Score based method estimates imaging limits in high dimensional spaces
Latent Score-Based Bayesian Cramér-Rao Bound Estimation for High-Dimensional Imaging Systems
Information TheoryMachine Learning
Summary
High-quality imaging systems often have many unknowns, making it hard to understand the best possible accuracy for reconstructing images. The authors developed a way to estimate this accuracy limit, called the Bayesian Cramér-Rao bound, by simplifying the problem into a smaller set of components using a neural network called a variational autoencoder. They also improved how to learn complex image properties using a new method that reduces weird artifacts in the results. This approach was tested on a large-scale simulated breast imaging problem, showing it can estimate difficult limits accurately and stably.
What this means in practice
- •For medical imaging engineers: Evaluate theoretical limits on image reconstruction accuracy for complex breast imaging systems to compare and design better hardware and algorithms.
- •For computational photography developers: Estimate fundamental noise and resolution boundaries in high-dimensional image recovery problems to improve designs of cameras and sensors.
Authors
Evan Scope Crafts, Thomas Wynn, Seonyeong Park, Mark Anastasio, Umberto Villa
Abstract
We propose a data-driven framework for estimating the Bayesian Cramér-Rao bound (CRB) in high-dimensional imaging systems with complex, analytically intractable priors. Direct CRB computation is challenging in this setting due to the need to model the prior score and to form and invert the Bayesian Fisher information matrix in very high dimensions. To address these issues, we first reformulate the inverse problem in the latent space of a pre-trained variational autoencoder, thereby dramatically reducing the dimensionality of the bound estimation problem while preserving the spatial structure of the images. The Bayesian CRB is formed in this latent space and mapped back to the native parameter space using a change-of-variables formula. Second, to learn the latent prior score, we introduce a new sliced score matching objective defined in a Bochner space endowed with an $H^1(Ω)$ Sobolev norm in space. This "Bochner-space sliced score matching" objective is consistent with standard sliced score matching, but penalizes errors in the spatial gradients of the score, suppressing high-frequency artifacts that otherwise contaminate the resulting CRB estimates. We validate the approach on a stylized quantitative photoacoustic computed tomography (qPACT) breast imaging problem with over one million unknown parameters, using a foundation-model autoencoder derived from Stable Diffusion. The proposed method yields stable, artifact-free Bayesian CRB estimates that reflect the highly non-Gaussian structure of the learned prior and reveal the substantial impact of the prior on the relative performance of competing qPACT design schemes.