High-speed reconstruction of low-dose CT using iterative techniques for image-guided interventions

Minimally invasive image-guided interventions (IGIs) lead to improved treatment outcomes while significantly reducing patient trauma and recovery time. Ultrasound and fluoroscopy have been traditionally used for image guidance. But these imaging modalities do not provide a comprehensive three-dimensional (3D) view of the anatomy. Because of features such as fast scanning, high spatial resolution, 3D view and ease of operation, computed tomography (CT) is increasingly the choice of intra-procedural imaging technique during IGIs. The risk of radiation exposure, however, limits its current and future use.;We perform ultra low-dose scanning to overcome this limitation. To address the image quality problem with ultra low-dose CT, we reconstruct images using the iterative Paraboloidal Surrogate (PS) algorithm. As iterative techniques are generally computationally intensive, we have accelerated the PS algorithm on a cluster of CPUs and also a GPU. Here, we first compare the quality of the low-dose images reconstructed using the PS algorithm and the standard filtered-back projection (FBP) algorithm. Using actual scanner data, we demonstrate visually acceptable improvement in the quality of reconstructed images using the iterative algorithm.;We further demonstrate a fast implementation of the Ordered Subsets version of the PS algorithm for axial scans on a cluster of 32 processors using the MPI (Message Passing Interface) and an NVIDIA 8800 GTX GPU using CUDA (Compute Unified Device Architecture). Several studies in the recent past have reported computing forward and back projection on GPU using the rasterization framework. However, the GP-GPU (General Purpose GPU) framework used in our implementation is more generic and accommodates a wider variety of penalty functions on the GPU as compared to the rasterization framework. This obviates the need to transfer data between the GPU and CPU during reconstruction.;We have compared the GPU and the cluster implementations using the ray-tracing method to the exact implementation using a pre-computed weight matrix on a single CPU. We demonstrate about 20 times speedup using a cluster of 32 processors and over two orders of improvement in speed using the GPU, while the image quality remains comparable to that of the exact implementation.