TY - JOUR AU - Kabri, Samira AU - Auras, Alexander AU - Riccio, Danilo AU - Bauermeister, Hartmut AU - Benning, Martin AU - Moeller, Michael AU - Burger, Martin TI - Convergent Data-driven Regularizations for CT Reconstruction JO - Communications on applied mathematics and computation VL - 6 IS - 2 SN - 2096-6385 CY - Singapore PB - Springer Singapore M1 - PUBDB-2023-06365 SP - 1342-1368 PY - 2024 N1 - Funding should be my DFG project BU 2327/19/-1 Grundlagen vollüberwachter Deep Learning Verfahren für Inverse ProblemePaper accepted, not yet published (written while SK and MB were still with FAU) AB - The reconstruction of images from their corresponding noisy Radontransform is a typical example of an ill-posed linear inverse problemas arising in the application of computerized tomography (CT). Asthe (naive) solution does not depend on the measured data continu-ously, regularization is needed to re-establish a continuous dependence.In this work, we investigate simple, but yet still provably convergentapproaches to learning linear regularization methods from data. Morespecifically, we analyze two approaches: One generic linear regularizationthat learns how to manipulate the singular values of the linear oper-ator in an extension of our previous work, and one tailored approachin the Fourier domain that is specific to CT-reconstruction. We provethat such approaches become convergent regularization methods as wellas the fact that the reconstructions they provide are typically muchsmoother than the training data they were trained on. Finally, wecompare the spectral as well as the Fourier-based approaches for CT-reconstruction numerically, discuss their advantages and disadvantagesand investigate the effect of discretization errors at different resolutions LB - PUB:(DE-HGF)16 UR - <Go to ISI:>//WOS:001169095400002 DO - DOI:10.1007/s42967-023-00333-2 UR - https://bib-pubdb1.desy.de/record/596988 ER -