Purpose: Compressed sensing‐based iterative cone beam CT (CBCT) reconstruction techniques can reconstruct CBCT from under‐sampled noisy projection data, allowing for imaging dose reduction. The long computation time prevents them from clinical applications. Although GPU dramatically improves computational efficiency, the computation time is still too long. The purpose of this project is to develop a reconstruction algorithm on a multi‐GPU platform. Methods: We have developed tight‐frame(TF) based CBCT reconstruction system on a workstation with 4 NVIDIA GTX590 GPUs. The algorithm iterates two steps: a conjugate gradient least square step (CGLS) enforcing projection condition and a regularization step improving image quality through TF domain. The first step involves frequent forward and backward x‐ray projections, which is accelerated by distributing tasks corresponding to different projection angles among GPUs. A parallel‐reduction algorithm is employed to accumulate data at all GPUs. The regularization step is achieved by having each GPU processing a sub‐volume. Boundary‐layer data between sub‐volumes are kept to maintain correct boundary conditions. A half‐fan reweighting technique is also invented to mitigate ring artifacts caused by imperfect scanning geometry. Results: Under the quad‐GPU system, the CGLS step, the regularization step are accelerated by 3.2∼3.6 times and 1.6∼2.6times compared to single‐GPU version, respectively. The overall speed‐up factor is 3.06∼3.51 times. As for the absolute time, it takes 0.41∼3.90 sec per iteration step depending on the image resolution and number of projections. Considering it usually takes about 10 iteration steps for the algorithm to achieve satisfactory image quality, the total reconstruction time ranges from a few seconds to up to 40 seconds. High quality CBCT images have been obtained in our system. The reweighing strategy also removes the ring artifacts in half‐fan cases. Conclusion: A TF‐based CBCT reconstruction on a multi‐GPU platform has been successfully developed. The achieved efficiency and image quality facilitates clinical implementations. This work is supported in part by NIH (1R01CA154747‐01), Varian Medical Systems through a Master Research Agreement, the Early Career Award from Thrasher Research Fund, and the University of California Lab Fees Research Program.
ASJC Scopus subject areas
- Radiology Nuclear Medicine and imaging