Complex Confounder-Corrected R2* Mapping for Liver Iron Quantification with MRI

Abstract

Objectives:

MRI-based R2* mapping may enable reliable and rapid quantification of liver iron concentration (LIC). However, the performance and reproducibility of R2* across acquisition protocols remain unknown. Therefore, the objective of this work was to evaluate the performance and reproducibility of complex confounder-corrected R2* across acquisition protocols, at both 1.5T and 3.0T.

Methods:

In this prospective study, 40 patients with suspected iron overload and 10 healthy controls were recruited with IRB approval and informed written consent, and imaged at both 1.5T and 3.0T. For each subject, acquisitions included four different R2* mapping protocols at each field strength, and an FDA-approved R2-based method performed at 1.5T as a reference for LIC. R2* maps were reconstructed from the complex data acquisitions including correction for noise effects and fat signal. For each subject, field strength, and R2* acquisition, R2* measurements were performed in each of the nine liver Couinaud segments, and the spleen. R2* measurements were compared across protocols, and field-strength (1.5T and 3.0T), and R2* was calibrated to LIC for each acquisition and field strength.

Results:

R2* demonstrated high reproducibility across acquisition protocols (p>0.05 for 96/108 pairwise comparisons across 2 field strengths and 9 liver segments, ICC > 0.91 for each field strength/segment combination), and high predictive ability (AUC > 0.95 for four clinically relevant LIC thresholds). Calibration of R2* to LIC was [LIC=−0.04+2.62×10⁻² R2*] at 1.5T and [LIC=0.00+1.41×10⁻² R2*] at 3.0T.

Conclusions:

Complex confounder-corrected R2* mapping enables LIC quantification with high reproducibility across acquisition protocols, at both 1.5T and 3.0T.

Introduction

Abnormal deposition of iron is a common clinical disease state resulting from two major causes: hereditary hemochromatosis[1; 2] and transfusional hemosiderosis in patients requiring repeated blood transfusions[3–7]. Excess iron is toxic[1; 8–11], and can lead to liver damage[8; 12], cirrhosis, liver failure, and hepatocellular carcinoma, as well as many other comorbidities such as pancreatic dysfunction[13], pituitary dysfunction[14], and cardiomyopathy[15]. Current treatment for iron overload depends on its etiology, and includes phlebotomy (typically in patients with hemochromatosis)[12; 16], and chelation therapy (typically in patients with transfusional hemosiderosis)[17; 18]. A critical component of the treatment for iron overload is accurate assessment of body iron stores.

Most clinicians rely on serum ferritin for the assessment of body iron stores. Although serum ferritin concentration correlates with body iron stores[8], ferritin levels are confounded by other factors including inflammation, and may not accurately reflect body iron levels[19–22]. Liver iron concentration (LIC, mg Fe/g dry tissue) is widely considered the best biomarker to assess total body iron stores, as the liver is the only organ whose iron content is consistently increased with increased body iron[9; 20; 23; 24]. Historically, non-targeted liver biopsy with biochemical quantification of iron concentration has been used for detection and treatment monitoring. However, biopsy is expensive, invasive, and many patients have coexisting thrombocytopenia, a contraindication to biopsy, due to the risk of uncontrolled bleeding. In addition, biopsy-based LIC measurement has wide sampling variability[25; 26], which is highly undesirable for treatment monitoring.

MRI is highly sensitive to abnormal deposition of iron[27–30]. Spin echo-based R2 mapping methods have been calibrated against biopsy and an R2-based method[31] is currently commercially available and FDA-approved. The main disadvantage of R2 relaxometry is the lengthy acquisition time (10–20 minutes), incomplete coverage of the liver, and ghosting artifacts related to the free breathing acquisition.

As an alternative, gradient echo (GRE)-based R2* methods have a linear relationship with iron concentration, and 3D GRE data can be acquired rapidly over the entire liver in a single breath hold[32; 33]. However, previous magnitude-based R2* relaxometry methods have suffered from several major confounding factors, including the presence of noise floor effects[34], fat[35; 36], and background magnetic field inhomogeneity effects near air-tissue interfaces[37; 38]. Recently developed complex chemical shift encoded (CSE)-MRI techniques that use both the magnitude and phase of the GRE signal have shown promise to address these challenges and enable confounder-corrected R2* mapping[11; 33; 34; 39; 40]. Therefore, the purpose of this work was to validate a complex-based confounder-corrected CSE-MRI method to measure R2* as a quantitative biomarker of liver iron concentration.

Materials and Methods

Study Participants

This was a prospective, HIPAA compliant study performed after obtaining approval from the local Institutional Review Board (IRB). After obtaining informed consent, 40 study participants with known or suspected iron overload from the local hematology clinic were recruited. Inclusion criteria included those with known or suspected iron overload, age 10 and up. Determination of iron overload was based on known elevated serum ferritin (>500ng/mL), or an established diagnosis of a genetic condition that predisposed patients to iron overload, such as hereditary hemochromatosis. Exclusion criteria included contraindications to MRI. Patients undergoing phlebotomy or chelation therapy were not excluded.

In addition, 10 healthy controls with no known history of iron overload were recruited from an IRB-approved database of healthy volunteers. The same exclusion criteria applied to the control subjects. The number of subjects in this study was driven by budget and feasibility constraints.

43 of the 50 subjects recruited for this study have been previously reported for separate analyses, including validation of a method to measure magnetic susceptibility directly from B0 field maps[41], validation of an algorithm for quantitative susceptibility mapping[42], and assessment of fat quantification methods in the presence of liver iron[43]. In contrast, this study evaluates confounder-corrected R2* as a reproducible biomarker of iron overload.

Study Visit

Study visits occurred between 1/2012 and 5/2015. Each participant underwent a single visit, which included MRI at 1.5T and 3.0T, performed within approximately one hour. Imaging was performed on clinical MRI systems (1.5T: HDxt; 3.0T: Discovery MR750, GE Healthcare) using a phased array torso coil (1.5T: 8-channel; 3.0T: 32-channel). In addition, all participants underwent phlebotomy for serum ferritin levels immediately before or after MRI.

Confounder-Corrected R2* Quantification

An investigational version of a 3D multi-echo spoiled gradient echo acquisition was implemented at both 1.5T and 3.0T[44]. In order to assess reproducibility across acquisition parameters, multiple R2* mapping acquisitions with different parameters were obtained by constructing four unique protocols at both 1.5T and 3.0T (see Table 1). These protocols were chosen to obtain data with varying spatial resolution, image orientation, echo spacing and number of echoes. The goal of varying these parameters is to determine their effect on R2* bias and variability, as introduced by noise floor effects, fat signals, and macroscopic B0 field variations.

Table 1:

MRI acquisition protocols.

	1.5T					3.0T
	Ferriscan	Protocol 1	Protocol 2	Protocol 3	Protocol 4	Protocol 1	Protocol 2	Protocol 3	Protocol 4
Orientation	Axial	Axial	Axial	Coronal	Axial	Axial	Axial	Coronal	Axial
Slice (mm)	6	8	6	5	8	8	6	5	8
Slices	11	32	36	40	32	32	32	40	32
FOV (cm)	44×33	40×36	40×36	40×36	40×36	40×32	40×32	40×40	40×32
Matrix size	256×256	256×160	256×160	256×160	144×128	256×144	256×144	256×144	128×128
Voxel dimensions (mm) AP×RL×SI	1.7×1.7×6.0	2.5×1.6×8.0	2.5×1.6×6.0	5.0×2.5×l.6	3.1×2.8×8.0	2.8×1.6×8.0	2.8×1.6×6.0	5.0×2.8×l.6	3.1×3.1×8.0
Flip angle (°)	90	5	5	5	5	4	4	4	4
TR (ms)	1000	14.1	14.1	14.1	11.0	8.6	8.6	8.6	5.9
Number of echoes	5	6	6	6	12	6	6	6	8
ETL	1	6	6	6	6	3	3	3	4
TE1	6	1.2	1.2	1.2	0.9	1.2	1.2	1.2	0.6
ΔTE	3	2.0	2.0	1.0	0.7	1.0	1.0	1.0	0.6
Acquisition time (min)	17:00	0:19	0:21	0:21	0:22	0:22	0:22	0:23	0:20

For each R2* mapping acquisition, R2* maps were reconstructed using five different algorithms (Table 2). These algorithms included four combinations of: magnitude-based (using magnitude echo images while discarding the phase) versus complex-based fitting, and fat-uncorrected versus fat-corrected (including simultaneous fat-water separation and R2* fitting). Note that complex-based fitting has been shown to avoid noise floor related bias, whereas fat-correction has been shown to avoid R2* bias due to the presence of fat-related signal oscillations[34]. In order to address the instability of fat-corrected R2* fitting at very high R2* values[34], a complex-based hybrid R2* fitting algorithm was also used to return fat-corrected R2* fitting at moderate R2* values (for R2*<500s⁻¹ at 1.5T, and for R2*<1000s⁻¹ at 3.0T), and fat-uncorrected R2* fitting at higher R2* values (R2*≥500s⁻¹ at 1.5T, and for R2*≥1000s⁻¹ at 3.0T). Given that eight different R2* mapping acquisitions were performed on each subject (two field strengths, four protocols per field strength), these five reconstructions result in a total of 40 R2* maps per subject.

Table 2:

R2* measurement methods.

	Magnitude fat-uncorrected	Magnitude fat-corrected	Complex fat-uncorrected	Complex fat-corrected	Complex hybrid
Signal fitting method	Least-squares	Least-squares	Least-squares	Least-squares	Least-squares
Magnitude vs. complex	Magnitude	Magnitude	Complex	Complex	Complex
Fat correction	None	Multi-peak fat model	None	Multi-peak fat model	Hybrid: multi-peak fat model up to R2* threshold

R2-Based LIC Quantification as Reference

Each patient underwent LIC measurement using a commercially available R2-based spin echo technique (FerriScan, Resonance Health)[31] at 1.5T. Briefly, this standardized protocol consists of free breathing, 2D spin echo images obtained with 11 axial slices over the liver, using a repetition time (TR) of 1,000ms and five distinct echo times (TE) (6,9,12,15,18ms).

Data Analysis

After reconstruction of R2* maps, images were loaded into an analysis software package (Osirix, Pixmeo). Circular regions of interest (ROI) were placed within each of the nine Couinaud segments of the liver (ie: segments I, II, III, IVa, IVb, V, VI, VII, VIII) as well as the spleen to estimate the average local R2* value. ROIs as large as possible to avoid large vessels, bile ducts, liver lesions, or obvious image artifacts were measured using a standardized approach described by Campo et al[45]. For R2* maps reconstructed using the same source echo data but in a different manner (eg: magnitude vs complex, Table 2), ROIs were copied and pasted to ensure identical regions were interrogated. For R2* maps generated from different acquisitions, attempts were made to co-register the ROIs across the three axial acquisitions at each field strength, using the copy/paste function, and any slight manual adjustment needed to account for slight differences in breath-holding. From each ROI, the mean R2* value was recorded. Finally, the average liver R2* value across the nine Couinaud segments was also computed.

In addition, co-localized measurements of proton-density fat-fraction (PDFF) were performed from protocol 4 acquired at 1.5T using a complex fat-corrected reconstruction to determine the prevalence of hepatic steatosis in this population and the potential confounding effects of fat.

Also, 1.5T spin echo images were uploaded to an independent core lab (Resonance Health), using secure web transfer. The core lab returned a summary LIC value (mg Fe/g dry tissue), which was used as the reference value for each subject.

Statistical Analysis

To assess reproducibility and bias in measured R2* between protocols, two-tailed, paired Wilcoxon rank sum tests for each protocol pair (six pairs) within each segment, reconstruction and field strength combination.

For brevity, subsequent analyses were limited to the complex hybrid R2* reconstruction. To further assess the reproducibility across R2* acquisition protocols, intraclass correlation coefficients (ICC) were estimated from linear mixed models[46]. To estimate the correlation between liver-averaged R2* and LIC, for each protocol and field strength, Spearman’s ρ, a rank-based correlation coefficient, was calculated between R2* and LIC (along with 95% BCa-bootstrapped confidence intervals). Pearson’s correlation coefficient was estimated to assess the relationship between serum ferritin and LIC at the subject level. Corresponding calibration curves (slope and intercept) were estimated using simple linear regression. To compare protocols across liver-averaged R2* measurements (complex hybrid reconstruction only, as seen in Figure 4) at 1.5T versus 3.0T, a linear mixed model was fit to the data where individual patient was modeled as a random effect. The resulting model p-values were estimated using Satterwhaite’s approximation.

Figure 4:

Regression analysis between liver-averaged R2* measurements (complex hybrid reconstruction) at 1.5T versus 3.0T. High correlation with a slope close to 2.0 is observed for each of the acquisition protocols. The slope of protocol 3 was found to be significantly lower than protocols 1, 2 and 4 (p<.001).

To assess between-segment variability in R2* (eg: due to the presence of segment-specific macroscopic B0 field-related bias in R2*), eight linear mixed effect models (one for each protocol in each field strength) were fit to the complex hybrid reconstruction R2* measurements. Each model adjusted for segment (a nine-level factor) as a fixed effect, where R2* measurements in segment VI were used as the reference (as this segment is generally free of macroscopic B0 related R2* bias or other obvious artifacts), and the individual patient was modeled as a random effect.

The association between LIC and both R2* and ferritin was also investigated using receiver operating characteristic (ROC) curves. The LIC variable was dichotomized (above or below) according to four threshold values, 1.8, 3.2, 7 and 15 mg/g dry, which are commonly used thresholds for normal (<1.8), mild (1.8–3.2), moderate (3.2–7), severe (7–15) and extreme (>15) iron overload, respectively[31; 47; 48]. To calculate common model metrics assessing discriminatory abilities such as sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV), an optimal threshold was calculated using Youden’s index (J) which equally weights sensitivity and specificity.

Finally, the association between liver-averaged PDFF and liver-averaged R2* (both measured at 1.5T from protocol 4) was assessed using Spearman’s rank correlation. All statistical analyses were conducted using R (V 3.6.1, R Core Team)[49; 50].

Results

Forty patients with known or suspected iron overload (age (mean±std): 43.9±21.1, age range: 11–79, 28 men and 12 women) were successfully recruited. Eleven patients had hereditary hemochromatosis, one patient with hyperferritinemia not otherwise specified, and 28 patients with transfusional hemosiderosis from a variety of causes. Causes of transfusional hemosiderosis included myelodysplastic syndrome (7), acute myelocytic leukemia (5), acute leukocytic leukemia (3), aplastic anemia (2), lymphoma (2), leukemia (1), sickle cell disease (1), myelofibrosis (1), and other anemias (6). Table 3 summarizes the diagnosis and/or etiology of hepatic iron overload. Further, we recruited 10 healthy control subjects (age (mean±std): 41.0±15.6, age range: [24,73], 5 men and 5 women). One patient was unable to undergo MRI due to large body habitus. For the remaining 39 patients and 10 healthy control subjects, acquisition of all R2* protocols at 1.5T and 3.0T, as well as spin echo R2-based relaxometry at 1.5T was successful in all subjects with no technical failures. The measured ROIs in the liver (across all segments) had area 3.6±1.9 cm². The sizes of the ROIs by liver segment, as well as in the spleen, are reported in Table S1 (Supplemental Material).

Table 3:

Subject characteristics (HH: hereditary hemochromatosis, TH: transfusional hemosiderosis, MDS: myelodysplastic syndrome, AA: aplastic anemia, AML: acute myeloid leukemia, ALL: acute lymphoblastic leukemia, SCD: sickle cell disease, NOS: not otherwise specified).

Subject#	Reason for suspected iron overload (grouped)	M/F	Age (years)	LIC (mg Fe/g dry)	Serum ferritin (ng/mL)
Patients
1	TH (MDS)	F	70	4.5	1045
2	TH(MDS)	M	66	13.8	4074
3	TH (Other anemia)	M	23	1.4	183
4	TH (AA)	F	19	2.2	523
5	TH (AML)	M	50	7.2	926
6	TH (ALL)	F	12	19.2	2634
7	HH	M	48	1.6	293
8	HH	M	59	0.9	1206
9	TH (MDS)	F	66	7.6	1022
10	TH (AML)	M	61	12.6	2806
11	HH	M	35	3.1	397
12	TH (AML)	F	14	1	76
13	TH(AA)	M	11	2.3	133
14	TH (SCD)	F	16	7.2	2484
15	TH (ALL)	F	24	3.4	1003
16	TH (Lymphoma)	M	79	1.7	951
17	TH (Lymphoma)	M	57	1.9	596
18	TH (Other anemia)	M	15	4.4	1697
19	TH (AML)	F	19	12.7	4852
20	TH (MDS)	F	64	2.1	4178
21	HH	M	33	7.1	1565
22	HH	M	21	6	372
23	HH	F	68	8	867
24	TH (MDS)	M	69	4.7	2144
25	HH	M	60	2	686
26	TH (Leukemia)	M	69	3.6	1165
27	HH	M	41	0.4	562
28	TH	M	57	10.8	2382
29	TH (MDS)	M	67	9.2	3366
30	TH(MDS)	M	n/a: large body habitus, did not complete MRI
31	TH (Other anemia)	M	35	2.7	4
32	TH (AML)	F	53	14.1	7427
33	TH (Other anemia)	M	63	3	2830
34	HH	M	50	5.7	2555
35	TH (ALL)	F	20	9.8	1865
36	HH	M	57	15.7	1240
37	TH (AML)	M	17	4.5	709
38	TH (Other anemia)	M	54	3.8	2094
39	TH (Other anemia)	M	21	7	1515
40	Hyperferritinemia (NOS)	M	51	3.3	531
Healths controls
1		M	26	0.6	139
2		F	63	1.4	39
3		F	47	1.1	30
4		M	25	1	118
5		M	44	0.3	14
6		M	29	1	38
7		F	62	0.8	11
8		F	57	1.4	110
9		M	24	1.3	150
10		F	33	0.6	23

Figure 1 shows representative examples of two subjects with moderate iron overload, and severe iron overload, showing R2* maps acquired at 1.5T, 3.0T, as well as R2*-based LIC maps (produced using the R2*-LIC calibration described below). R2* maps without obvious artifacts were obtained at both field strengths over a wide range of LIC values. Further, even though R2* values are field strength-dependent, LIC maps (obtained based on the calibration described below) are field strength-independent.

Figure 1:

Representative R2* maps (top two rows) and corresponding LIC maps (bottom two rows) for a patient with moderate iron overload (left) and a patient with severe iron overload (right). The LIC maps were obtained from the R2* maps using the calibration derived in this manuscript. R2* and LIC maps are shown for both 1.5T and 3.0T acquisitions, demonstrating the field strength dependence of R2*, and the field strength independence of LIC maps. (SF: Serum ferritin concentration).

Figure 2 evaluates the reproducibility of each R2* reconstruction method across acquisition protocols by assessing pairwise protocol biases. In this figure, six pairwise comparisons between R2* measured with four different acquisition protocols are shown, for different combinations of field strength (1.5T and 3.0T), measurement location (nine Couinaud liver segments, spleen and overall liver mean), and reconstruction method. For each of these comparisons, a colored tile denotes a significant result (p<0.05) from the paired Wilcoxon rank-sum test at the corresponding field strength, reconstruction and segment combination. A significant p-value provides evidence of a possible protocol-level bias for that particular combination. For most segments and at each field strength, complex fitting based R2* mapping results in no significant difference (p>0.05) in R2* measurements between each pair of protocols, providing no evidence of a bias and demonstrating reproducibility of R2* mapping. In contrast, magnitude fitting R2* measurements, both fat-corrected or uncorrected, result in significant differences (p<0.05) between acquisition protocols for a large number of liver segments as well as the spleen.

Figure 2:

R2* mapping with complex reconstruction is highly reproducible across acquisition protocols. The plots show p-values evaluating the presence of systematic differences in R2* for each pair of acquisition protocols. This comparison was performed for each reconstruction algorithm (magnitude fat-uncorrected, magnitude fat-corrected, complex fat-uncorrected, complex fat-corrected, and complex hybrid), for each liver segment as well as the spleen and liver mean, and for each field strength (1.5T and 3.0T). For most segments and at both field strengths, complex fitting based R2* mapping results in no significant difference (p>0.05) in R2* measurements between each pair of protocols. In contrast, magnitude fitting R2* measurements result in significant differences (p<0.05) between acquisition protocols for many field strength and segment combinations.

Subsequent analysis was focused on the complex hybrid R2* reconstruction. This reconstruction leads to R2* that is highly reproducible across acquisition protocols, with an ICC between 0.91–1.00 for each liver segment (between 0.98–1.00 for liver-averaged R2*) at both 1.5T and 3.0T. To illustrate these comparisons across protocols, Figure S1 (Supplemental Material) shows the six Bland-Altman plots comparing protocols 1, 2 and 3 to protocol 4 for the mean liver R2*, in the complex hybrid reconstruction (for both field strengths).

Figure 3 shows correlation analysis between complex hybrid fitting mean liver R2* (obtained by averaging R2* measurements across all nine Couinaud liver segments) and LIC, for each of the four R2* mapping protocols acquired at each field strength. The calibration and correlation coefficient for each of these protocols between R2* and LIC are shown in Table 4.

Figure 3:

Regression analysis and calibration between liver-averaged R2* measurements (complex hybrid reconstruction) from each of the acquisition protocols and LIC. This analysis is shown separately for 1.5T and 3.0T based R2* measurements. Similar calibrations are obtained at each field strength across acquisition protocols.

Table 4:

Regression analysis between liver-averaged R2* and LIC, for various R2* acquisition protocols, using complex hybrid reconstruction.

Field Strength (T)	Protocol	Intercept ± SE (mg/g)	Slope ± SE (×10−2 mg/g/s−1)	Correlation
1.5	1	−0.19 ± 0.40	2.71 ± 0.16	0.94
1.5	2	−0.20 ± 0.39	2.75 ± 0.16	0.94
1.5	3	−0.26 ± 0.42	2.78 ± 0.17	0.94
1.5	4	−0.04 ± 0.39	2.62 ± 0.15	0.95
3.0	1	−0.16 ± 0.40	1.47 ± 0.09	0.94
3.0	2	−0.12 ± 0.41	1.46 ± 0.09	0.94
3.0	3	−0.29 ± 0.54	1.56 ± 0.13	0.92
3.0	4	−0.00 ± 0.42	1.42 ± 0.09	0.93

Figure 4 shows correlation between mean liver R2* measured at 1.5T versus R2* measured at 3.0T, for each of the four acquisition protocols reconstructed with the complex hybrid approach. High correlation was observed for each of the protocols (1: r=0.997, 2: r=0.992, 3: r=0.983, 4: r=0.996), with slope close to 2.0, as expected[51; 52]. The slope of the acquisition protocol 3 was significantly lower than that of protocols 1, 2 and 4, as assessed by the linear mixed model (p<.001 for each comparison).

Figure 5 shows ROC analysis for the diagnostic accuracy of R2* to diagnose different thresholds of iron overload using LIC as the reference. For this plot, mean liver R2* from the acquisition protocol 4 and complex hybrid reconstruction 4 were used. An area under the ROC curve higher than 0.95 was obtained for all LIC thresholds under consideration (1.8, 3.2, 7.0, 15.0 mg Fe/g dry), at either field strength. Detection of these four LIC thresholds can be performed from 1.5T R2* measurements with thresholds 97.2, 131.3, 273.0, and 459.1 s⁻¹, respectively. Similarly, detection of these four LIC thresholds can be performed from 3.0T R2* measurements with thresholds 152.7, 236.6, 350.6, and 839.7 s⁻¹, respectively (see table S2 in Supplemental Material for details).

Figure 5:

ROC analysis for R2*-based detection of clinically relevant LIC thresholds (1.8, 3.2, 7.0, 15.0 mg/g dry) demonstrates excellent predictive ability for each of the LIC thresholds, using liver-average R2* measured from protocol 4 (complex hybrid reconstruction) at both 1.5T and 3.0T.

No significant segment variability was observed for R2* measurements in the four 1.5T models, or in the 3.0T models for protocols 2 and 4, when comparing to the reference R2* measurements from liver Couinaud segment VI (p>0.05 for each Couinaud segment in each model). Within the model fit to the 3.0T and protocol 3 data, segment IVa and segment VIII had significantly higher mean R2* measurements compared to segment VI (p<0.05). With regard to the model fit to the 3.0T and protocol 1 data, segment II had significantly lower mean R2* measurements compared to segment VI (p<0.01).

Figure 6 shows a scatter plot between serum ferritin concentration and LIC with the associated linear regression curve, as well as ROC plot for serum ferritin concentration to diagnose different thresholds of iron overload using LIC as the reference. Although a significant correlation exists between serum ferritin concentration and LIC (r=0.68, p<.0001), serum ferritin concentration results in a moderate area under the ROC curve for detection of higher LIC thresholds (eg: 0.76 for LIC>15 mg Fe/g dry). Detection of the four LIC thresholds considered in this work (1.8, 3.2, 7.0, and 15.0 mg Fe/g dry) can be performed from serum ferritin measurements with thresholds 372, 709, 867, and 1240 ng/mL, respectively (see table S2 in Supplemental Material).

Figure 6;

Correlation analysis shows moderate correlation between serum ferritin concentration and LIC (r=0.68). ROC analysis for serum ferritin-based detection of clinically relevant LIC thresholds demonstrates good predictive ability at low LIC thresholds, and moderate accuracy at high thresholds (eg: 15.0 mg/g dry).

Liver PDFF measurements were 4.7±6.6% across all subjects (min PDFF: −2.3%, max PDFF: 34.4%). No significant correlation (based on Spearman’s rank correlation analysis) was found between PDFF and R2* across the reconstructed data sets (p=0.84).

Discussion

In this work we investigated the performance of a complex confounder-corrected R2* mapping method to quantify hepatic iron overload. We evaluated eight different acquisition protocols (four at 1.5T and four at 3.0T), using an FDA-approved spin echo based LIC quantification technique as the reference. A linear relationship with excellent correlation was observed between all R2* protocols at both 1.5T and 3.0T with the LIC reference. Further, R2* values across the different protocols were highly reproducible, even over a wide range of spatial resolution, echo times, number of echoes, and image plane orientation, which indicates insensitivity to the confounding effects of macroscopic susceptibility, as well as noise floor effects. Therefore, we have successfully validated confounder-corrected CSE-MRI to measure R2* as an accurate and reproducible (across protocols) quantitative biomarker of liver iron overload.

In addition, we confirmed the known relationship of R2* at 1.5T versus 3.0T, as predicted theoretically[53] and as measured in previous studies[51; 52]. Further, comparison of R2* values obtained at both 1.5T and 3.0T on the same day demonstrated very strong correlation, with regression parameters similar to those observed in previous works, eg: slope of 1.86 and intercept of −6.4 s⁻¹ in our study (protocol 4) compared to slope of 1.92 and intercept of −9.1 s⁻¹ in Meloni et al [52]. In combination with the excellent agreement of R2* measured from different acquisition parameters at a single field strength (also shown in this manuscript), this provides indirect evidence that the precision of R2*-based relaxometry may be very high. Further work with test-retest repeatability studies will be needed to confirm this speculation.

In addition, this work demonstrated excellent diagnostic accuracy of R2* for the detection and differentiation of different stages of iron overload based on clinically accepted LIC thresholds. Comparison of serum ferritin levels obtained on the same day as MRI showed a significant but moderate correlation with LIC.

A widely recognized R2* versus LIC calibration performed by Wood et al[32], demonstrates similar, although slightly different calibration compared to our results. It is important to note that the technique described by Wood was markedly different, as it did not correct for the presence of fat, noise floor effects, used a large number of echo times, and used relatively large voxels (15 mm slices) and single slice 2D acquisition. Similarly, a comparison between (1.5T) R2* and FerriScan-based LIC was also recently performed using magnitude-based R2* fitting[40], also leading to a similar calibration to the one shown in this work. Due to these technical differences in R2* mapping methods, it is difficult to make a direct comparison of the complex confounder-corrected CSE-MRI method used in this work. We do note, however, that we have successfully demonstrated excellent reproducibility across different protocols acquiring data with different echo spacing, different echo times, and different voxel sizes. Thus, the generalizability of the calibration determined in this work should be very broad. Despite the high reproducibility across acquisition protocols, R2* mapping in the presence of high iron overload will benefit from acquisitions with high SNR and short TEs. For this reason, from the acquisitions performed in this study, we expect protocol 4, which has the shortest echo times, to provide the broadest dynamic range.

There are several limitations of this study. First, detailed effects on the etiology of iron overload on the R2*-LIC calibration, as well as the effects of phlebotomy or chelation therapy were not examined. Another limitation is the lack of biopsy data. A follow-up study including biopsy calibration would be highly informative. However, liver biopsies for LIC measurement are rarely performed at our institution, so conducting such biopsy calibration studies may be challenging. Importantly, the reference standard used in this study is FDA-approved, based on previous liver biopsy-MRI correlation studies. Thus, the LIC values used in this reference are equivalent to those obtained from liver biopsy. Finally, this prospective study reports the results from a single center study using a single MRI vendor. Future validation across multiple centers, vendors, and platforms are needed to demonstrate the widespread reproducibility of this technique across sites and MRI vendors and platforms.

Importantly, this study included relatively few patients with severe iron overload. Further work validating the methods in populations with higher iron burdens (e.g. thalassemia major) are needed. One additional technical limitation of the R2* mapping described in this work is that it is based on a Cartesian acquisition with an initial echo time of approximately 1ms. As a result, accurate measurement of R2* in patients with extreme iron overload may be compromised, i.e.: the dynamic range of this technique may be limited. Non-Cartesian radial ultra-short TE (UTE) techniques may enable a broader dynamic range for those patients with extreme iron overload[54–56].

In conclusion, we have successfully validated a complex confounder-corrected CSE-MRI method to quantify R2* as an accurate and reproducible biomarker of liver iron concentration in a prospective single center study. Further, we have provided calibration between R2* and LIC, at both 1.5T and 3.0T, enabling estimation of LIC at these two important clinical field strengths, using complex gradient echo methods which can be acquired in a single breath-hold. Given the widespread commercialization of 3D multi-echo gradient echo methods for R2*-corrected fat quantification, translation of the proposed technique in this work into a commercially available application on all vendor platforms should be relatively straightforward. Further work validating such methods in multi-center studies will be needed.

Key points:

• Confounder-corrected R2* of the liver provides reproducible R2* across acquisition protocols, including different spatial resolutions, echo times, and slice orientations, at both 1.5T and 3.0T.

• For all acquisition protocols, high correlation with R2-based liver iron concentration (LIC) quantification was observed.

• The calibration between confounder-corrected R2* and LIC, at both 1.5T and 3.0T is determined in this study.

List of abbreviations:

AA

aplastic anemia

ALL

acute lymphoblastic leukemia

AML

acute myeloid leukemia

CSE

Chemical Shift-Encoded

GRE

Gradient-Recalled Echo

HH

hereditary hemochromatosis

IRB

Institutional Review Board

LIC

Liver Iron Concentration

MDS

myelodysplastic syndrome

NOS

not otherwise specified

SCD

sickle cell disease

TH

transfusional hemosiderosis