by Robert Frysch, Tim Pfeiffer, Georg Rose
Abstract:
Many applications in X-ray imaging, especially iterative reconstruction for computed tomography (CT), require fast and accurate forward projections of 3D volumes. In this study we present an implementation of a highly accurate voxel-based footprint method for cone beam CT (CBCT). The vast computational effort is compensated by means of efficient approximations as well as massively parallel GPU programming. As a trade-off between accuracy and computation speed we use the separable trapezoid approach by Y. Long et al. (IEEE Trans Med Imaging, 2010). Our implementation focuses on using shared memory to quickly compute disjoint detector parts in parallel. Simultaneous read-write accesses within such a sector are handled using atomic functions which are part of the OpenCL 1.1 programming language used. The proposed routine transfers a large amount of the computational load to all available GPU hardware while direct communication between single devices is not required. As a consequence the computational time scales well with additional devices. The proper functionality is evaluated within an iterative statistical reconstruction algorithm. A typical volume for CBCT of 25x25x25 cm³ (512³ voxels) can be projected in less than 300 ms per view with a resolution of 616x480 pixels. Those results demonstrate that the proposed method is a valid alternative to the common ray-based projectors. With respect to this properties our approach is well suited for the demanding CBCT applications (e.g. perfusion imaging), which benefit from iterative reconstructions due to its sparsity and poor quality of the measured flat-panel data.
Reference:
Scalable OpenCL Accelerated Multi-GPU Projection for Cone Beam Computed Tomography with a Highly Accurate Separable Footprint Method (Robert Frysch, Tim Pfeiffer, Georg Rose), In WFITN, 2013.
Bibtex Entry:
@inproceedings{frysch_scalable_2013,
address = {Buenos Aires},
title = {Scalable {OpenCL} {Accelerated} {Multi}-{GPU} {Projection} for {Cone} {Beam} {Computed} {Tomography} with a {Highly} {Accurate} {Separable} {Footprint} {Method}},
abstract = {Many applications in X-ray imaging, especially iterative reconstruction for computed tomography (CT), require fast and accurate forward projections of 3D volumes. In this study we present an implementation of a highly accurate voxel-based footprint method for cone beam CT (CBCT). The vast computational effort is compensated by means of efficient approximations as well as massively parallel GPU programming. As a trade-off between accuracy and computation speed we use the separable trapezoid approach by Y. Long et al. (IEEE Trans Med Imaging, 2010). Our implementation focuses on using shared memory to quickly compute disjoint detector parts in parallel. Simultaneous read-write accesses within such a sector are handled using atomic functions which are part of the OpenCL 1.1 programming language used. The proposed routine transfers a large amount of the computational load to all available GPU hardware while direct communication between single devices is not required. As a consequence the computational time scales well with additional devices. The proper functionality is evaluated within an iterative statistical reconstruction algorithm. A typical volume for CBCT of 25x25x25 cm³ (512³ voxels) can be projected in less than 300 ms per view with a resolution of 616x480 pixels. Those results demonstrate that the proposed method is a valid alternative to the common ray-based projectors. With respect to this properties our approach is well suited for the demanding CBCT applications (e.g. perfusion imaging), which benefit from iterative reconstructions due to its sparsity and poor quality of the measured flat-panel data.},
booktitle = {{WFITN}},
author = {Frysch, Robert and Pfeiffer, Tim and Rose, Georg},
year = {2013}
}