As a quantitative functional imaging technique, dynamic contrast enhanced (DCE) MRI plays an important role in the detection of liver metastases, characterization of tumors, assessing tumor response and studying diffuse liver disease. According to reference (1), an ideal study of liver perfusion requires a high spatial-temporal resolution. Existing liver DCE reconstruction schemes, such as conjugate-gradient SENSE (CG-SENSE) (2), iGRASP (3) and L+S (4), realize their functions through an iterative minimization of regularized or non-regularized least square objective function. Though these solutions are generally acceptable, yet suffer from significantly high computational cost. The aim of this work is to propose a computationally efficient reconstruction scheme, which results in comparable image quality relative to the above iterative schemes. The testing results demonstrate the effectiveness of the prosed method.

This study is based on the GRAPPA-like parallel imaging (PI) scheme, which offers an enhanced performance when the to-be reconstructed image is sparse in the image domain, namely artificial sparsity. Artificial sparsity is capable of effectively improving the results of PI (5, 6). In this work, GRAPPA-based PI is exploited with the aid of dynamic artificial sparsity. The proposed k-t Artificial Sparsity enhanced GROWL (k-t ARTS-GROWL) consists of three steps.

Step 1: Apply sliding window and PI to the time series k-space data for a temporally averaged intermediate result with high signal-to-noise ratio;

Step 2: Use a k-t sparse (7) de-nosing method to remove certain noise-like streaking artefacts and noise, and then extract the artificial sparsity by a subtraction operator;

Step 3: Restore the final reconstruction from PI artificial sparsity result.

To explain the proposed scheme, DCE-MRI with free breathing golden angle radial acquisition was taken as an example. GRAPPA Operator for Wider Lines (GROWL) (8) is utilized as the GRAPPA-like PI scheme. Then denoising along temporal dimension is applied to reduce the noise/artifact level.

The liver DCE-MRI data used in this work is the same as the original iGRASP reference for the convenience of comparison (http://cai2r.net/resources/software). As depicted, the entire dataset is divided into 28 image time frames, 21 radial spokes per frame, the temporal resolution is ~3s/volume.

For iterative schemes, the same parameters and implementations were used as suggested in their original references (2-4). All algorithms are implemented in Matlab (R2012b) running on a HP Elite Desk 800 with quad Intel Core i5-4570 CPU 3.20GHz and 4GB of Memory.

For temporal resolution evaluation, the signal intensity to time curve of aorta (AO) region and portal vein (PV) region were plotted to assess the reconstruction results, as shown in Fig.1 (e), (f). AO and PV are the main contrast medium sources of liver perfusion, which may greatly influence the perfusion effect. AO is labelled by the red arrow in Fig.1 (d); PV is labelled by the blue arrow. CG-SENSE result is used as golden-standard instead of NUFFT reconstruction in reference (3), since the latter contains a variety of streaking artifacts at high reduction factors.

Fig.1 (a) - (d) depict the result of reconstructed images by various methods. Fig.1 (e) and (f) display signal-intensity time courses of two ROIs of each method. As shown in Fig.2, the proposed k-t ARTS-GROWL achieves similar image quality to iGRASP for the same randomly picked frame. The reconstruction time of GROWL, iGRASP, L+S, and k-t ARTS-GROWL methods are 140s, 1309s, 425s and 305s, respectively. Compared to L+S and iGRASP, our approach provides comparable image quality, but much faster.

From the reconstruction results shown in Fig.1, it can be seen that our approach preserves small vessel as the purple arrows pointed, and other image details comparable to iterative schemes. Fig.2 also shows similar results. In addition, our approach provides reliable time resolution as in the signal-intensity time courses in Fig.1 (e) and (f) which is similar to iGRASP and L+S. So the spatial and temporal resolutions offered by our approach are both comparable to iterative schemes. With a comparison between the result of initial GROWL reconstruction and our approach, it can be seen that the dynamic artificial sparsity does improve the performance of PI. Most importantly, our method is computationally efficient. For example, under the circumstance in this study, GROWL operator costs approximately 5s per frame.

Compared with iterative schemes, the k-t ARTS-GROWL reconstruction can result in comparable image quality and temporal resolution with greatly reduced computational time. The proposed method improves the clinical applicability of the high spatiotemporal resolution DCE-MRI.