Regridding reconstruction algorithm for real-time tomographic imaging

TLDR

Sub‑second temporal‑resolution tomographic microscopy is becoming feasible at third‑generation synchrotron sources, but data rates near 10 GB s⁻¹ make efficient handling and post‑processing difficult, limiting the exploitation of ultrafast acquisition. This study introduces gridrec, a fast reconstruction algorithm optimized for conventional CPU technology. Gridrec resamples Fourier space from polar to Cartesian coordinates via a regridding procedure, achieving excellent performance with negligible accuracy loss. Gridrec delivers high‑quality reconstructions with up to a 20‑fold speed increase, outperforms filtered back‑projection when the tomographic problem is well sampled, and enables real‑time monitoring of sub‑second acquisition.

Abstract

Sub-second temporal-resolution tomographic microscopy is becoming a reality at third-generation synchrotron sources. Efficient data handling and post-processing is, however, difficult when the data rates are close to 10 GB s−1. This bottleneck still hinders exploitation of the full potential inherent in the ultrafast acquisition speed. In this paper the fast reconstruction algorithm gridrec, highly optimized for conventional CPU technology, is presented. It is shown that gridrec is a valuable alternative to standard filtered back-projection routines, despite being based on the Fourier transform method. In fact, the regridding procedure used for resampling the Fourier space from polar to Cartesian coordinates couples excellent performance with negligible accuracy degradation. The stronger dependence of the observed signal-to-noise ratio for gridrec reconstructions on the number of angular views makes the presented algorithm even superior to filtered back-projection when the tomographic problem is well sampled. Gridrec not only guarantees high-quality results but it provides up to 20-fold performance increase, making real-time monitoring of the sub-second acquisition process a reality.