TomoSelfDEQ: a Self-Supervised DEQ approach for Sparse-Angle CT

Title

TomoSelfDEQ: a Self-Supervised DEQ approach for Sparse-Angle CT

Speaker

Andrea Sebastiani - University of Modena and Reggio Emilia - Unimore

Abstract

Most Deep Learning approaches for solving imaging inverse problems require a significant amount of training data. In medical applications, such as computed tomography (CT), acquiring data, including ground truth images, presents significant challenges due to radiation safety concerns and data regulation policies. Self-supervised techniques emerged as promising alternatives that exploit the acquired measurements, without requiring paired ground truth data. In this talk, we present a self-supervised Deep Equilibrium (DEQ) framework for sparse-angle CT reconstruction that performs training using only undersampled measurements. The proposed model, TomoSelfDEQ, extends the SelfDEQ method, previously developed for undersampled MRI problems, to non-unitary operators, addressing a key limitation in medical imaging applications. We establish theoretical guarantees demonstrating that, under suitable assumptions, the self-supervised updates match those of fully-supervised training with a modified loss, that includes the forward operator (in our setting the CT forward map). Numerical experiments validate our theoretical finding, showing that TomoSelfDEQ achieves state-of-the-art results even in severely limited-angle scenarios. These results demonstrate the potential of this approach for practical CT reconstruction applications where training data is scarce or expensive to obtain.

Bio

Andrea Sebastiani is a Research Fellow in the Department of Mathematics at the University of Modena and Reggio Emilia. He obtained his PhD in Mathematics at the University of Bologna in 2024. His research focuses on the intersection of deep learning and computational imaging, particularly solving inverse problems in medical imaging through physics-informed machine learning techniques that combine model-based and data-driven approaches to enhance algorithm reliability and interpretability.

When

Thursday, May 29th, 12:00

Where

Room 326, UniGe DIBRIS/DIMA, Via Dodecaneso 35