Low latency iterative reconstruction of first pass stress cardiac perfusion with physiological stress using graphical processing unit
Journal of Cardiovascular Magnetic Resonance volume 15, Article number: E10 (2013)
Cardiac MR perfusion has been shown to provide high diagnostic accuracy in detection of the coronary artery disease . We have recently installed an MR-compatible supine bicycle mounted on the scanner table, which allows performing CMR perfusion immediately after physiologic stress. However, patients are unable to sustain a breathold after physical exercise, limiting the choice of acceleration techniques such as k-t approaches. Additionally, due to subject motion during exercise, coil sensitivity map are inaccurate resulting in imaging artifacts in conventional parallel imaging reconstruction. Compressed sensing (CS) is an alternative acceleration technique that enables high acceleration even without exploiting temporal dimension or need for coil maps. However, iterative CS reconstruction of randomly undersampled k-space is lengthy, performed off-line and is not usually integrated into the workflow of a clinical scan which requires viewing and initial assessment on the scanner console and storing the clinical images on the hospital PACS system. In this proposal, we aim to develop an accelerated iterative CS reconstruction workflow for reconstruction of CS acquired perfusion data using physical stress perfusion.
Figure 1 shows the workflow of the accelerated CMR perfusion reconstruction. After completion of the CMR perfusion sequence, the reconstruction process is manually started by CMR technologist using an in-house graphical user interface. All the subsequent reconstruction steps are then performed automatically without any user interaction. The raw data are pre-processed on the scanner workstation and sent to a dedicated computer for reconstruction (equipped with graphic processing unit (GPU) NVIDIA Tesla) and finally sent back to the scanner workstation and the PACS database. Pre- and post-processing are performed using the ReconFrame platform (Gyrotools, Zurich, Switzerland). The GPU-based CS reconstruction is implemented using a fast alternating minimization approach . Since this reconstruction is iterative and voxel-independent for each iteration, the parallelization level of the GPU implementation was set to the voxel level. The presented workflow has been tested during bicycle ergometer stress CMR perfusion exams in healthy subjects using a 1.5T Philips scanner and a prospective 4× CS-accelerated CMR perfusion sequence. Computation time and latency of each reconstruction step was measured and compared to a non-parallelized implementation.
The proposed workflow allows reconstruction and viewing of the CS accelerated perfusion on the scanner console. The GPU-based implementation provides a 12-fold reduction in reconstruction time with an overall latency of 5 min 15s (Table 1).
GPU based CS-reconstruction significantly improved the reconstruction time and guarantee a minimal latency required for optimized clinical MR protocol for CMR perfusion during physical stress.
Greenwood : . The Lancet. 2012
Yang : . IEEE JSTSP. 2010
About this article
Cite this article
Roujol, S., Basha, T.A., Schülke, C. et al. Low latency iterative reconstruction of first pass stress cardiac perfusion with physiological stress using graphical processing unit. J Cardiovasc Magn Reson 15 (Suppl 1), E10 (2013). https://doi.org/10.1186/1532-429X-15-S1-E10