The relationship between liver triglyceride composition and proton density fat fraction as assessed by ¹H MRS

Abstract

Purpose:

To estimate parameters determining liver triglyceride composition (TC) using ¹H MRS and to assess how TC estimability is affected by proton density fat fraction (PDFF) in adults with nonalcoholic fatty liver disease (NAFLD).

Material and Methods

In this prospective single-site study, 199 adults with known or suspected NAFLD in whom other causes of liver disease were excluded underwent two ¹H MRS STimulated Echo Acquisition Method (STEAM) sequences at 3T. A respiratory-gated water-suppressed free breathing sequence (TE 10 ms, 16 signal averages) was used to assess triglyceride composition in terms of number of double bonds (ndb) and methylene-interrupted double bonds (nmidb), and a single breath-hold long-TR, multi-TE sequence (TR 3500 ms) which acquired 5 single average spectra over TE 10–30 ms was used to estimate liver PDFF. Ndb and nmidb estimability was qualitatively assessed for each case and summarized descriptively. Consistency of ndb and nmidb estimation was examined using ROC analysis. Relationship between ndb and nmidb values and PDFF was presented graphically. Quality-of-fit of ndb and nmidb vs. PDFF was evaluated by Pearson-r correlation. A significance level of 0.05 was used.

Results:

In 263 ¹H MRS exams performed on 199 adult participants, ndb and nmidb were successfully estimated in 7/53 (13.2%) exams with PDFF < 4%, 13/30 (43.3%) exams with PDFF between 4% and 7%, 33/41 (80.5%) exams with PDFF between 7% and 10%, and 124/139 (89.2%) exams with PDFF > 10% (max PDFF 38.1%). Liver TC could be estimated consistently for PDFF > 6.7 %. Both ndb and nmidb decreased with increasing PDFF (ndb = 2.83 - 0.0160·PDFF, r = −0.449 p < 0.0001); nmidb = 0.75 −0.0088·PDFF, r = −0.350, p < 0.0001).

Conclusion:

In a cohort of adults with known or suspected NAFLD, liver triglyceride composition becomes more saturated as PDFF increases.

Graphical Abstract

Adults with known or suspected NAFLD underwent two liver ¹H MRS STEAM sequences at 3T to investigate how triglyceride composition varies with proton density fat fraction (PDFF). A respiratory-gated water-suppressed sequence was used to characterize liver triglyceride composition in terms of number of −CH=CH− and −CH=CH−CH₂−CH=CH− bonds and a long-TR, multi-TE sequence was used to estimate liver PDFF. ¹H MRS indicates that liver triglyceride composition becomes more saturated with increasing PDFF in adults with NAFLD.

INTRODUCTION

Magnetic resonance (MR)-based methods can noninvasively, accurately and precisely estimate liver proton density fat fraction (PDFF), a quantitative imaging biomarker of hepatic steatosis.^1–3 Both MR imaging (MRI)- and MR spectroscopy (MRS)-based estimates of PDFF correlate strongly with histologic hepatic steatosis grade,^4–7 and MRI- and MRS-PDFF agree well with each other.^3,8,9 Although liver fat can be accurately quantified using MR-derived PDFF, the lack of a direct correlation between the amount of liver fat and the development of nonalcoholic steatohepatitis (NASH), the progressive form of nonalcoholic fatty liver disease (NAFLD), may to some extent limit its clinical applications.^10,11 The noninvasive characterization of the mixture of fatty acids in liver triglycerides (i.e., the triglyceride composition [TC]) may potentially help us understand why the majority of patients have simple benign steatosis while a minority develop progressed forms of disease. However, the ability of MR techniques to characterize the composition of fatty acid chains in triglycerides is not well established. Ex vivo analysis of biopsy samples have suggested TC may independently affect clinical outcomes such as risk or severity of NASH or Type-2 diabetes.^12–14

Standard MRI-PDFF techniques are not able to assess TC because they assume a predetermined triglyceride spectrum,^15,16 but specialized MRI and MRS techniques have been developed to allow TC to be characterized in terms of three variables: number of -CH=CH- double bonds per molecule (ndb), number of double bonds separated by a single -CH₂- (nmidb; number of methylene-interrupted double bonds), and fatty-acid chain length (CL) (Table 1).^17–20 However, those studies have focused on fat depots outside the liver (yellow bone marrow, adipose tissue),^18–24 have been acquired in small subject groups,^25,26 or were acquired with sequences designed for estimating MRS-PDFF which cannot accurately assess TC except at the highest PDFF values.¹⁷

Table 1.

Composition of the peaks in the liver ¹H MR spectrum with mean T2 from Hamilton et al¹⁷

Peak	Location	Assignment	Relative Magnitude	In vivo ppm	T2 (ms)
Water (& 1+2)	5.29 ppm	-CH=CH-	2·ndb	4.7 ppm	23
	5.19 ppm	-CH-O-CO-	1
	4.7 ppm	H2O	(Water Signal)
	4.20 ppm	-CH2-O-CO-	4
3	2.75 ppm	-CH=CH-CH2-CH=CH-	2·nmidb	2.75 ppm	51
4	2.24 ppm	-CO-CH2-CH2-	6	2.1 ppm	52
	2.02 ppm	-CH2-CH=CH-CH2-	4·(ndb-nmidb)
5	1.6 ppm	-CO-CH2-CH2-	6	1.3 ppm	62
	1.3 ppm	-(CH2)n-	6·(CL-4) - 8·ndb + 2·nmidb
6	0.90 ppm	-(CH2)n-CH3	9	0.9 ppm	83

In one study based on MRS, subjects with NAFLD had higher saturation and lower unsaturation indices than that seen in healthy controls.²⁷ In another study, TC (specifically ndb) was indirectly assessed in a large adult population in which liver PDFF bias estimated by confounder-corrected chemical-shift-encoded MRI was studied.²⁸ To explain the observed changes in bias with PDFF, that study hypothesized that TC was not fixed, with ndb decreasing as PDFF increased,²⁸ but that hypothesis was not directly tested. Thus, there is a need to determine whether the main parameters characterizing TC (i.e., ndb and nmidb) are correlated with PDFF. Correlation of liver TC with PDFF may act as a confounder when trying to compare liver TC in NASH and non-NASH subjects and if liver TC correlates very strongly with PDFF, this may suggest that liver TC, like PDFF, may not be a good predictor of NASH.

To address these issues, in this study we use a ¹H MRS water-suppressed, respiratory-gated sequence in a large subject cohort of adults with known or suspected NAFLD to estimate TC, to assess how TC estimability is affected by PDFF and to determine whether ndb and nmidb are correlated with PDFF.

MATERIALS AND METHODS

This was a prospective, single-site study approved by our Institutional Review Board and compliant with the Health Insurance Portability and Accountability Act. Adult human subjects were recruited from clinical NAFLD studies being conducted at our institution. Prospectively recruited subjects either had biopsy-proven NAFLD or were at risk for NAFLD due to family history or obesity. Patients with other causes of liver disease were excluded by the hepatology investigator (RL). Written informed consent was obtained from all subjects. Subjects were scanned between January 2011 and December 2013 and were included in this analysis if both liver MRS-PDFF, and MR spectra to assess TC were acquired.

MRS Acquisition

Subjects were scanned at 3T (GE Signa EXCITE HDxt, GE Healthcare, Waukesha, WI) with an 8-channel torso array coil. Multi-planar localization images were acquired, and a 20 × 20 × 20 mm voxel was selected within the right lobe of the liver, avoiding the liver edges as well as large biliary or vascular structures. To minimize J coupling effects, ¹H MR spectra were acquired using the STimulated Echo Acquisition Method (STEAM) sequence which allows a shorter minimum TE than the Point REsolved SpectroScopy (PRESS) sequence, and the mixing time (TM) for the STEAM sequence was fixed at a minimum value of 5 ms.²⁹

Two different STEAM acquisitions were acquired for each subject. For PDFF estimation, a long-TR, multi-TE breath-hold acquisition was used,^8,17,30 while for TC estimation a free-breathing sequence with respiratory gating and water suppression was used. Scanning parameters are given in Table 2. All spectra were shimmed during free breathing. There was no water suppression, and spatial saturation bands around the voxel were disabled to ensure a uniform spectral response across the frequency range of interest.

Table 2.

STEAM acquisition parameters for ¹H MRS estimation of PDFF and triglyceride composition

	PDFF	Triglyceride composition
TR (ms)	3,500	Respiratory gated
TE (ms)	10, 15, 20, 25, 30	10
TM (ms)	5	5
NSA	1 (per TE)	16
Voxel Size (mm)	20 × 20 × 20	20 × 20 × 20
BW (Hz)	5,000	5,000
Number of points	2,048	2,048
Water suppression	None	CHESS
Scan Time	21 s	~ 2 minutes

Notes: STEAM = stimulated echo acquisition method; MRS = magnetic resonance spectroscopy; PDFF = proton density fat fraction; TR = repetition time; TE = time to echo; TM = mixing time; NSA = number signal averages; BW = bandwidth; Hz = Hertz; CHESS = chemically selective saturation

For PDFF estimation, following a single pre-acquisition excitation, TR 3500 ms single average spectra at TE 10, 15, 20, 25 and 30 ms were acquired in a single 21 s breath-hold. This sequence was chosen as it is regarded as the standard for PDFF estimation.^3,9 For the ndb- and nmidb-estimation sequence to assess liver TC, accurate determination of the area of the individual fat peaks in the range 0–3 ppm (peaks 3, 4, 5 and 6) is required (Table 1). This is challenging for peak 3 at 2.75 ppm, a marker of polyunsaturated triglyceride. This peak is smaller than the other fat peaks, and cannot be accurately determined using a breath-held sequence except at the highest PDFF values.¹⁷ Thus, a non-breath-hold sequence with respiratory gating and water suppression (16 signal averages, approx. 2 min acquisition) was used at a single TE (10 ms) to maximize signal-to-noise ratio (SNR). The T1 relaxation times of the fat peaks are short enough (~400–800 ms) such that variation in TR due to normal respiration (TR 2.5~6 s) will not introduce T1 bias.^18,31,32 Similarly the T2 relaxation times of all the fat peaks (~50–100 ms) are long enough such that the short TE (10 ms) used here is independent of subject T2 variability.^29,31,32

MRS Analysis

Identical prior-knowledge-based analysis was used to fit all spectra from both sequences using an approach previously defined.³¹ Rather than analyze spectra from each of the eight elements in the abdomen coil, spectra from the individual channels were combined using singular value decomposition, so there was only one spectrum per TE to analyze.³³ A single experienced observer (GH) analyzed the spectra using prior knowledge in the Advanced Method for Accurate, Robust and Efficient Spectral fitting (AMARES) algorithm³⁴ included in the Magnetic Resonance User Interface (MRUI) software package.³⁵ No post-processing, such as line-broadening was applied. Each of the peaks were modeled by multiple Gaussian resonances. While the frequency of the fat peaks was fixed relative to the main CH₂ peak at 1.3 ppm (peak 5), the water and fat signals in the 4–6 ppm range were not fixed in frequency.^17,31

For the long-TR, multi-TE breath-hold acquisition, the T2-corrected peak areas of fat (0-3 ppm) and water (4–6 ppm) were estimated by non-linear least-square fitting that minimized the difference between the observed peak areas and values given by theoretical decay. PDFF was estimated as the ratio of T2-corrected fat signal to the sum of T2-corrected water and fat signals, adjusted for fat included in the ‘water’ peak from a previously established standard liver spectrum.¹⁷

For the respiratory-gated, water-suppressed acquisition, spectra were visually inspected and judged not to be evaluable, and were excluded from further analysis if TC could not be estimated due to obvious artefact, technical failure (such as failure of water suppression), or inadequate spectral quality. If one or more of the individual (i.e., fat peaks 3, 4, 5 and 6 at 2.75, 2.1, 1.3 and 0.9 ppm respectively) was not distinguishable from background noise the spectrum was classified as an SNR failure, whereas if the fat peaks could not be clearly distinguished from each other but were visible from background noise, the spectrum was classified as having failed due to indistinguishable fat. To remove possible temporal bias in determining whether spectra were evaluable, all spectra were re-inspected within a single week in large batches a year or more after the original analysis were performed, by the same observer who originally analyzed the spectra. The results of the re-inspections were used to determine spectral evaluability, rather than the assessment made at the time of the original analysis. If the spectrum was judged acceptable, the results from spectral fitting that used the same MRUI-based approach as that used for long-TR, multi-TE spectra were recorded. The areas of fat peaks 3, 4, 5 and 6 were corrected for relaxation effects using T2 values from literature (Table 1).¹⁷

A description of the MRS method employed to calculate ndb and nmidb has already been published ^17,18. Briefly, the relative area of each peak was found by adding the number of hydrogen nuclei with its associated type of bond in the triglyceride molecule. This gives the relative area of the peaks shown in Table 1. As allowing the CL to vary freely produced unstable ndb and nmidb estimates, CL was fixed at 17.5 to match observed values in previous studies.^17,21 The ndb and nmidb values were calculated by non-linearly minimizing the difference between the estimated areas of the fat peaks 3,4,5 and 6, and that given by the theoretical model.

Statistical Analysis

Study sample characteristics were summarized. The relationship between success rate of TC assessment (i.e., of the success rate of ndb and nmidb estimation) and PDFF was summarized. A receiver operating characteristic (ROC) curve examining the success rate of TC assessment was generated and used to identify the PDFF threshold from the optimal combination of sensitivity and specificity that separated success from failure. Logistic regression was used to examine the relationship between PDFF above the selected threshold and the probability of failure. For comparison of TC and PDFF, if a subject had scans at multiple time points, only the first scan was used as that was the scan most likely to occur without treatment or intervention. A scatterplot of ndb and nmidb values vs. PDFF was generated, and the relationship was examined using linear regression and Pearson’s correlation. Ndb and nmidb estimations were modeled as a constant plus linear term in PDFF: ndb = ndb₀ + ndb₁·PDFF and nmidb = nmidb₀ + nmidb₁·PDFF. The relationship between ndb and nmidb was examined using linear regression and Pearson’s correlation. Partial correlation between ndb and nmidb without the PDFF effect was assessed. A significance level of 0.05 was used in all calculations in this analysis.

RESULTS

Two hundred sixty-three MRS exams were performed on 199 adult participants (93 males, mean age at first scan of 49.0 yrs [range 22 to 75 yrs]). Of the participants who underwent multiple scans, 33 were scanned twice, five were scanned three times, and six were scanned five times.

TC could be assessed for 177/263 (67.3 %) of the acquired exams. A successful acquisition and three examples of spectra where TG could not be assessed are shown in Figure 1. The success rate of TC assessment increased with increasing PDFF: ndb and nmidb could be estimated in 7/53 (13.2%) exams with PDFF < 4%, 13/30 (43.3%) exams with PDFF between 4% and 7%, 33/41 (80.5%) exams with PDFF between 7% and 10%, and 124/139 (89.2%) exams with PDFF > 10%. Low SNR was the main cause for rejection of 42 exams (16.0%), with 31 (11.8%) rejected for indistinguishable fat peaks and 13 (4.9%) rejected for failure of water suppression or other artefact. For PDFF < 4%, low SNR was the prime reason for failure with 38 exams rejected for low SNR, 5 for indistinguishable fat peaks and 3 for artefact. For PDFF > 4% the most common reason for failure was indistinguishable fat peaks with 26 exams rejected for indistinguishable fat peaks, 10 for artefact and 4 for low SNR.

Figure 1.

Four respiratory-gated STEAM, TE 10 ms spectra with 5 Hz Lorentzian line broadening showing successful and inadequate acquisitions: A) successful (PDFF 7.6%); B) inadequate due to low SNR (PDFF 3.4%) with noise between 2.5 and 3.0 ppm obscuring the 2.75 ppm polyunsaturated peak; and C) inadequate due to indistinguishable fat peaks (PDFF 7.7%) obscuring the 0.9 ppm peak and D) inadequate due to failure of water suppression making it challenging to fit the small 2.75 ppm peak (PDFF 7.7%). The fat peak assignments are as in Table 1. Cho - choline containing compounds.

An ROC curve for classifying the success of liver TC assessment is shown in Figure 2. From the Youden Index, the threshold separating success from failure is PDFF = 6.7% which gives a sensitivity of 0.910 (95% CI: 0.857–0.947) and a specificity of 0.721 (95% CI: 0.614–0.812). For PDFF > 6.7% there was no significant relationship between PDFF and TC assessment success rate (coefficient p = 0.29). A scatterplot showing the dependence of success and failure on PDFF is shown in Figure 3, with the threshold for the Youden Index (PDFF 6.7%).

Figure 2.

ROC curve classifying the success rate of liver TC assessment (AUROC = 0.848). The optimal threshold separating success from failure (PDFF = 6.7%) is indicated by a black dot.

Figure 3.

Scatterplot showing the dependence of success and failure on PDFF with the thresholds for the Youden Index (PDFF 6.7%) shown.

One hundred thirty-eight (69.3%) adult participants had successful TC assessment on their first exam. The relationships between the determinants of TC (ndb and nmidb) in these subjects and PDFF are shown in Figure 4. Both ndb and nmidb decreased with increasing PDFF: ndb = 2.83 - 0.0161·PDFF, r = −0.449 (p < 0.0001) and nmidb = 0.75 – 0.0088·PDFF, r = −0.350 (p < 0.0001).

Figure 4.

Comparison of measures of TC assessment (ndb and nmidb) with PDFF. There is correlation between both ndb and nmidb with PDFF, showing liver becoming more saturated as PDFF increases. ndb₀ = 2.83 (95% CI 2.74 – 2.93), ndb₁ = −0.0161 (95% CI 0.0107 – 0.0215), r = −0.449 (p < 0.0001) and nmidb₀ = 0.752 (95% CI 0.681 – 0.823), nmidb₁ = −0.0088 (95% CI −0.0128 – −0.0048), r = −0.350 (p < 0.0001).

Ndb and nmidb were highly correlated (nmidb = −0.921 + 0.594·ndb, r = 0.847 p < 0.0001). Partial correlation between ndb and nmidb with the effect of PDFF removed was similarly strong (r=0.823, p < 0.0001).

DISCUSSION

In this study, we confirmed that both measures of liver TC (ndb and nmidb) decreased with increasing PDFF. Further, we found that the ¹H MRS TC assessment success rate increased with increasing PDFF, and failure was not significantly associated with PDFF above 6.7%.

Variation in liver TC with PDFF was not observed in previous studies.^12–14 However, those were ex vivo studies on small patient cohorts. Our data showed considerable scatter of ndb (and nmidb) compared to PDFF, so it is possible that this/these relationships would have been unapparent for smaller numbers of subjects.

Our data suggest that the correlation of measures of liver TC with PDFF may be confounding comparison of liver TC in NASH and non-NASH groups, as these groups typically have different mean PDFFs. However, the relationship between liver TC and PDFF displays considerable scatter and when comparing ndb and nmidb, PDFF has almost no effect on their relationship suggesting that there are factors other than just PDFF driving the behavior of liver TC.

A previous study using similar MRS techniques also did not detect a change in ndb with PDFF.¹⁷ However, as the authors of that study stressed, they did not use a ¹H MRS sequence optimized for assessing TC, rather using a sequence optimized for PDFF estimation. Thus, their data would have been weighted towards values from spectra with higher PDFF values. Further, as fat peak 3 (at 2.75 ppm) was only evident in their sequence for subjects with PDFF > 30%, the authors assumed that the relative area of this peak compared to other fat peaks was the same for all subjects. Our study suggests that this was an incorrect assumption. This may explain why measures of liver TC assessed in that study (ndb = 1.92, nmidb = 0.32) were most similar to values seen at the highest PDFF values in our study. The low-PDFF TC observed in our study is similar to that assessed in adipose tissue using a similar technique.^18,21

Chemical-shift-encoded MRI (CSE-MRI) uses a multi-peak spectral fat model to estimate PDFF¹⁷ that has been widely applied for CSE-MRI PDFF estimation both in research and in some commercially available clinical MRI-based techniques. While this standard model differs from our findings, the effect on PDFF estimation of varying the liver spectral model across a range of biological plausible TC values has been examined and found to be minimal.³⁶ Although the maximum ndb in that study (2.7)³⁶ was slightly lower than our ndb₀ value of 2.83, that study suggests that the standard model is good enough, and spectral models that account for changing liver TC with PDFF are not required for CSE-MRI PDFF estimation.

A previous study indirectly observed that TC changed with PDFF whilst examining bias in MRI-PDFF measurements.²⁸ To provide a consistent estimate of PDFF, that study hypothesized that liver triglyceride became more saturated as PDFF increased. That study’s estimated value for 0^th-order change (ndb₀ = 2.74) is close to our estimated value (ndb₀ = 2.83), and although that study was not able to estimate the 1^st-order change rate in ndb with PDFF, that change rate in our study (nmidb₁ = −0.0160) was similar to their hypothesized value (ndb₁ = −0.01). The prior study was unable to estimate change in nmidb. Another MRI study on volunteers at 7T also observed an increase in saturation with increasing PDFF, but this change did not reach significance due to a small cohort size.²⁶

Further fine tuning of the TC-assessment sequence may improve SNR by increasing the voxel size or acquiring MR spectra with more averages. However, larger voxel size will produce poorer shims and wider linewidths, which may adversely affect the ability to differentiate fat peaks from each other, and the SNR increase available at clinically feasible scan times is marginal as SNR is proportional to the square root of the number of averages.

Assessments of spectral evaluability were subjective. For this approach to apply in a wider setting, standardized criteria for evaluability for assessing liver TC will be required and eventually, as with spectral analysis, the evaluability analysis would be automated, removing operator bias.

Although for PDFF > 6.7% the success rate of TC assessment was independent of PDFF, this technique has lower success rates at low, but clinically important PDFF values. For example, thresholds for detecting NAFLD with PDFF in the range 3.4 to 6.9% have been proposed,³⁷ and so there is interest in achieving a better understanding of PDFF behavior in this range. However, our technique has only a moderate success rate in this range.

Our study is an exploratory study. As such, it did not attempt to adjust the estimated PDFF or assessments of TC for subject demographics, or for the presence of NASH. An MR imaging study using the same triglyceride model that we used here found that liver triglyceride in NASH subjects was more saturated than in those with simple NAFLD.¹⁹ Future studies are required to examine the importance, if any, of our findings that ndb and nmidb decreased (i.e., that liver triglyceride became more saturated) with increasing PDFF. Although biopsies were collected on some of the subjects in our study, these biopsies were performed for clinical care and thus were not available for assessing TC. Hence, we cannot confirm that the values of ndb and nmidb estimated by ¹H MRS reflect the actual TC. Though this technique has previously been verified in phantoms,¹⁷ in vivo studies comparing MR estimates of TC with biopsy-derived TC have been limited to adipose tissue³⁸. Further studies in liver comparing MRS with biopsy-derived estimates of TC and disease severity are required, as the challenges in estimating TC in liver are greater than those in adipose tissue, and liver studies are required to investigate the clinical applicability of this technique.

In conclusion, we have confirmed in a population of adults with known or suspected NAFLD that liver triglyceride becomes more saturated as PDFF increases, have shown at 3T that the ability of ¹H MRS to assess liver TC successfully increased as PDFF increased, and that for PDFF > 6.7%, the success rate of TC assessment was independent of PDFF.

Abbreviations:

AMARES

advanced method for accurate, robust and efficient spectral fitting

CL

fatty-acid chain length

MRUI

Magnetic Resonance User Interface

NAFLD

nonalcoholic fatty liver disease

NASH

nonalcoholic steatohepatitis

ndb

number of double bonds

nmidb

number of methylene-interrupted double bonds

PDFF

proton density fat fraction

PRESS

point resolved spectroscopy

ROC

receiver operating characteristic

STEAM

stimulated echo acquisition method

TC

triglyceride composition