4D-CT reconstruction with unified spatial-temporal patch-based regularization

Kazantsev, Daniil and Thompson, William M. and Lionheart, William R.B. and Van Eyndhoven, Geert and Kaestner, Anders P. and Dobson, Katherine J. and Withers, Philip J. and Lee, Peter D. (2015) 4D-CT reconstruction with unified spatial-temporal patch-based regularization. Inverse Problems and Imaging, 9 (2). pp. 447-467. ISSN 1930-8345 (https://doi.org/10.3934/ipi.2015.9.447)

[thumbnail of Kazantsev-etal-IPI-2015-4D-CT-reconstruction-with-unified-spatial-temporal-patch-based-regularization]
Preview
Text. Filename: Kazantsev_etal_IPI_2015_4D_CT_reconstruction_with_unified_spatial_temporal_patch_based_regularization.pdf
Accepted Author Manuscript

Download (10MB)| Preview

Abstract

In this paper, we consider a limited data reconstruction problem for temporarily evolving computed tomography (CT), where some regions are static during the whole scan and some are dynamic (intensely or slowly changing). When motion occurs during a tomographic experiment one would like to minimize the number of projections used and reconstruct the image it-eratively. To ensure stability of the iterative method spatial and temporal constraints are highly desirable. Here, we present a novel spatial-temporal regularization approach where all time frames are reconstructed collectively as a unified function of space and time. Our method has two main differences from the state-of-the-art spatial-temporal regularization methods. Firstly, all available temporal information is used to improve the spatial resolution of each time frame. Secondly, our method does not treat spatial and temporal penalty terms separately but rather unifies them in one regularization term. Addition- ally we optimize the temporal smoothing part of the method by considering the non-local patches which are most likely to belong to one intensity class. This modification significantly improves the signal-to-noise ratio of the reconstructed images and reduces computational time. The proposed approach is used in combination with golden ratio sampling of the projection data which allows one to find a better trade-off between temporal and spatial resolution scenarios.