Energy-Based Generative Models for Inverse Problems in Medical Imaging

Description

In the past decades, Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) have established themselves as the primary imaging methods in medicine.
Variational reconstruction of these methods still rely on hand-crafted regularizers or patch-based approaches, which break down in cases of severe undersampling.
In this work, we introduce a novel reconstruction scheme, where a parametrized regularizer with global field-of-view is learned on data using maximum likelihood.
Our variational formulation improves interpretability when compared to feed-forward approaches, by allowing experts to judge the full posterior distribution of any reconstruction problem.
We show how our regularizer is capable of generating realistic images \emph{without} any data, and apply it to limited-angle and (very) few-view CT tasks as well as accelerated MRI reconstruction, where it outperforms traditional reconstruction algorithms by a large margin.
Additionally, we analyze how posterior variance relates to unnatural reconstructions, as is the case for certain pathologies.

Period	21 Mar 2022
Event title	SIAM Conference on Imaging Science: IS 2022
Event type	Conference
Location	VirtualShow on map
Degree of Recognition	International