Hepatic steatosis is an important global health problem, affecting 20-30% of adults in the Western Hemisphere and is central to the pathogenesis of type 2 diabetes¹. Techniques such as the 3-point Dixon technique and IDEAL have proven to be sensitive non-invasive protocols for quantifying hepatic steatosis in interventional type 2 diabetes trials². However, both chemical shift imaging and T2-corrected spectroscopic acquisitions typically require breath hold times of 18-29s, which are challenging for most patients, and impossible for some. Scan acceleration by undersampling data and reconstructing with combined compressed sensing and parallel imaging (CS-PI) may substantially reduce the required breath hold duration³. However, to date no study has quantitatively optimized the reconstruction parameters required for CS-PI, or evaluated quantitative fat fraction results from prospectively undersampled data. The aim of our study was to determine optimized reconstruction parameters and determine the limits of agreement of hepatic fat fraction quantified by accelerated MRI compared to conventional acquisitions. This will allow us to determine the maximum acceleration factor advisable and justify applying the technique prospectively in clinical research.Eleven subjects with type 2 diabetes and 1 healthy subject were recruited. The MRI examination was performed on a 3T Philips Achieva using a 6 channel cardiac array, with the subject in a supine position. Fully-sampled and prospectively undersampled k-space data at acceleration factors of 2.6x, 2.9x, 3.8x, 4.8x (Figure 1) were acquired during consecutive expiration breath holds. A custom 3D spoiled gradient echo was used with variable density Poisson disk undersampling in the ky-kz encoding plane with a central fully-sampled region (20x20 encodes for the 2.6x acceleration, 20x14 encodes for higher accelerations). The matrix size was 221x160x16 yielding a voxel resolution of 1.90x1.90x7mm, and 5 unipolar gradient echoes per repetition time: repetition time (msec)/echo times (msec)/flip angle (deg)=7.0ms/0.91,2.16,3.42,4.67,5.92ms/3^o, bandwidth 2144Hz/pixel. The fully sampled acquisition time was 17.7s and the duration of the single breath hold for the accelerated acquisitions was 6.8s, 6.1s, 4.7s and 3.7s respectively. Data using conventional PI (SENSE) at 2.6x acceleration were also collected.The 3D raw data were reconstructed with a compressed-sensing L1-ESPIRiT formulation with the Daubechies-4 wavelet used as the sparsifying transform⁴. For each dataset, the optimal weighting parameter was determined by comparing the reconstruction of each echo with the reconstruction of the fully-sampled data for several different values of the parameter using both the root mean square error (RMSE) and the structural similarity index (SSIM)⁵. Proton density fat fraction maps were calculated by a mixed fitting method⁶ described fully elsewhere⁷, and analysed by multi-slice region of interest analysis in ImageJ⁷. KGH and LWM independently marked ROIs to create an inter-rater analysis for fat fraction and LWM repeated all ROIs after a period of a week without the original ROIs to assess intra-rater variability. Comparisons were performed by Bland-Altman analysis. Image quality was qualitatively assessed⁷.All subjects successfully completed the imaging protocol (Figure 2a-c). 10/11 T2D subjects showed steatosis (>5%). The number of slices with delineated ROIs per subject was 12 ± 2 (range 9-15). 143 ROIs were delineated in total and were available for statistical analysis. The optimum reconstruction parameter was found to be 0.3 for the 2.6x CS-PI (Figure 2d) and 0.5 for all other accelerations: these values were consistent across the patient cohort. The fat fractions for the fully sampled and undersampled reconstructions are given for each individual in Table 1. There was no significant bias of fat fraction for any of the accelerations and the 95% limits of agreement were tight and similar for the 2.6x, 2.9x and 3.8x CS-PI reconstructions (1.2%, 1.2% and 1.1%, Figure 2e and f), while they were higher for the 4.8x CS-PI reconstruction (1.5%). The inter-rater analysis of fat fraction showed negligible bias (0.2%) and a 95% limit of agreement of 0.8%, while the intra-rater analysis had a 95% limit of agreement of 0.7% with negligible bias (0.1%). Image quality was sufficient for ROI analysis for all scans except some 4.8x CS-PI images (Table 2). All conventional PI images were severely artefacted and unsuitable for analysis (Table 2).Our study demonstrates that the measurement of liver fat fraction can be substantially accelerated by prospective undersampling and reconstruction by combined CS-PI, with good fidelity of the fat fractions compared to conventional full sampling. Optimal reconstruction parameters are fixed for a given protocol, avoiding the need for patient by patient calibration. As demand for measurement of liver fat fraction grows, the time efficiency of accurate quantification is likely to become of greater importance. The uncertainty in fat fraction caused by undersampling up to a factor of 3.8x was similar to that of the inter-rater uncertainty in the unmodified technique, with only a small degree of image distortion. The 95% limits of agreement for the 4.8x undersampling were substantially larger than either the repeatability or inter-observer limits of agreement. The image quality score at 4.8x showed marked degradation, with uncertainty over the delineation of a number of regions of interest using this scan alone (not shown⁷). From these data, we would recommend that 3.8x should be regarded as the practical limit of acceleration for this protocol and reconstruction method.