Estimation of hepatic proton-density fat fraction by using MR imaging at 3.0 T

Abstract

Purpose: To compare the accuracy of several magnetic resonance (MR) imaging-based methods for hepatic proton-density fat fraction (FF) estimation at 3.0 T, with spectroscopy as the reference technique.

Materials and methods: This prospective study was institutional review board approved and HIPAA compliant. Informed consent was obtained. One hundred sixty-three subjects (39 with known hepatic steatosis, 110 with steatosis risk factors, 14 without risk factors) underwent proton MR spectroscopy and non-T1-weighted gradient-echo MR imaging of the liver. At spectroscopy, the reference FF was determined from frequency-selective measurements of fat and water proton densities. At imaging, FF was calculated by using two-, three-, or six-echo methods, with single-frequency and multifrequency fat signal modeling. The three- and six-echo methods corrected for T2*; the two-echo methods did not. For each imaging method, the fat estimation accuracy was assessed by using linear regression between the imaging FF and spectroscopic FF. Binary classification accuracy of imaging was assessed at four reference spectroscopic thresholds (0.04, 0.06, 0.08, and 0.10 FF).

Results: Regression intercept of two-, three-, and six-echo methods were -0.0211, 0.0087, and -0.0062 (P <.001 for all three) without multifrequency modeling and -0.0237 (P <.001), 0.0022, and -0.0007 with multifrequency modeling, respectively. Regression slope of two-, three-, and six-echo methods were 0.8522, 0.8528, and 0.7544 (P <.001 for all three) without multifrequency modeling and 0.9994, 0.9775, and 0.9821 with multifrequency modeling, respectively. Significant deviation of intercept and slope from 0 and 1, respectively, indicated systematic error. Classification accuracy was 82.2%-90.1%, 93.9%-96.3%, and 83.4%-89.6% for two-, three-, and six-echo methods without multifrequency modeling and 88.3%-92.0%, 95.1%-96.3%, and 94.5%-96.3% with multifrequency modeling, respectively, depending on the FF threshold. T2*-corrected (three- and six-echo) multifrequency imaging methods had the overall highest FF estimation and classification accuracy. Among methods without multifrequency modeling, the T2-corrected three-echo method had the highest accuracy.

Conclusion: Non-T1-weighted MR imaging with T2 correction and multifrequency modeling helps accurately estimate hepatic proton-density FF at 3.0 T.

Figure 1:

Typical MR spectrum of fatty liver. Liver MR spectra in 38-year old woman, acquired at echo time (TE) 10, 15, 20, 25, and 30 msec. Fat peaks (1–5) are centered at 0.9, 1.3, 2.1, 4.2, and 5.3 ppm, respectively. Water peak (w) is at 4.7 ppm. Calculated fat fraction (FF) based on total fat and water proton densities was 0.409. The tiny 2.75-ppm peak (*) was not included in the calculation because it was not consistently observed at 3.0 T.

Figure 2:

Fat quantification accuracy of imaging FF calculation methods. Linear regression analysis of imaging FF against spectroscopic FF. Top: Two-, three-, and six-echo method with single-frequency fat modeling (S-2, S-3, and S-6, respectively). Bottom: Two-, three-, and six-echo method with multifrequency fat modeling (M-2, M-3, and M-6, respectively). Red line = best-fit line. Gray line with intercept 0 and slope 1 = perfect agreement between imaging and spectroscopy. * = statistically significant deviation from the null hypotheses (slope = 1, intercept = 0) at Hochberg-adjusted 95% confidence level for multiple comparisons. corr = correction. r = Pearson correlation coefficient.

Figure 3:

Fat quantification accuracy of six-echo multifrequency imaging according to hepatic segments. Linear regression analysis of imaging FF against spectroscopic FF according to hepatic segments (II/III, IV, V, VI, VII, and VIII). Red line = best-fit line. Gray line with intercept 0 and slope 1 = perfect agreement between imaging and spectroscopy. No slope or intercept estimates were significantly different from 1 or 0, respectively, at Hochberg-adjusted 95% confidence level for multiple comparisons. r = regression coefficient.

Figure 4:

Cross-section gray-scale FF (in percentages) maps (dynamic range, 0%–50%) through the upper liver of patients suspected of having fatty liver disease were generated by using an automated online postprocessing algorithm (six-echo multifrequency method). Maps are in (left) segments IVa, VII, and VIII in 26-year-old woman, (middle) segments II, IVa, and VII in 45-year-old woman, and (right) segments IVa, VII, and VIII in 27-year-old man. Spectroscopic (white) and imaging (black) FF values at coregistered locations are shown.