RSNA 2015

Abstract Archives of the RSNA, 2015


SSE22-04

Fast Implementation of the Katsevich Reconstruction Algorithm for Dedicated Breast CT

Monday, Nov. 30 3:30PM - 3:40PM Location: S403B



Zhenzhen Jiang, Erlangen, Germany (Presenter) Nothing to Disclose
Marcel Beister, Erlangen, Germany (Abstract Co-Author) Employee, CT Imaging GmbH
Christian Steiding, MSc, Erlangen, Germany (Abstract Co-Author) Employee, CT Imaging GmbH
Martin Hupfer, PhD, Erlangen, Germany (Abstract Co-Author) Employee, CT Imaging GmbH
Daniel Kolditz, PhD, Erlangen, Germany (Abstract Co-Author) Employee, CT Imaging GmbH
Willi A. Kalender, PhD, Erlangen, Germany (Abstract Co-Author) Consultant, Siemens AG Consultant, Bayer AG Founder, CT Imaging GmbH Scientific Advisor, CT Imaging GmbH CEO, CT Imaging GmbH
PURPOSE

We designed a dedicated breast computed tomography (BCT) system with photon-counting technology with a small detector size in cone direction and use dynamic pitch spiral trajectories starting directly at the patient table in order to cover the whole breast. The Katsevich image reconstruction algorithm is suited for such trajectories, but no fast dynamic pitch implementation is available. The aim of this study was to investigate if the algorithm can be accelerated sufficiently to allow for routine clinical workflow.

METHOD AND MATERIALS

The Katsevich algorithm is an exact filtered backprojection type algorithm suitable for both constant and dynamic pitch spiral cone beam trajectories. The algorithm consists of two major parts: preprocessing of the 2D projection data and 3D backprojection. Both were adapted to support dynamic pitch datasets and to allow for fast and parallelized computation. The algorithm was accelerated by graphics processing units (GPU) using the CUDA framework. The datasets used for the measurements consisted of 6000 projections with 2816x512 pixels. We performed two reconstructions tasks: A fast preview mode with 256 images with 512² pixels and a high-resolution mode with 1024 images with 1536² pixels. Speed and image quality measurements were performed on a high-end system with an NVIDIA Quadro K5200 GPU. Image quality of the Katsevich-reconstructed images was compared to that of a standard Feldkamp-type spiral reconstruction (SFDK) algorithm.

RESULTS

2D preprocessing took 19 s and 253 s for preview and high-resolution mode, respectively. 3D backprojection took 48 s and 420 s which resulted in a total reconstruction time of 93 s and 12 min 15 s for preview and high-resolution mode, respectively. Katsevich reconstructed images were cone-beam artifact-free in contrast to SFDK images. Resolution and image noise were equivalent to the results of SFDK.

CONCLUSION

The proposed GPU implementation improved speed of reconstruction markedly and provided artifact-free high-resolution BCT images. Less than 2 minutes for the preview and less than 13 minutes for the high-resolution reconstruction appear sufficient for routine clinical workflow and can be further reduced by using multiple GPUs.

CLINICAL RELEVANCE/APPLICATION

Fast dynamic pitch spiral reconstruction is available and allows for adequate clinical workflow.