Generalized SART Methods for Tomographic Imaging
This work provides a computationally efficient and versatile reconstruction framework for tomographic imaging, addressing the need for flexible methods across diverse experimental settings.
The authors show that a general class of tomographic Kaczmarz iterations can be efficiently evaluated via updates similar to SART, enabling regularized reconstructions with non-trivial image-formation models and non-quadratic data-fidelity terms at low computational cost. The method is applicable to various tomographic settings, including polychromatic CT and X-ray phase contrast tomography.
Nowadays, the field computed tomography (CT) encompasses a large variety of settings, ranging from nanoscale to meter-sized objects imaged by different kinds of radiation in various acquisition modes. This experimental diversity challenges the flexibility of tomographic reconstruction methods. Kaczmarz-type methods, which exploit the natural block-structure of tomographic inverse problems, are a promising candidate to provide the required versatility in a computationally efficient manner. In the present work, it is shown that indeed a surprisingly general class of tomographic Kaczmarz-iterations may be efficiently evaluated via computational schemes of a similar structure as updates of the so-called simultaneous algebraic reconstruction technique (SART). This enables regularized reconstructions with non-trivial image-formation models as well as non-quadratic or even non-convex data-fidelity terms at low computational costs. Moreover, the proposed generalized SART schemes are equally applicable in parallel- and cone-beam settings and regardless of the choice of tomographic incident directions. Their potential is illustrated by outlining applications in several non-standard tomographic settings, including polychromatic CT and X-ray phase contrast tomography.