System matrix computation vs storage on GPU: A comparative study in cone beam CT

Abstract

Partial system matrix storage was shown to yield the lowest relative performance. On-the-fly ray tracing was shown to be the most flexible method, yielding reasonable execution times. A fully stored system matrix allowed for the lowest backprojection and OSC iteration times and may be of interest for certain performance-oriented applications.

Distributed Iterative CT Reconstruction using Multi-Agent Consensus Equilibrium

“…We specify F more precisely in Algorithm 2. In Theorem II.1, we show that any fixedpoint (w * , X * ) of this partial-update algorithm is a solution to the exact MACE method of (11). Further, in Theorem II.2 we show that for the specific case where f i defined in (3) is strictly quadratic, the partial-update algorithm has guaranteed convergence to a fixed-point.…”

“…as specified by Algorithm 2. Then any fixed-point (w * , X * ) of the Partial-update MACE approach represented by ( 13) is a solution to the exact MACE approach specified by (11).…”

“…One approach to speeding MBIR is to precompute and store the system matrix [9], [10], [11]. In fact, the system matrix can typically be precomputed in applications such as scientific imaging, non-destructive evaluation (NDE), and security scanning where the system geometry does not vary from scan to scan.…”

Technical Note: FreeCT_ICD: An open‐source implementation of a model‐based iterative reconstruction method using coordinate descent optimization for CT imaging investigations

“…(roughly 27 GB, Blob‐based approach). Based on the size of our stored system matrix and the recent analysis provided by Matenine et al., an efficient GPU refinement of our code may be possible in the near future for high end GPU cards. This warrants further investigation and development.…”

“…Taking into account the reconstruction and acquisition parameters, our matrix sizes were larger than those achieved by Guo et al 21 (roughly 1-10 GB, Siddon-based), however, smaller than those achieved by Xu et al 12 (roughly 27 GB, Blobbased approach). Based on the size of our stored system matrix and the recent analysis provided by Matenine et al, 29 an efficient GPU refinement of our code may be possible in the near future for high end GPU cards. This warrants further investigation and development.…”