Effect of noise and estimator type on bias for analysis of liver proton density fat fraction

Abstract

PurposeProton-density fat-fraction (PDFF) is typically measured from PDFF maps by calculating the mean PDFF value within a region of interest (ROI). However, the mean estimator has been shown to result in bias when signal-to-noise ratio (SNR) is low, resulting from a skewed distribution of PDFF noise statistics. Thus, the purpose of this work was to determine the relative performance of three estimation methods (mean, median, maximum likelihood estimators (MLE)) for analysis of liver PDFF maps. MethodsObservational study of adult patients (n = 56) undergoing abdominal MRI. Both 2D-sequential CSE-MRI (‘low-SNR’) and 3D CSE-MRI (‘high-SNR’) acquisitions were obtained. Single-voxel MRS formed the independent reference measurement of hepatic PDFF. Intra-class correlation was tested on a subset of ‘low-SNR’ acquisitions. ROIs were semi-automatically co-registered across all acquisitions. Bland-Altman analysis and intra-class correlation coefficients were used for statistical analysis. A p-value of <0.05 was considered significant. ResultsFor in vivo low-SNR acquisitions, the mean estimator had a larger error than either the median or MLE values (bias ~ −1% absolute PDFF). The intra-class correlation coefficient was significantly greater for median and maximum likelihood estimators (0.992 and 0.993, respectively) compared to the mean estimator (0.973). ConclusionAlternative ROI analysis strategies, such as MLE or median estimators, are useful to avoid SNR-related PDFF bias. Median may be the most clinically practical strategy given its ease of calculation.