There are currently multiple free-breathing quantitative 3D liver perfusion imaging techniques under development including GRASP¹ and non-Cartesian through-time spiral GRAPPA². While these kinds of completely free-breathing technologies show great potential in quantitative characterization of focal liver lesions³, application of these high spatiotemporal acquisition methods in routine clinical scans is greatly limited by the prolonged image processing times, which are often orders of magnitude longer than the clinical scan session. While the image reconstruction times are significantly shorter for the through-time spiral GRAPPA technique compared to compressed sensing technologies, due to the large number of volumes acquired in such a free-breathing acquisition, the image reconstruction for these many volumes in MATLAB also requires approximately 5-6 hours. Thus computing is an important practical limitation to extending the technology for many practical clinical settings. To make this technique feasible for more routine clinical use, the objectives of this study were to 1) to accelerate image reconstruction using C++ and 2) to accelerate image registration using a new parallel image registration algorithm that is optimized to run on GPUs.The proposed image analysis methods were performed on five image sets, two from asymptomatic subjects and three from patients with metastatic adenocarcinoma. For each subject, time series of 3D DCE-MRI images were obtained from a Siemens 3T Skyra scanner with 30 coil elements. A total of 100 to 120 volumes were acquired continuously in approximately 4 min for each subject while the subjects were breathing freely. Each volume includes 60 slices with an in-plane matrix size of 208×208 to 224×224. To accelerate image reconstruction, the whole process was implemented using C++ on a hybridized system with a multi-core CPU (an Intel Xeon X5650 CPUs, 48GB of RAM) and a graphical processing unit (GPU; NVIDIA Fermi M2090). The results were compared to those obtained using MATLAB on a standalone desktop computer (Intel Xeon E5-2630 v2 CPUs, 128GB of RAM). To eliminate respiratory motion between different volumes for perfusion quantification, non-rigid image registration was performed with BROCCOLI, a parallel-processing image registration algorithm that heavily employs GPU hardware⁴. Both the speed and quality of registration were compared to volumes registered with FMRIB's Software Library (FSL). Registration using both algorithms was run on the same hardware configuration: CPU: Intel i5 4670K, 8GB of RAM; GPU: Nvidia GTX Titan X. The quality of the BROCCOLI registered images to FSL registered images was quantified using the Dice similarity coefficient (DSC), a commonly used measure of image similarity⁵. A dual-input single-compartment model was then applied to retrieve liver perfusion parameters and the results were compared to those obtained with the original processing methods.

Fig. 1 shows representative post-contrast through-time spiral GRAPPA images reconstructed using both MATLAB and C++ from a subject with metastatic breast cancer. While no clear visual difference was observed in the images obtained from the two methods, the image reconstruction time was reduced by 32 times with the C++ approach (Table 1) to approximately 10-11 minutes for the 100-120 volumes that are acquired over 3-4 minutes for a free-breathing patient exam.

Fig. 2 shows the effectiveness of using BROCCOLI to eliminate respiratory motion from the reformatted coronal view. The mean DSCs for the five BROCCOLI-registered datasets as compared to the unregistered dataset and FSL-registered dataset were 0.7166±0.0407 and 0.9994±0.0005, respectively, which suggests similar registration performance between FSL and BROCCOLI methods. However, the average registration time was reduced from 1012.50±116.78 min to 1.11±0.09 min with BROCCOLI, which represents a 900 fold speedup in processing time (Table 1). The practical implications are that while registering the volumes for the five subjects took an average of 16.9 hours previously, the non-rigid registration of the same volumes, with no loss in fidelity of registration, now requires 1.1 minutes. Thus the time from acquiring the free-breathing images, to availability of registered, time resolved data is approximately 12 minutes. These data can then be viewed clinically or further processed for quantitative analysis.

Fig. 3 and Table 2 present perfusion quantification obtained with the proposed method as well as the original method and a close agreement in all perfusion parameters was observed between the two methods (P>0.9).

In this study, a rapid image reconstruction and registration method was developed for free-breathing 3D liver perfusion imaging using through-time spiral GRAPPA acceleration. A marked improvement in the time to availability of reconstructed and registered volumes (~12 minutes compared to >22.5 hours) was observed with extremely minor differences in both image quality and perfusion quantification. This moves the technology from a research novelty to clinical feasibility.