Diagnostic Accuracy of Magnetic Resonance Imaging Hepatic Proton Density Fat Fraction in Pediatric Nonalcoholic Fatty Liver Disease

Abstract

We assessed the performance of magnetic resonance imaging (MRI) proton density fat fraction (PDFF) in children to stratify hepatic steatosis grade before and after treatment in the Cysteamine Bitartrate Delayed-Release for the Treatment of Nonalcoholic Fatty Liver Disease in Children (CyNCh) trial, using centrally-scored histology as reference. Participants had multi-echo 1.5T or 3T MRI on scanners from three manufacturers. Of 169 enrolled children, 110 (65%) and 83 (49%) had MRI and liver biopsy at baseline and at end-of-treatment (EOT; 52-weeks), respectively. At baseline, 17% (19/110), 28% (31/110), and 55% (60/110) of liver biopsies showed grades 1, 2, and 3 histologic steatosis; corresponding PDFF (mean ± standard deviation) values were 10.9 ± 4.1%, 18.4 ± 6.2%, and 25.7 ± 9.7%, respectively. PDFF classified grade 1 vs. 2–3 and 1–2 vs. 3 steatosis with areas under receiving operator characteristic curves (AUROCs) of 0.87 (95% confidence interval [CI]: 0.80, 0.94) and 0.79 (0.70, 0.87), respectively. PDFF cut-offs at 90% specificity were 17.5% for grades 2–3 steatosis, and 23.3% for grade 3 steatosis. At EOT, 47% (39/83), 41% (34/83), and 12% (10/83) of biopsies showed improved, unchanged, and worsened steatosis grade, respectively, with corresponding PDFF (mean ± standard deviation) changes of −7.8 ± 6.3%, −1.2 ± 7.8% and 4.9 ± 5.0%, respectively. PDFF change classified steatosis grade improvement and worsening with AUROCs (95% CIs) of 0.76 (0.66, 0.87) and 0.83 (0.73, 0.92), respectively. PDFF change cut-off values at 90% specificity were −11.0% and +5.5% for improvement and worsening.

Conclusion:

MRI-estimated PDFF has high diagnostic accuracy to both classify and predict histologic steatosis grade, and change in histologic steatosis grade in children with NAFLD.

Nonalcoholic fatty liver disease (NAFLD) is the most common pediatric chronic liver disease (1). Common co-morbidities of NAFLD include diabetes and cardiovascular disease (2–7). The long-term natural history of NAFLD is not well understood but the disease has the potential to progress to cirrhosis, hepatocellular carcinoma, and early death (8–11), and is rapidly becoming the leading cause of liver transplantation (12).

Hepatic steatosis, a key feature of NAFLD, is histologically graded on biopsy according to the proportion of hepatocytes containing fat macrovesicles on hematoxylin and eosin staining (grade 0: < 5%; grade 1: 5–33%; grade 2: 34 to 66%; and grade 3: > 66% (13). However, liver biopsy is invasive and samples only a very small portion of the liver, and thus may not be ideal for longitudinal clinical trials or for clinical monitoring and care in the early stages of the disease. Computed tomography is not well suited for use in serial-monitoring because it uses ionizing radiation and shows only modest association with liver fat content, and conventional ultrasound is of limited value because it provides only semi-quantitative estimates of liver fat content and does not permit assessment of all liver segments (14). Hepatic steatosis may also be assessed non-invasively by a transient elastography-derived controlled attenuation parameter, which is an estimate of total ultrasound attenuation (15). Although we did not perform a direct comparison of the controlled attenuation parameter and PDFF, the accuracy of MRI to estimate a fat fraction has been reported to be better than that obtained from the controlled attenuation parameter in adults, using histology as reference standard (16,17). The accuracy of the controlled attenuation parameter to assess steatosis in children is not known.

In contrast, multi-echo magnetic resonance imaging (MRI) is emerging as the non-invasive method of choice to estimate, on a continuous scale, hepatic proton density fat fraction (PDFF) (10). Single-center cross-sectional studies have shown MRI-estimated PDFF to be accurate, using magnetic resonance spectroscopy (18–22) or histology (23–27) as reference standards, and to be reproducible across field strengths (28–31) and MRI scanner manufacturers (28,31), and a recent meta-analysis has shown that MRI-estimated PDFF is reproducible across field strength and scanner manufacturer (32).

Several single-center studies have examined the diagnostic performance of hepatic PDFF to grade histologic steatosis. In a single-center study of 89 adults (33), PDFF correlated with histologic steatosis grade, and classified steatosis grade 0 vs. grade 1 at the 6.4% PDFF cutoff reported by Tang et al (23), with a sensitivity of 68%, a specificity of 98%, and an area under receiving operator characteristic curve (AUROC) of 0.82. In adults with nonalcoholic steatohepatitis (NASH) in the multi-center Farnesoid X Receptor Ligand Obeticholic Acid in NASH Treatment (FLINT) trial, PDFF correlated with histologic steatosis grade, and accurately classified baseline dichotomized steatosis grades 0–1 vs. grades 2–3, and grades 0–2 vs. grade 3 with AUROCs of 0.95 and 0.96 at cutoffs of 16.3% and 21.7%, respectively, with corresponding sensitivities of 83% and 84% at 90% specificity (34,35). In the same trial, PDFF change correlated with histologic steatosis grade change, and accurately classified steatosis grade change longitudinally (35); at end-of-treatment, PDFF change classified steatosis grade change, with cutoffs at 90% specificity of −5.1% for improvement and +5.6% for worsening, both with AUROCs of 0.81.

Assessing multi-center PDFF diagnostic performance in children is necessary to further validate PDFF as a biomarker of hepatic steatosis prior to its widespread use in clinical care or as an endpoint in pediatric clinical trials (34). Few studies assessing PDFF validation metrics have been conducted in children (36–40), and none in a multi-center setting. In the MRI Rosetta Stone Project, 174 children at a single center had hepatic MRI-estimated PDFF and histology evaluated (36). PDFF correlated with steatosis grade (r = 0.725, p < 0.01), and the overall accuracy of predicting histologic steatosis grade from PDFF was 56%. However, the relationship between change in hepatic PDFF and change in histologic steatosis grade has not been evaluated in children. Therefore, the purpose of this study was to assess cross-sectional and longitudinal diagnostic performance of hepatic PDFF to grade histologic steatosis in children with NAFLD using centrally-scored histology as the reference standard.

Patients and Methods

STUDY DESIGN

We performed a prospectively-designed study of the diagnostic performance of PDFF estimated by multi-echo MRI as part of the Cysteamine Bitartrate Delayed-Release for the Treatment of Nonalcoholic Fatty Liver Disease (NAFLD) in Children (CyNCh) trial (NCT01529268), a multi-center, randomized, double-masked, placebo-controlled, phase 2b clinical trial of treatment with either cysteamine bitartrate delayed release or placebo in children with NAFLD (41). Liver biopsy and MRI were performed at baseline and after 52 weeks of treatment allowing paired comparisons of PDFF and histologic hepatic steatosis grade, and their longitudinal changes. PDFF was a secondary endpoint, for which centrally-scored histologic hepatic steatosis grade from percutaneous biopsy served as the reference standard.

Eligibility criteria for the CyNCh trial are published elsewhere and were based on well-established biopsy criteria for pediatric NAFLD (41). Inclusion criteria for MRI were enrollment in the CyNCh trial, and willingness and ability to complete both MRI exams (the baseline exam prior to randomization and within 90 days of baseline biopsy, and the end-of-treatment (EOT) MRI exam within 120 days of EOT biopsy). Exclusion criteria for MRI were contraindication to MRI, extreme claustrophobia, weight or girth exceeding MRI scanner capability, or any condition or circumstance that, in the opinion of the clinical trial site investigator would interfere with completion of the exam.

The CyNCh study protocol, including the MRI portion conformed to the ethical guidelines of the 1975 Declaration of Helsinki as reflected in a priori approval by the appropriate institutional review committee at each participating clinical trial site, and was in compliance with the Health Insurance Portability and Accountability Act. Children aged 8 to 17 years at enrollment were included in this study, with all children providing written informed assent with written informed consent by a parent or guardian.

All authors had access to the study data, and reviewed and approved the final manuscript.

MRI CLINICAL TRIAL SITES

Nine of the ten participating CyNCh clinical trial sites contributed MRI data to this study using 1.5T (five sites), 3T (three sites), or both 1.5 and 3T (one site) MR scanners (Table 1).

Table 1.

MRI scanners and techniques

Clinical trial site MR scanners	1.5T	3T
Site A		General Electric HDxt
Site B	General Electric Signa HDx, and Philips Ingenia	-
Site C	General Electric Signa HDxt	-
Site D	Siemens Aera	Siemens TrioTim
Site E	-	Siemens TrioTim
Site F	Siemens Aera	-
Site G	Siemens Avanto	-
Site H	-	General Electric Signa HDxt
Site I	Siemens Avanto	-
Site J	-	-
Parameter
TR (ms)	≥ 120	≥ 120
First TE (ms)	2.3	1.15
Delta TE (ms)	2.3	1.15
Number of echoes	6	6
Flip angle (°)	10	10
Bandwidth (Hz/Px)	≥ 500	≥ 1,000
Slice thickness (mm)	8 or 10	8 or 10
Slice gap (mm)	0	0
Phase encoding steps	192	128
Frequency encoding steps	192	128

Notes: TR = repetition time; TE = time to echo; General Electric Healthcare (Waukesha, WI, USA); Philips Healthcare (Best, The Netherlands), Siemens Healthcare (Erlangen, Germany)

MRI CLINICAL TRIAL SITE QUALIFICATION AND QUALITY CONTROL

The NASH Clinical Research Network (CRN) Radiology Coordinating Center (RCC), in conjunction with the NASH CRN Data Coordinating Center managed the MRI portion of this study. Each individual site was approved by the RCC based on approval of technically adequate phantom or volunteer MRI data. The RCC also provided central imaging quality control.

MRI ACQUISITION AND ANALYSIS

MRI acquisition was identical to that recently reported for the MRI sub-study to the FLINT trial (35), using a non-contrast, breath-hold, gradient-recalled-echo, two-dimensional axial sequence. Images of the entire liver were obtained using a torso array coil centered over the upper abdomen. Parameters were selected to correct for or avoid confounding factors (e.g., T1 bias, T2* decay, multi-frequency interference) that could introduce fat quantification error (10,18–22) (Table 1). Images in Digital Imaging and Communications in Medicine format were transferred from clinical trial sites to the RCC.

MRI analysis was performed at the RCC in the following fashion. For each MRI, signal intensities from the 6-echo magnitude spoiled-gradient-echo source images were analyzed pixel-by-pixel using a custom MATLAB™ (The MathWorks, Natick, MA, USA) non-linear, least-squares fitting algorithm to produce a PDFF parametric map at each image level throughout the liver (18–22, 42). The algorithm computed the PDFF values of those parametric maps by assuming exponential T2* signal decay and applying a multi-peak spectral model to account for fat-fat and fat-water multi-frequency interference effects, based on the work of Hamilton et al (43). One circular 1-cm radius region of interest (ROI) was placed the on 5th echo (out-of-phase) source images in each of the nine anatomical liver segments. Those ROIs were propagated to the corresponding PDFF parametric map images, and the mean PDFF value from each evaluable parametric map-based ROI was recorded. For each MRI, the number of ROIs that were evaluable was recorded. For MRIs that were not evaluable, reasons were recorded.

LIVER BIOPSY

In the CyNCh trial, liver biopsies were performed for clinical care within 90 days of the start of screening and no more than 120 days before randomization and demonstrated histologic evidence of NAFLD as well as a NAFLD activity score (NAS) of ≥ 4, as scored by the individual NASH CRN Pathology Committee member at each study site. Liver biopsies were performed at 52 weeks to determine the response to therapy (41).

Pathologists of the NASH CRN Pathology Committee reviewed the biopsies conjointly at a multi-head microscope as part of the CyNCh trial (41). The pathologists scored steatosis grade (proportion of hepatocytes containing fat macrovesicles: grade 0 for < 5%, grade 1 for 5 to 33%, grade 2 for 34 to 66%, and grade 3 for > 66% (13)) according to the NASH CRN histologic NAFLD scoring system that also includes scoring criteria for accompanying histologic features such as steatosis location, lobular inflammation, portal inflammation, hepatocellular ballooning, fibrosis stage, iron grade and location, and global biopsy diagnosis. The NAFLD Activity Score was calculated from unweighted sums of steatosis, lobular inflammation, and hepatocellular ballooning (13).

BLINDING

RCC analysts and other staff at the RCC were blinded to histology results, and pathologists were blinded to all clinical and imaging results. RCC analysts were blinded to treatment assignment. Pathologists knew each slide was obtained from a participant enrolled in or being considered for enrollment in a NASH CRN study, but they were blinded to all clinical information, including age, time point (baseline or EOT), or enrollment in the CyNCh MRI sub-study.

OTHER DATA

Participant demographics, laboratory, anthropomorphic measurements, and medical history were collected at each clinical trial site.

STATISTICAL ANALYSIS

Statistical analyses were done with SAS (SAS Institute 2011, Base SAS 9·3 Procedures Guide) and Stata (StataCorp 2013, Stata Statistical Software: release 13).

A single composite PDFF value was calculated for each MRI as the mean of the PDFF values for the nine anatomical liver segments. Demographic, histologic and imaging information were summarized with categorical variables expressed as numbers and percentages and continuous variables expressed by mean(± standard deviation [SD]). The proportion of participants with, and without MRIs at baseline, and for those with MRIs at baseline, the proportion of those with, and without MRIs at EOT were compared with regard to treatment group, study site, demographics, liver enzymes, lipids, metabolic factors, co-morbidities, concomitant liver medications, and histology findings.

Histologic components at baseline were linearly regressed on PDFF at baseline, and 52-week changes in histologic features were linearly regressed on 52-week changes in PDFF adjusting for baseline value of histologic feature. Analyses of follow-up data at 52 weeks were pooled across treatment groups. Beta values (mean histologic component score change per 1% increase in PDFF, 95% confidence intervals (CIs), and p-values were estimated between PDFF and histologic components (steatosis score, lobular and portal inflammation scores, hepatocellular ballooning score, and fibrosis score) at baseline and for changes from baseline to 52 weeks.

Diagnostic accuracy of PDFF to classify hepatic steatosis grade at baseline was tested for grades 1 vs. 2–3, and grades 1–2 vs. 3. Diagnostic accuracy of change in PDFF to classify change in hepatic steatosis grade from baseline to EOT was tested for reduction vs. no change/increase, and increase vs. no change/decrease. Cross-validated AUROCs using a jack-knife procedure and 95% CIs were estimated for each of these dichotomizations (44). Cut-off PDFF values were estimated using the lowest threshold value for which there was ≥ 90% specificity to distinguish between these dichotomized categories. Sensitivity, positive predictive values, and negative predictive valves were calculated along with 95% CIs fixing specificity at 90%.

Interaction between subgroups for MRI-determined steatosis vs. histologic steatosis was tested at baseline and longitudinally for the following dichotomized subgroups: age (8–12 vs. 13–18 yrs), sex, lobular inflammation score (grade 1 inflammation at baseline vs. grade 2–3 inflammation at baseline), fibrosis (no fibrosis at baseline vs. any fibrosis at baseline), scanner type (1.5T vs. 3T), and time from baseline biopsy to baseline MRI (≤ 60 days vs. > 60 days, based on the midpoint of the distribution of these times). Because of the exploratory nature of these new analyses and the increased likelihood of Type I error due to multiple comparisons, we chose the cut-off for statistical significance as 0.01.

Units of PDFF change are expressed in absolute units of change in percent PDFF. Thus, a decrease from 10 to 5% PDFF would represent a change of 5% PDFF.

Results

Of 169 children enrolled in the CyNCh trial from June 2012 to January 2014 at ten participating CyNCh clinical trial sites, all (n = 169) underwent liver biopsies at baseline, and 146 (86%) had liver biopsies at EOT. One hundred ten participants (65%) had MRI at baseline, and 85 (50%) had MRI at baseline and EOT. Of the 195 MRIs obtained for study participants, one EOT MRI was excluded because of poor image quality (phase artifact, and low signal-to-noise ratio). PDFF values were derived from all nine anatomical liver segments for 192/194 of the remaining MRIs; one baseline-only MRI had 2/9 segments that were not evaluable, and one EOT MRI had 1/9 segments that were not evaluable. Eighty-three participants (49%) had MRI and liver biopsy at both time points. Baseline MRIs occurred from 0 to 115 days after initial biopsy (mean 61 days). EOT MRIs that were paired with baseline MRIs ranged from 78 days pre-EOT biopsy to 61 days post-EOT biopsy (mean 2 days pre-EOT biopsy).

There were no statistically significant differences in study parameters among children with, and without MRI at baseline, except for differences by clinical site (p < 0.0001), and that children who had baseline MRIs had a lower mean fasting glucose than those who did not have baseline MRIs (p = 0.04). There were also no statistically significant differences in study parameters among children who had MRIs at baseline, and those with, and without MRIs at EOT, except again for differences by clinical site (p < 0.0001), and that children who had no EOT MRI were older (p = 0.02) and had lower mean alkaline phosphatase (p = 0.008) compared to those who had only baseline MRIs. Study parameter differences for these groups are given in Table 2.

Table 2.

Baseline characteristics of: a) those obtaining vs. not obtaining baseline MRI, and b) of those obtaining baseline MRI, those obtaining vs. not obtaining EOT MRI

	Baseline vs. no baseline MRI (n = 169)			EOT vs. no EOT MRI for those who had baseline MRI (n = 110)
	Number participants who had baseline MRIs (n=110)	Number participants who did not have baseline MRIs (n=59)	p-value	Number participants who had EOT MRIs (n=85)	Number participants who did not have EOT MRIs (n=25)	p-value
Study site			< 0.0001			< 0.0001
Site A	10	13		4	6
Site B	16	5		14	2
Site C	17	4		13	4
Site D	6	10		2	4
Site E	12	1		7	5
Site F	6	6		4	2
Site G	2	3		2	0
Site H	40	2		39	1
Site I	1	3		0	1
Site J	0	12		0	0
Treatment group
cysteamine bitartrate delayed release – no. (%)	53 (48.2)	35 (59.3)	0.20	39 (45.9)	14 (56.0)	0.50
Demographics
Age (yrs)	13 ± 3	13 ± 3	0.68	12.7±2.5	14.2±2.8	0.02
Male sex - no. (%)	78 (70.9)	41 (69.5)	0.86	60 (70.6)	18 (72.0)	1.00
White race - no. (%)	64 (58.2)	41 (69.5)	0.18	45 (83.3)	19 (100.0)	0.10
Hispanic ethnicity - no. (%)	78 (70.9)	46 (78.0)	0.37	60 (70.6)	18 (72.0)	1.00
Liver enzymes
Aspartate aminotransferase - U/L	64 ± 48	84 ± 74	0.07	64 ± 52	63 ± 32	0.86
Alanine aminotransferase – U/L	110 ± 85	146 ± 124	0.05	108 ± 91	119 ± 64	0.48
γ-Glutamyltransferase – U/L	48 ± 34	45 ± 25	0.46	46 ± 34	57 ± 36	0.19
Alkaline phosphatase - U/L	223 ± 106	214 ± 114	0.63	237 ± 105	174 ± 97	0.008
Total bilirubin – mg/dL	0.5 ± 0.3	0.6 ± 0.3	0.09	0.46 ± 0.33	0.58 ± 0.23	0.05
Metabolic factors
Weight - kg	85 ± 25	85 ± 25	0.92	83 ± 26	89 ± 24	0.30
Body-mass index - kg/m2	32 ± 6	33 ± 7	0.50	32 ± 6	33 ± 6	0.27
Waist circumference - cm	104 ± 14	102 ± 16	0.34	104 ± 15	104 ± 13	0.80
Waist to hip ratio	0.98 ± 0.06	0.97 ± 0.08	0.26	0.99 ± 0.06	0.97 ± 0.07	0.48
Fasting serum glucose - mg/dL	86 ± 12	90 ± 11	0.04	85 ± 12	89 ± 11	0.19
Insulin - umol/mL	37 ± 34	47 ± 93	0.40	35 ± 28	44 ± 51	0.40
HOMA-IR - (mg/dL * umol/mL) / 405	7.8 ± 7.7	10.2 ± 19.2	0.35	7.2 ± 5.8	9.7 ± 12.1	0.34
Hemoglobin A1c - %	5.4 ± 0.4	5.4 ± 0.4	0.47	5.5 ± 0.5	5.4 ± 0.4	0.27
Systolic blood pressure - mmHg	119 ± 11	120 ± 11	0.44	119 ± 12	119 ± 10	0.87
Diastolic blood pressure - mmHg	67 ± 9	67 ± 8	0.66	67 ± 9	69 ± 8	0.26
Comorbidities
Hyperlipidemia - no. (%)	19 (17.3)	6 (10.2)	0.26	17 (20.0)	2 (8.0)	0.23
Hypertension - no. (%)	12 (10.9)	3 (5.1)	0.26	10 (11.8)	2 (8.)	0.73
Diabetes - no. (%)	6 (5.4)	2 (3.4)	0.71	6 (7.1)	0 (0.0)	0.33
Liver histology findings
Definite steatohepatitis - no. (%)	30 (27.3)	17 (28.8)		22 (25.9)	8 (32.0)	0.61
Fibrosis - stage1	1.2 ± 1.1	1.2 ± 1.1	0.60	1.3 ± 1.0	1.2 ± 1.1	0.82
NAFLD Activity Score2	4.7 ± 1.4	4.7 ± 1.5	0.96	4.6 ± 1.4	4.8 ± 1.2	0.48
Hepatocellular ballooning - score	0.6 ± 0.7	0.6 ± 0.8	0.77	0.6 ± 0.8	0.6 ± 0.6	0.94
Steatosis - score	2.4 ± 0.8	2.4 ± 0.7	0.77	2.3 ± 0.8	2.5 ± 0.7	0.26
Lobular inflammation - score	1.7 ± 0.7	1.7 ± 0.7	0.62	1.7 ± 0.7	1.7 ± 0.6	0.99
Portal inflammation - score3	1.1 ± 0.5	1.0 ± 0.5	0.05	1.2 ± 0.5	1.1 ± 0.6	0.54

Notes: Plus-minus values are means ± SD; BMI = body mass index; PDFF = proton density fat fraction; SD = standard deviation.

¹Fibrosis was assessed on a scale of 0 to 4, with higher scores indicating more severe fibrosis

²NAFLD Activity Score was assessed on a scale of 0 to 8, with higher scores indicating more severe disease; the components of this measure are steatosis (assessed on a scale of 0 to 3), lobular inflammation (assessed on a scale of 0 to 3), and hepatocellular ballooning (assessed on a scale of 0 to 2)

³Portal inflammation was assessed on a scale of 0 to 2 with higher scores indicating more severe inflammation

CROSS-SECTIONAL ANALYSIS

The distribution of PDFF in the 110 children at baseline, all of whom were diagnosed with NAFLD is shown in Figure 1: PDFF mean ± SD was 21.1 ± 9.8%, and ranged from 5.3% to 46.8%.

Figure 1.

PDFF distribution of study population at baseline.

At baseline, biopsy findings included grade 1 steatosis in 17% (19/110), grade 2 steatosis in 28% (31/110), and grade 3 steatosis in 55% (60/110) of participants. Corresponding mean ± SD PDFF values were 10.9 ± 4.1%, 18.4 ± 6.2%, and 25.7 ± 9.7%, respectively (Figure 2).

Figure 2.

Bar and whisker plot at baseline of PDFF vs. steatosis grade.

Linear regressions of PDFF values on histologic components at baseline are summarized in Table 3. PDFF was positively associated with steatosis score [0.068 mean difference in histologic steatosis score per 1% increase in PDFF; 95% CI: 0.056, 0.081; p < 0.001]. No associations with PDFF at baseline were found for lobular or portal inflammation scores, hepatocellular ballooning score, or fibrosis score (p-values 0.10 to 0.66), and all of the CIs for those regressions included 0.

Table 3.

Linear regression of PDFF with histologic components at baseline, and change in PDFF with change in histologic components at 52 weeks

Histologic component	PDFF (%) at baseline (n=110)			Change in PDFF (%) at 52 weeks (n=83)
	β*	95% CI	p-value	β†	95% CI	p-value
Steatosis (score)	0.068	0.056, 0.081	< 0.001	0.057	0.034, 0.079	< 0.001
Lobular inflammation (score)	0.012	−0.002, 0.025	0.10	0.019	0.000, 0.037	0.73
Portal inflammation (score)	0.011	−0.005, 0.018	0.24	0.005	−0.011, 0.015	0.77
Hepatocellular ballooning (score)	0.003	−0.012, 0.019	0.66	0.013	−0.010, 0.025	0.40
Fibrosis (stage)	0.013	−0.010, 0.036	0.28	0.008	−0.023, 0.030	0.80

Notes: PDFF = proton density fat fraction; CI = confidence interval

^*Mean difference in histologic component per 1% increase in PDFF

^†Mean change in histologic component per 1% increase in change of PDFF adjusted for baseline value of histologic component

Table 4 summarizes the diagnostic accuracy of PDFF for classifying steatosis. Using PDFF as a classifier, the AUROCs from logistic regression were 0.87 (95% CI: 0.80, 0.94) for classifying steatosis grade 1 vs. 2–3, and 0.79 (95% CI: 0.70, 0.87) for classifying steatosis grade 1–2 vs. 3. PDFF cut-off values at 90% specificity were 17.5% for grades 2–3, and 23.3% for grade 3 discrimination.

Table 4.

Diagnostic accuracy of PDFF for classifying steatosis

Cross-sectional steatosis grade classification (n=110)	PDFF threshold (%) at 90% specificity		Sensitivity (%)	PPV (%)	NPV (%)	Correctly classified (%)	AUROC (95% CI)
1 vs. 2–3	17.5		74 (67/91)	97 (67/69)	41 (17/41)	76 (84/110)	0.87 (0.80, 0.94)
1–2 vs. 3	23.3		60 (36/60)	88 (36/41)	65 (45/79)	74 (81/100)	0.79 (0.70, 0.87)
Longitudinal steatosis grade change classification (n=83)	Mean (SD) PDFF change (%)	Cutoff PDFF change at 90% specificity (%)	Sensitivity (%)	PPV (%)	NPV (%)	Correctly classified (%)	AUROC (95% CI)
Decrease in steatosis grade (improvement) (47%)	−7.8 ± 6.3	−11.0	31 (12/39	75 (12/16)	60 (40/67)	63 (52/83)	0.76 (0.66, 0.87)
No change in steatosis grade (41%)	−1.2 ± 7.8	-	-	-	-		-
Increase in steatosis grade (worsening) (12%)	4.9 ± 5.0	5.5	40 (4/10)	33 (4/12)	92 (65/71)	83 (69/83)	0.83 (0.73, 0.92)

Notes: PDFF = proton density fat fraction; PPV = positive predictive value; NPV = negative predictive value; AUROC = area under receiver operating characteristic curve; CI = confidence interval; SD = standard deviation

LONGITUDINAL ANALYSIS

Figure 3 shows box plots of PDFF values for each histologic hepatic steatosis grade change category (reduction, no change, and increase in steatosis grade). Mean ± SD of change in PDFF values for children with decreased histologic steatosis grade at EOT was −7.8 ± 6.3% (range: −20.1 to 1.7%) in 39 children with improvement (reduction) in steatosis grade (28 reduced one grade, eight reduced two grades, and three reduced three grades); −1.2 ± 7.8% (range: −6.8 to 12.1%) in 34 children with no change in steatosis grade; and 4.9 ± 5.0% (range: −0.5 to 12.1%) in 10 children with worsening (increase) in steatosis grade (nine increased one grade, and one increased two grades).

Figure 3.

Bar and whisker plot of change in PDFF vs. change in histologic hepatic steatosis grade. Mean (± SD) PDFF change values for steatosis grade reduction, no change in steatosis grade , and increase in steatosis grade were, respectively: −7.8 ± 6.3% (n = 39), −1.2 ± 7.8% (n = 34), and 4.9 ± 5.0 % (n = 10).

Linear regressions of changes in PDFF values on changes in histologic components at 52 weeks are summarized in Table 3. Change in PDFF was positively associated with change in steatosis score [0.057 mean change in histologic steatosis score per 1% increase in change of PDFF adjusted for baseline value of histologic steatosis; 95% CI: 0.034, 0.079; p < 0.001]. No associations with change in PDFF were found for changes in lobular or portal inflammation scores, hepatocellular ballooning score, or fibrosis score (p-values 0.40 to 0.80), and all of the CIs for those regressions included 0.

Table 4 summarizes diagnostic accuracy of change in PDFF vs. change in histologic hepatic steatosis grade from baseline to EOT. The AUROCs using PDFF change to classify histologic hepatic steatosis grade improvement and worsening, respectively, were 0.76 (95% CI: 0.66, 0.87) and 0.83 (95% CI: 0.73, 0.92). Cut-off values for PDFF change at 90% specificity were −11.0% for improvement and +5.5% for worsening hepatic steatosis grade.

TESTS OF INTERACTION BETWEEN SUBGROUPS

Table 5 summarizes subgroup analyses of regression of baseline steatosis grade on baseline PDFF, and Table 6 summarizes subgroup analyses of regression of worsening or improvement in steatosis grade on 52-week change in PDFF. None of the interaction p-values were below 0.01 (only 2 out 20 were below 0.05), although the power to detect subgroup effects in this study was low.

Table 5.

Subgroup analyses of regression of baseline steatosis grade on baseline PDFF

Event	Subgroup	N	AUROC	Odds of event per 1% increase in PDFF	95% CI	p-value
Steatosis grade 2/3 vs 1
	Age 8–12 yrs	52	0.88	1.24	1.07, 1.43	0.004
	Age 13–18 yrs	58	0.89	1.34	1.11, 1.62	0.003
	Interaction†					0.51
	Female	32	0.95	1.44	1.02, 2.05	0.04
	Male	78	0.86	1.25	1.10, 1.43	0.0008
	Interaction					0.45
	Grade 1 lobular inflammation at BL	44	0.84	1.24	1.06, 1.47	0.009
	Grade 2/3 lobular inflammation at BL	66	0.91	1.32	1.11 1.56	0.002
	Interaction					0.64
	No fibrosis at BL	29	0.78	1.23	0.97, 1.56	0.09
	Any fibrosis at BL	81	0.92	1.40	1.17, 1.66	0.0002
	Interaction					0.39
	1.5T MRI scanner	46	0.88	1.28	1.08, 1.50	0.003
	3T MRI scanner	64	0.89	1.28	1.09, 1.52	0.003
	Interaction					0.95
	BL MRI ≤ 60 days of BL biopsy	54	0.96	1.53	1.18, 1.97	0.001
	BL MRI > 60 days from BL biopsy	56	0.82	1.18	1.04, 1.35	0.01
	Interaction					0.08
Steatosis grade 3 vs 1/2
	Age 8–12 yrs	52	0.76	1.12	1.04, 1.20	0.003
	Age 13–18 yrs	58	0.84	1.21	1.09, 1.33	0.0002
	Interaction					0.20
	Female	32	0.78	1.13	1.02, 1.24	0.01
	Male	78	0.82	1.18	1.09, 1.27	<0.0001
	Interaction					0.52
	Grade 1 lobular inflammation at BL	44	0.78	1.13	1.04, 1.24	0.004
	Grade 2/3 lobular inflammation at BL	66	0.82	1.16	1.07, 1.26	0.0002
	Interaction					0.68
	No fibrosis at BL	29	0.69	1.08	1.00, 1.16	0.06
	Any fibrosis at BL	81	0.85	1.22	1.12, 1.34	<0.0001
	Interaction					0.03
	1.5T MRI scanner	46	0.78	1.15	1.05, 1.26	0.003
	3T MRI scanner	64	0.81	1.15	1.07, 1.24	0.0003
	Interaction					0.99
	BL MRI ≤ 60 days of BL biopsy	54	0.79	1.16	1.06, 1.27	0.001
	BL MRI > 60 days from BL biopsy	56	0.82	1.16	1.07, 1.26	0.0004
	Interaction					0.99

^†Test of interaction of subgroup with PDFF

Table 6.

Subgroup analyses of regression of worsening or improvement in steatosis grade on 52-week change in PDFF

Event	Subgroup	N	AUROC	Odds of event per 1% increase in change in PDFF	95% CI	p-value
Worsening of steatosis grade
	Age 8–12 yrs	43	0.94	1.24	1.00, 1.56	0.05
	Age 13–18 yrs	40	0.77	1.16	1.01, 1.33	0.03
	Interaction					0.60
	Female	25	0.86	1.16	0.98, 1.38	0.08
	Male	58	0.86	1.22	1.05, 1.41	0.008
	Interaction					0.68
	Grade 1 lobular inflammation at BL	34	0.69	1.08	0.94, 1.22	0.27
	Grade 2/3 lobular inflammation at BL	49	0.96	1.46	1.12, 1.92	0.006
	Interaction					0.04
	No fibrosis at BL	20	NC*	NC	NC	NC
	Any fibrosis at BL	63	0.88	1.20	1.07, 1.35	0.002
	Interaction					NC
	1.5T MRI scanner	32	0.90	1.29	1.04, 1.59	0.02
	3T MRI scanner	51	0.82	1.16	1.01, 1.33	0.04
	Interaction					0.40
	BL MRI ≤ 60 days of BL biopsy	39	0.88	1.32	1.03, 1.69	0.03
	BL MRI > 60 days from BL biopsy	44	0.86	1.19	1.02, 1.39	0.02
	Interaction					0.50
Improvement in steatosis grade
	Age 8–12 yrs	43	0.70	0.88	0.79, 0.97	0.01
	Age 13–18 yrs	40	0.80	0.83	0.73, 0.95	0.006
	Interaction					0.54
	Female	25	0.81	0.82	0.69, 0.98	0.02
	Male	58	0.76	0.86	0.78, 0.94	0.002
	Interaction					0.67
	Grade 1 lobular inflammation at BL	34	0.81	0.81	0.70, 0.94	0.005
	Grade 2/3 lobular inflammation at BL	49	0.71	0.87	0.79, 0.97	0.008
	Interaction					0.40
	No fibrosis at BL	20	0.68	0.89	0.76, 1.04	0.14
	Any fibrosis at BL	63	0.82	0.83	0.75, 0.92	0.0003
	Interaction					0.49
	1.5T MRI scanner	32	0.71	0.88	0.78, 1.00	0.05
	3T MRI scanner	51	0.82	0.81	0.72, 0.92	0.0007
	Interaction					0.34
	BL MRI ≤ 60 days of BL biopsy	39	0.77	0.86	0.77, 0.97	0.01
	BL MRI > 60 days from BL biopsy	44	0.79	0.83	0.74, 0.94	0.002
	Interaction					0.66

^*Not calculable due to no events

Discussion

In CyNCh, a multi-center, randomized controlled trial of children with known NAFLD, we found that MRI-derived PDFF values and histologic hepatic steatosis scores were associated (p < 0.001), that changes in MRI-derived PDFF values and changes in histologic hepatic steatosis scores were associated (p < 0.001), and that baseline and longitudinal PDFF values were not associated with other histologic components (p-value range 0.10 to 0.80).

Some discordance between PDFF values and histologic steatosis scores is expected since PDFF and histology do not measure the same quantity. As emphasized previously by others (14), PDFF is a quantitative marker of MRI-visible hepatic fat content, while histology scoring by grade is a semi-quantitative assessment of the proportion of microscopically-assessed steatotic hepatocytes. PDFF and histology assess liver fat on different scales ranging from 0 to 50% (rarely > 50%), and 0 to 100%, respectively. PDFF percentages are usually less than half of histologic steatosis percentages because PDFF quantifies the ratio of MRI signal from fat to the MRI signal from water, while histologic steatosis grade scores reflect the estimated percentage of hepatocytes that contain microscopically-visible fat globules. If all histologically-examined hepatocytes were half filled with fat globules, the histologic steatosis percentage would be 100% since all cells show fat globules, and the steatosis grade would be 3, but the MRI PDFF value would be about 50%. Another possible reason for discordance between PDFF and histologic measures of steatosis is that by MRI, it is possible to estimate PDFF throughout the liver, whereas biopsy samples only a small volume of a diffuse process.

Our results show that PDFF can distinguish histologic steatosis grade 1 vs. grades 2–3 with an AUROC of 0.87 (95% CI: 0.80, 0.94). This is similar to the larger single-center study of 174 children in which Schwimmer et al (36) showed that PDFF could classify histologic steatosis grades 0 and 1 with an AUROC of 0.82. Our results additionally suggest that, at 90% specificity a PDFF cut-off value of 17.5% provides 74% sensitivity for discriminating histologic steatosis grade 1 from grades 2–3, and a value of 23.3% provides 60% sensitivity for discriminating grades 1–2 from grade 3. However, we also found that, at 90% specificity a decrease of ≥ 11.0% PDFF provided only 31% sensitivity for identifying histologic steatosis grade improvement, and an increase of ≥ 5.5% provided only 40% sensitivity for identifying histologic steatosis grade worsening.

These PDFF (17.5%, 23.3%) and change-in-PDFF (−11.0%, 5.5%) cutoffs are similar to the corresponding PDFF (16.3%, 21.7%) and change-in-PDFF (−5.1%, 5.6%) cutoffs reported in adults in the FLINT MRI sub-study (35). The associated sensitivities in this study (74% and 60% for PDFF; 31% and 40% for change in PDFF) are lower, however, than those reported for the FLINT MRI sub-study (83% and 84% for PDFF; 58% and 57% for change in PDFF). The reason for the lower sensitivity in the current study is unclear, as identical qualification procedures and imaging protocols were used. One possible explanation is that two sites changed scanners between baseline and EOT, which may have reduced the precision with which PDFF change can be measured, but this would not have impacted the performance of PDFF at baseline. Another possible explanation is that children may be less cooperative than adults. Image quality was checked as MRIs were done, but it is still possible that, although scans were deemed adequate, image quality may have been reduced compared to adults.

We found no evidence of interaction between dichotomized subgroups (age, sex, lobular inflammation, fibrosis, scanner field strength, and time between baseline biopsy and baseline MRI) for MRI-determined steatosis vs. histologic steatosis at baseline or longitudinally at or better than a conservative p < 0.01 cutoff significance level, although we acknowledge that the power to detect subgroup effects in this study was low. The lack of an interaction for magnetic field strength was expected based on theory, and is in agreement with other studies (32).

Results regarding the role of fibrosis in the previous literature differ from ours with regard to the the relationship between MRI-determined PDFF and histologic steatosis. We found that the correlation of PDFF with histologic steatosis was stronger for higher fibrosis stages, for histologic grade 1 vs. 2–3 (interaction p-value 0.39), and for histologic grade 1–2 vs. 3 (interaction p-value 0.03), although neither reached our conservative significance cutoff of p < 0.01. Idilman et al (45) found in a study of 70 adults with diagnosed NAFLD (60% with no fibrosis) that this correlation was stronger when fibrosis was absent (r = 0.86) than when it was present (r = 0.60; p = 0.02). Idilman et al (46) in a second, smaller study (18 adults with diagnosed NAFLD; 67% with no fibrosis) also found that this correlation was stronger when fibrosis was absent (r = 0.83) compared to their cohort of all cases (r = 0.76). In a larger single-center study, Schwimmer et al (36) (174 children, 29% with no fibrosis) reported that correlation between MRI-determined PDFF and histologic steatosis was weaker in children with higher stages of fibrosis (r = 0.61 for stages 2–4, p < 0.001) compared to those with stage 1 fibrosis (r = 0.78) or no fibrosis (r = 0.76). This difference for the two studies by Idilman et al may be due to the lower fibrosis stage distribution in our study (no fibrosis in 26% of subjects), or to NAFLD differences between children and adults. It is not clear why our results differ from those reported by Schwimmer et al (36); perhaps other factors may affect this correlation, such as differences across sites.

Results regarding the role of sex in the previous literature also differ from ours with regard to the the relationship between MRI-determined PDFF and histologic steatosis. Schwimmer et al (36) reported that this correlation was stronger in girls (r = 0.86) than in boys (r = 0.70; p < 0.01). We found the opposite, that these correlations were stronger in boys than in girls, although again the comparison of these subgroups in our analysis did not meet our threshold cutoff of p < 0.01. Perhaps this reflects under-powering of our study, although there is no physical basis to expect correlations of MRI-determined PDFF vs. histologic steatosis to be affected by sex.

Strengths of this study were the prospective longitudinal design, the well-characterized cohort of children with a racial/ethnic makeup representative of pediatric NAFLD in the United States, the availability of paired biopsies at baseline and at EOT, rigorous central scoring of histology by the NASH CRN Pathology Committee and central reading of MRIs by the NASH CRN RCC, and the utilization of a range of MRI scanner manufacturers at two field strengths and across multiple scanner types and study sites. Thus, our study results are likely to be generalizable to the entire pediatric NAFLD population in the United States, while also establishing the feasibility of MRI in multi-center pediatric trials.

While not in the design, nor a study goal, a possible limitation of this study is that none of the participants included in this MRI sub-study had histologic steatosis grade 0 at baseline, as, by definition, to be enrolled, participants all had NAFLD. The relevance of this is that PDFF cut-off values that might be used at baseline to separate participants with NAFLD, from those without NAFLD cannot be defined by this study. A detailed assessment of proposed cut-points for this purpose was reported in the MRI Rosetta Stone Project (36). Further multi-center studies in populations including participants with grade 0 steatosis are needed to continue to define associations of PDFF with hepatic steatosis in grade 0–1 range.

A more relevant limitation of this study was that not all participants enrolled in the CyNCh trial had MRIs, and that not all participants who had baseline MRIs also had EOT MRIs. Participation in the MRI sub-study was reduced in part because clinical trial sites for the sub-study were brought in after the main study started, as they completed the MRI qualification process. The only statistically significant differences between participants who did and who did not participate in the MRI sub-study were clinical site (p < 0.0001), and slightly lower mean fasting serum glucose (86 vs. 90 mg/dL; p = 0.04) among subjects who did not participate in the MRI sub-study. Comparing baseline characteristics between participants with MRI at baseline and EOT (n = 85) vs. those with MRI only at baseline (n = 25), statistically significant differences were clinical site (p < 0.0001), higher mean age (14.2 vs. 12.7 yrs; p = 0.02) and lower mean alkaline phosphatase (174 vs. 237 U/L; p = 0.008) in children without EOT MRI. Potentially confounding factors such as those in Table 2 were not investigated because the study was not powered to permit those investigations. While unlikely to have affected study results, there may have been a metabolic difference in these participants that is not accounted for.

In conclusion, in a well-controlled, multi-center study in children with NAFLD, PDFF and change in PDFF estimated by multi-echo MRI at sites using scanners from different manufacturers and of different field strengths showed high diagnostic accuracy (i.e., agreement) with histologic steatosis grade and change in histologic steatosis grade, respectively.

List of Abbreviations:

MRI

magnetic resonance imaging

PDFF

proton density fat fraction

CyNCh

Cysteamine Bitartrate Delayed-Release for the Treatment of Nonalcoholic Fatty Liver Disease (NAFLD) in Children

EOT

end of treatment

SD

standard deviation

AUROC

area under receiving operating characteristic curve

CI

confidence interval

NAFLD

nonalcoholic fatty liver disease

NASH

nonalcoholic steatohepatitis

FLINT

Farnesoid X Receptor Ligand Obeticholic Acid in NASH Treatment

CRN

Clinical Research Network

RCC

Radiology Coordinating Center