Dynamic contrast enhanced MRI involves tradeoffs between spatial and temporal resolution. High temporal resolution (TRES) is needed to capture the arterial phase or accurately measure the arterial input function (AIF) while high spatial resolution is needed for diagnostic quality images. The golden angle radial stack-of-stars trajectory is a flexible, motion robust acquisition scheme that has been used for liver imaging [1]. We used a multiresolution compressed sensing (CS) reconstruction scheme: AIFs were generated using a very small temporal window while dynamic data were reconstructed with larger temporal window for high spatial resolution. We validated the accuracy of the reconstruction method on a realistic phantom and then applied it to in vivo data.

Phantom validation: Validation was performed on a synthetic phantom proposed by Yoruk et al. [2], where the true AIF, k_ep (1.5) and GFR (74 ml/min) are known a priori. The phantom was synthesized from a dynamic contrast enhanced MRI dataset acquired on a pediatric patient with 4s TRES. An aortic enhancement curve with 1s temporal resolution were then fused in and a new data set created using interpolation. The k-space data was then sampled using a golden angle stack-of-stars radial trajectory.

Reconstruction schemes: Dynamic MRI data were reconstructed at different temporal resolutions ranging from 1s to 12s using a sliding window. A high temporal resolution (HTR) AIF was generated using a 1s temporal window and Non Uniform Fourier Transform (NUFFT), followed by a 5-point moving average filter to reduce noise (Figure 1). For each TRES, three different reconstruction methods were performed: 1) NUFFT reconstruction with AIF estimated from the reconstructed data 2) NUFFT with the 1s temporal resolution HTR-AIF 3) Compressed Sensing reconstruction [1] with Total Variation sparsity constraint applied across the temporal dimension $$x = \underset{x} {\mathrm{argmin}}\left|\left|F.C.x-k\right|\right|_2^2+\lambda\left|\left|TV(x)\right|\right|_1$$ and using the HTR-AIF. After creating cortical and aortic ROIs, GFR and k_ep estimates were computed for the cortical ROIs using a 3-compartment model [3] for each of the three schemes.

The three schemes were also used on in vivo datasets for GFR estimation. Imaging was performed on a 3T MRI scanner (Skyra, Siemens Healthcare, Malvern, PA) using a radial golden angle stack-of-stars spoiled gradient echo pulse sequence on abdominal imaging patients, after informed consent. Acquisition parameters: TR = 3.52 ms; TE = 1.5 ms; FOV = 38 cm; flip angle = 10; receiver bandwidth=1565 Hz/pixel; acquisition matrix=288x288x44; 1.3x1.3 in-plane spatial resolution and 3 mm thick slices to achieve whole abdomen coverage. Free breathing data were acquired for 90s following injection of Gadolinium contrast.

Figure 2 shows NUFFT reconstructed images of the synthetic phantom for TRES of 4s (a) and 12s (b). CS reconstructed images for TRES of 1s (c) and 4s (d) are also shown. The superior image quality of the CS reconstruction and lack of streaking artifacts is apparent. CS can increase the TRES by at least 3X without compromising spatial resolution or image quality.

Table 1 reports errors in GFR and k_ep estimates for all three reconstruction schemes for the phantom with TRES of 1s, 4s and 12s. There is a tradeoff between spatial and temporal resolution and the GFR/ kep estimates are most accurate for midrange TRES (4s). It can be seen that the GFR estimates are poor for the NUFFT case for all TRES (top row). While HTR-AIF with 1s TRES improves the GFR estimates for the NUFFT (middle row), the image quality is still poor (Fig 2a-b). The CS reconstruction with HTR-AIF (bottom row) achieves both accurate GFR and k_ep estimates as well as diagnostic image quality with 4s TRES (Fig 2d).

Figure 3 shows pre-contrast (a), arterial/renal cortical (b) and renal medullary (c) phases from a 3D radial stack-of-star dataset acquired on a patient and reconstructed using 4s TRES CS reconstruction.The images are high quality with negligible artifacts compared to the 12s TRES NUFFT reconstruction (d), highlighting the usefulness of temporally constrained CS. Table 2 shows GFR estimates from an in vivo dataset, which follows the same trend as the phantom. The CS estimates are in line with expected GFR values for a normal kidney while the CS reconstruction removes streaking artifacts still present in the NUFFT reconstruction (Fig 3d).

We have demonstrated the feasibility of multiresolution imaging using a radial stack-of-stars scheme to accurately estimate the AIF as well as produce diagnostic quality dynamic images in vivo. The high temporal resolution AIF estimate can significantly reduce errors in GFR estimation for free breathing dynamic MR urography. The same technique can also be used for pharmacokinetic modeling of breast, liver and prostate cancers.