Argonne National Laboratory, Data Science and Learning Division, United States of America
As the sophistication and speed of today's X-ray experiments grow, collecting the most informative data has become ever more relevant, necessitating the development of algorithms that can provide good quality reconstructions from tomographic data streams. Almost all conventional reconstruction systems and algorithms work exclusively offline, however, requiring that complete datasets be collected and available before they can be processed. Further, these systems and algorithms provide limited consideration and support for challenging experiments, such as imaging samples with dynamic features, where both spatial and temporal properties of the features rapidly change.
We describe here a high-performance runtime system for analyzing and reconstructing streaming tomography datasets using sliding subsets of projection images, and evaluate the reconstruction quality of dynamic features with respect to different runtime configuration parameters using phantom and real-world tomography datasets. Our system enables runtime system parameters to be adjusted dynamically over the course of experiment, providing opportunities for balancing the quality and computational demands of tasks, better observation of phenomena and improving advanced experimental techniques such as autonomous experimental steering.