Complas 2025

COMPLAS 2025

Self-Supervised Deep Equilibrium Learning for Sparse-Angle CT Reconstruction

Bubba, Tatiana (University of Ferrara)
Santacesaria, Matteo (MaLGa Center, University of Genoa)
Sebastiani, Andrea (University of Modena and Reggio Emilia)

In session: 1202 - Inverse Problems: Theoretical Challenges, Numerical Solutions, and Cutting-Edge Applications in Biomedicine and Material Engineering

Please login to view abstract download link

Computed tomography (CT) image reconstruction is crucial, but traditional methods often struggle with limited data, like in sparse-angle CT. Deep learning offers solutions, but usually needs "ground truth" images for training, which are hard to get, especially in medical imaging. We introduce TomoSelfDEQ, a new self-supervised deep learning approach that trains directly on the limited, undersampled CT measurements. We prove theoretically, and demonstrate experimentally, that TomoSelfDEQ works as well as fully-supervised methods (which need ground truth), even with very sparse data – as few as 16 projection angles. Our method outperforms existing self-supervised techniques, achieving state-of-the-art results.