforming artificial priors. This advantage is demonstrated through numerical simulations of three-dimensional photoacoustic tomography under various conditions of sparsity, noise levels, and limited-view scenarios.