Study design and population
This was a retrospective study of MR data acquired as part of the UK Biobank imaging enhancement study between August 2014 and August 2015. The UK Biobank study recruited 500,000 volunteers between the ages of 40 and 69 years from across the UK between April 2007 and August 2010, based on proximity to the 21 assessment centers and registration with the National Health Service . A random subset of 100,000 subjects are being invited to participate in the imaging sub-study that images the brain, heart, abdomen, bones, and carotid arteries of these subjects . At the time this retrospective study was performed, scan data for 2921 subjects were available for processing. 6000 subjects had been scanned in the pilot phase, but following review of the pilot data, the MR sequence was further optimized. The 2921 subjects in this study were scanned between January 4, 2016 and September 8, 2016. The study protocol is shown in Fig. 1. Of the 2921 datasets available for processing, 105 (~ 4%) were deemed incomplete or of insufficient quality. As outlined in Fig. 1, the most common reason for failure was generally related to motion artifacts. This left 2816 (~ 96%) datasets that were analyzed to obtain quantities for liver fat fraction (proton density fat fraction—PDFF) and fibroinflammatory disease (cT1). Accompanying participant data [gender, age, and body mass index (BMI)] were acquired through a data access application to UK Biobank (Application 9914). The reference population was defined by excluding those participants with either PDFF ≥ 5%, or BMI ≥ 25 kg/m2. Data were analyzed for the reference cohort, comparison cohorts with PDFF ≥ 5% and/or BMI ≥ 25 kg/m2, as well as separately for gender and age ranges of 40–49, 50–59, 60–69, and ≥ 70 years.
Image acquisition took place at the dedicated Biobank imaging Centre at Cheadle (UK) using a Siemens 1.5 T MAGNETOM Aera, syngo MR D13. Two sequences were used to acquire data: a Shortened Modified Look Locker Inversion (ShMOLLI)  and a multiecho spoiled-gradient-echo .
A single transverse slice, located at the porta hepatis, was chosen to represent the liver for both acquisitions. Acquisition was performed in end-expiration breath-hold and without the aid of any contrast agent injection. This slice-based methodology has previously been shown to correlate well with histology  and predict liver-related outcomes . The rapid acquisition time (≤ 3 min) was an essential requirement to meet the high-throughput demands of the UK Biobank study .
The ShMOLLI sequence is based on an ECG-Triggered method that samples the T1 recovery curve in three epochs over 9 heartbeats using single-shot balanced steady-state free precession (bSSFP) acquisitions with the following parameters: TR = 4.8 ms, TE = 1.93 ms, flip angle = 35°, slice thickness = 8 mm, field of view = 44 × 33 cm2, acquisition matrix = 192 × 144 (zero padded to 384 × 288) yielding an interpolated pixel size of 1.1 mm × 1.1 mm. Parallel imaging acceleration factor of 2 with 24 reference lines was used with the integrated Parallel Acquisition Technique (iPAT). Trigger delay was set to 50 ms and the first ShMOLLI inversion time (TI) was set to 170 ms with five increments of 50 ms each.
The T1 relaxation map acquired using the ShMOLLI sequence is affected by excess iron, which reduces T1. Thus, an algorithm removes the bias introduced by excess iron, which can be calculated from the T2* maps , from the T1 measurements, providing the iron corrected T1 (cT1) . The specific methodology of this algorithm has been described previously , and simulates the T1 that would be measured on a 3T scanner in the absence of excess liver iron, with a T2* of 23.1 ms (at 1.5T) used as the threshold for excess iron.
The multiecho spoiled-gradient-echo chemical shift encoded acquisition was used to calculate T2* and PDFF maps of the liver. The following parameters were used: field of view = 40 × 40 cm2, acquisition matrix = 160 × 160 yielding a voxel size of 2.5 mm × 2.5 mm, slice thickness = 6 mm, flip angle = 20°, TR = 27 ms, and 2 signal averages, requiring 9 s to acquire. Ten echo times were selected such that the signals from fat and water were in phase and out of phase at 1.5T (TE = 2.38, 4.76, 7.14, 9.52, 11.90, 14.28, 16.66, 19.04, 21.42, and 23.80 ms). For PDFF, a three-point DIXON technique [22, 23] was used on the complex data from the second, third, and fourth echoes. This technique takes into account magnetic field inhomogeneity and assumes that the fat has a single-peak frequency. The flip angle of 20 degrees leads to some T1 bias, which has been shown to reduce PDFF values by a factor of 1.2, relative to using a lower flip angle with an in-house implementation of the IDEAL methodology . For R2* estimation, an exponential signal decay model  was fitted to the magnitude data from the second, fourth, sixth, eight, and tenth echo times (i.e., the five in-phase echoes).
Raw MR data were sent to a central reporting laboratory and transferred to an individual workstation (Mac OSX) for analysis. All processed data are available through application to UK Biobank.
Image data were analyzed, blinded to all other subject data, using LiverMultiScan Discover software . For each T2*, cT1, and PDFF image, three circular regions of interest (ROIs) of 15 mm diameter were selected, and a mean value from the pixels within the ROI was calculated. ROIs were manually placed by a trained analyst to capture the distribution of image intensities within the liver and avoid MRI artifacts, if present. The T2* and T1 values together provide the cT1. Characterization of liver fat (PDFF) in the UK Biobank cohort has been previously reported .
Inter- and intra-reader variability
Inter- and intra-reader variability in ROI placement was determined in an initial set of MR datasets collected from the UK Biobank imaging cohort. 39 MR datasets were selected to encompass the range of cT1 and PDFF values found in the population. Six technologists analyzed the anonymized datasets on day one and then again on day four. Intra and inter-reader agreement was calculated using intra-class correlation (two-way mixed model with fixed effects, referred to as Class 3 in the original work)  and Bland–Altman analyses .
Two sets of Bland–Altman analyses were performed. The first set compared values acquired on day one and day four for each of the six technologists. The mean and range of the limits of agreement from these six Bland–Altman are reported to demonstrate the intra-rater reliability. The second set of Bland–Altman analyses was an all-against-all analysis of data from day one. The results of each technologist were compared to those of each other technologist. The mean and range of the limits of agreement from these fifteen Bland–Altman analyses are reported to demonstrate the inter-rater reliability.
Normality of BMI, PDFF, and cT1 distributions was tested using D’Agostino–Pearson and Kruskal–Wallis tests. Summary data for cT1 are presented as medians with interquartile ranges (IQR) and means with standard deviation (SD). Mann–Whitney tests were used to test for statistically significant differences between groups. Correlation analysis was performed using linear regression.
Case studies are provided in the Appendix. Case studies given in Figs. 4, 5, 7, and 8 in Appendix are from the Patient Understanding of LiverMultiScan trial (ClinicalTrials.gov Identifier: NCT02877602) performed at the Oxford University NHS Foundation Trust, UK. The case study in Fig. 6 is from Massachusetts General Hospital, US.