|Home | About | Journals | Submit | Contact Us | Français|
To optimize R2*(1/T2*) measurements for cardiac iron detection in sickle cell and thalassemia patients.
We studied 31 patients with transfusion-dependent sickle cell disease and 48 patients with thalassemia major; myocardial R2* was assessed in a single midpapillary slice using a gated gradient-echo pulse sequence. Pixel-wise maps were coregistered among the patients to determine systematic spatial fluctuations in R2*. The contributions of minimum TE, echo spacing, signal-decay model, and region-of-interest (ROI) choice were compared in synthetic and acquired images.
Cardiac relaxivity demonstrated characteristic circumferential variations regardless of the degree of iron overload. Within the interventricular septum, a gradient in R2* from right to left ventricle was noted at high values. Pixel-wise and ROI techniques yielded nearly identical values. Signal decay was exponential but a constant offset or second exponential term was necessary to avoid underestimation at high iron concentration. Systematic underestimation of R2* was observed for higher minimum TE, limiting the range of iron concentrations that can be profiled. Fat-water oscillations, although detectable, represented only 1% of the total signal.
Clinical cardiac R2* measurements should be restricted to the interventricular septum and should have a minimum TE ≤ 2 msec. ROI analysis techniques are accurate; however, offset-correction is essential.
Iron-induced arrhythmia and heart failure remain the leading cause of death in transfusion-dependent anemia (1,2). Conventional cardiac monitoring, including physical exam, electrocardiogram (ECG), and echocardiogram, fails to predict iron-induced cardiac disease prior to the onset of clinical symptoms (3–7). In the presence of tissue iron, the shortening of the relaxation time constant T2* (1/R2*) has demonstrated considerable promise for preclinical detection of cardiac iron (8–11). Patients with short T2* or high relaxation rates R2*, corresponding with myocardial iron loading, have significantly greater risk of resting systolic dysfunction and need for medications than those with values in the normal range. Unfortunately, R2* measurements in the myocardium can be confounded by susceptibility artifacts near the boundaries with cardiac veins (12) and by high spatial variation of iron deposition (13–15).
The purpose of our study was to systematically identify technical and anatomic factors that influence cardiac R2* in sickle cell and thalassemia patients. Specifically we wished to characterize the following: 1) interpatient spatial variability of R2*, 2) contributions from magnetic susceptibility artifacts in iron-loaded patients, 3) appropriate signal relaxation model, 4) best choice of minimum echo time and echo spacing, and 5) the contribution of fat-water phase interactions to the observed signal decay.
We studied 31 patients with transfusion-dependent sickle cell disease (SCD) and 48 patients with thalassemia major (TM) from January 2002 until August 2004. Informed consent was obtained from all patients on a protocol approved by the Children’s Hospital Los Angeles Committee on Clinical Investigation. Magnetic resonance imaging (MRI) examinations were performed using a four-element phased array torso coil on a 1.5-T General Electric CVi clinical scanner running system software version 9.1 (GE Medical Systems, Waukesha, WI). Myocardial R2* was assessed in a single midpapillary short axis slice using a breath-hold, cardiac-gated, segmented gradient echo sequence. Echo times (TE) of 2, 3, 4, 6, 9, 12, 15, and 18 msec; fixed TR of 21 msec; four to eight views per segment; one excitation; and symmetric centrally ordered phase encode grouping (SCOPEG) method (16,17) were used, similar to the technique by Anderson et al (10). The number of excitations was increased to three for any patient who could not adequately breathhold.
Myocardial R2* measurements were performed using custom-written software developed in MATLAB® (The Mathworks, Natick, MA). Images at different TEs were coregistered by affine transformation according to mutual information criteria to correct for variations in end-expiratory cardiac position. A region of interest (ROI) was manually selected in the interventricular septum. Two models were used for fitting signal decay, consisting of a single exponential with and without a constant offset:
where S = fitted signal, S0 = initial amplitude, TE = echo time, R2* = transverse relaxation rate, and C = constant (offset correction). The curve fitting was performed using a interior-reflective Newton method (18) with parameters initialized as follows: S0 = signal amplitude at the shortest echo time, R2* = 100 Hz, and C = 0. Equations [1a] and [1b] were fit to every pixel in the ROI, producing a distribution of R2* values. The mean from this distribution (pixel-wise estimate) was compared with the value calculated by fitting the average signal intensity within the ROI.
T2* signal decay need not be exponential, even for homogeneous media, because of contributions from spatial variations in spin density and susceptibility (19). We hypothesized that signal decay would be nonexponential for low iron concentrations (high T2* values) but resemble an exponential function with progressive iron loading. To test this hypothesis, we mapped all of the patient data to a unit amplitude exponential having a T2* of 1. Mean and standard deviation of the residual signal decay was assessed for systematic nonexponential deviation.
Cardiac R2* is elevated near the cardiac veins in subjects without iron overload; however, these susceptibility artifacts are modest (20 Hz) (12) compared with the rise associated with iron cardiomyopathy (R2* > 100 Hz) (8,10,20,21). We hypothesized systematic circumferential variations in cardiac R2* would disappear at high iron burdens. To test this hypothesis, we subdivided the myocardium into four segments (indicated by 1–4 in Fig. 1i). Segment 1 represented the interventricular septum, Segment 2 predominantly bordered the liver, and Segments 3 and 4 formed the cardiac-lung interface. The relative ratios of the segment areas were approximately equal across patients. These four segments, can be translated into the conventional framework of the 16-segment model as described by Schiller et al (22). Mean and standard deviation of R2* within each segment was tabulated for each patient accounting for variations in myocardial dimensions, orientation, and iron loads.
True cardiac R2* is unknown in all patient studies. In order to test the efficacy of our various fitting algorithms, we needed to develop realistic synthetic cardiac images that could serve as a “gold standard.” From the pixel-wise fits in our iron-overload patients we made the following observations: 1) Mean R2* ranges from 20 to 500 Hz and has a coefficient of variation of 40%. 2) Signal decay appears to have a dominant (90% amplitude) “fast” component and a weak (10% amplitude) slowly varying or constant component. The overall proton density (S0) is Gaussian distributed with a coefficient of variation of 21.8%. 3) Noise power, when corrected for Rician bias in a multicoil acquisition, is 6.6% of S0. 4) A three-fold gradient in R2* value from endocardium to epicardium, either smooth or discontinuous, is often observed in iron-loaded patients.
Simulated pixels were generated over a septal ROI having the same average dimensions and resolution as our patient studies. Each pixel was calculated according to the following equation:
where, R2*h (heart tissue) and R2*b (blood) are the rapidly and slowly decaying components, respectively, and S0 is the total proton density representing random samples from a unit Gaussian distribution having standard deviation of 0.218. R2*b was assumed to be 5 Hz, similar to oxygenated blood (19,23).
The value of R2*h varied with pixel position in the ROI. We modeled relaxivity distributions for the synthetic images using three patterns: 1) uniform, 2) linear ramp, and 3) dualtone (a step change between two levels). The linear ramp and dualtone had 50% lower average R2* values at the endocardial margin and 50% higher values at the epicardial boundary. Mean value was chosen uniformly from the range observed in patients. True values were allowed to deviate from their “expected” results by a coefficient of variation of 40%.
In order to accurately model noise bias, it was necessary to synthesize complex data. This was done by projecting the magnitude image using a constant phase angle of 0, 15, 30, and 45 degrees for the four receiver channels, respectively. Complex noise was added to each channel and a magnitude image was recreated that had the desired rectified noise characteristics.
The synthetic data was fit in the same manner as the patient data. The accuracy of pixel-wise mapping and ROI-based techniques were compared with one another and with “true” R2* values. The accuracy of exponential (Eq. [1a]) and exponential with a constant term (Eq. [1b]) was similarly compared.
Using the synthetic data sets, it was straightforward to generate arbitrary combinations of TEs to determine the effect of TE and spacing on measurement accuracy. In particular, we wanted to assess the importance of minimum TE. We compared the following: 1) TE = 2, 3, 4, 6, 9, 12, 15, 18 msec; 2) TE = 1, 2, 3, 5, 7, 10, 14, 18 msec; 3) TE = 4, 4.9, 6.1, 7.6, 9.4, 11.7, 14.5, 18 msec; and 4) TE = 5.5, 6.5, 7.7, 9.1, 10.8, 12.8, 15.2, 18 msec, with the goal of matching the number of echoes, relative echo spacing, and the longest TE while varying the shortest TE.
Figure 1a–h shows the midpapillary short-axis gradient echo images of the heart at echo times TE = 2, 3, 4, 6, 9, 12, 15, 18 msec, respectively, for a 19.5-year-old patient with TM. The patient was asymptomatic from the cardiovascular perspective and had normal ventricular function (left ventricular ejection fraction [LVEF] of 62.6%). The ferritin was 1855 ng/mL and hepatic iron was 3.8 mg/g dry tissue weight. These values were consistent with good compliance with iron chelation therapy. The profound darkening of the ventricular muscle relative to the chest-wall muscle is indicative of strong iron-based magnetic susceptibility effect. Figure 1i shows a magnified view of the myocardium (TE = 3 msec) with manually traced endocardial and epicardial walls dividing the left ventricular (LV) and right ventricular (RV) regions. The R2* measurements were made within the indicated segments (1–4) for all patients.
Figure 2a is the pixel-wise R2* map computed from the LV, shown in Fig. 1, using Eq. [1b]; bright areas represent high values. A gradient from endocardium to epicardium can be seen through most of the circumference, except at about 4 o’clock where the gradient is reversed. Figure 2b and 2c show TE-dependent mean signal decay for the interventricular septum and corresponding fits using a pure exponential (Exp) and exponential with constant term (Exp+C), respectively. A signal “plateau” at late TE is evident having a magnitude much larger than noise bias. The simple exponential model cannot simultaneously capture the plateau and the rapid signal loss at short TE, leading to systematic T2* overestimation. Inclusion of an offset into the fitting equation visibly improves the overall agreement to the signal decay curve.
Iron-rich and iron-poor regions are clearly visible in Fig. 2a. Spatial heterogeneity in R2* could potentially account for the “plateau” observed in Fig. 2b. If so, calculation of R2* on a pixel-by-pixel basis would eliminate this bias. This hypothesis can only be tested by modeling because true R2* values are unknown in patient studies. Figure 3 compares values computed from synthetic maps using both signal decay models (Exp and Exp+C). To simplify the graph, results from all three R2* distribution models have been combined together; error bars denote standard error. Simple exponential models badly underestimate true values unless an offset is included, regardless of whether fitting was performed pixel-wise or not. Pixel-wise and ROI-based methods are comparable for relaxivities less than 175 Hz but diverge for larger values. This effect was larger for greater anisotropy (dualtone > ramp > uniform) but was still relatively modest. Pixel-wise mapping accurately tracked the true R2* values up to a relaxivity of 300 Hz. Systematic underestimation for R2* > 300 Hz represents effects from an inadequately short minimum TE; this will be demonstrated later.
Figure 4 superimposes model predictions onto our patient measurements. Comparison between Exp and Exp+C models is shown in Fig. 4a for pixel-wise mapping. Individual points represent mean R2* values from each patient, with error bars representing standard deviations of the R2* histograms. The bold line represents model prediction. The model accurately predicts the systematic difference between Exp and Exp+C models. Figure 4b demonstrates a similar patient-model comparison of ROI vs. pixel-wise mapping using exponential plus constant fitting. Patient data demonstrate a very close correlation between ROI and pixel-wise techniques throughout the R2* range. No systematic underestimation was observed; however, there is a paucity of patient data above 150 Hz where differences would be maximized. Alternatively, very high iron loads may tend to be more uniformly distributed, favoring agreement between pixel-wise and ROI-based measurements.
To determine the effect of minimum TE on the accuracy of R2* estimation (Fig. 5), synthetic images and resulting relaxivities were derived at four different minimum TEs, ranging from 1 to 5.5 msec (the range of TEs used clinically). True R2* is catastrophically underestimated once values exceed the reciprocal of the minimum TE, representing biased sampling of less heavily iron-laden tissue. The underestimation observed with a minimum TE of 2 msec is concordant with the underestimation seen in Fig. 3 (x-symbols in both graphs).
We postulated that signal decay curves would be non-exponential at low iron concentrations. Figure 6a demonstrates normalized signal decay curves for all 79 patients; intensity values were ROI-based from the interventricular septum. Data are perfectly superimposed on the unit exponential function, e−t, indicating that a monoexponential with a variable constant provides an excellent fit to the experimental data at all iron loads.
The fitting residuals (Fig. 6b) provide insight into systemic errors in the fitting model. Error bars represent standard error and the solid line represents a cubic-spline through the sampled TEs. Alternating dotted and dashed vertical lines indicate the TEs at which water and fat protons are out-of-phase and in-phase, respectively, at 1.5 T. The residual signal oscillates with a period of 4.6 msec and the first two minima and maxima align closely with predicted phase oscillations of fat and water. Our temporal sampling was too coarse (below Nyquist rate) to visualize oscillations at longer TE. Peak amplitude averages to about 1% of S0. Although small, signal modulation effects of water-fat proton spin differences can subtly alter the measured R2* depending on the choice of TE.
Figure 7a demonstrates mean-normalized circumferential R2* variability within the myocardium moving across each of the four segments 1–4 (Fig. 1i) with the conventional names given on the horizontal axis. All unloaded patients (T2* > 20 msec, roughly half of the TM and all of the SCD patients) had nearly identical relaxivity distribution patterns with a mean R2* of 38.03 Hz. Values were nearly three-fold greater at the boundary of segments 2 and 3 relative to segment 1 in these patients. Segment 4 values were 50% higher, in agreement with prior studies in normal controls (12). Surprisingly, strong circumferential variation was also observed in iron-loaded patients (T2* < 20 msec) having a mean R2* of 157.26 Hz. Oscillations were slightly less dramatic but had a similar pattern to the non iron-overload patients, suggesting that geometric interactions other than proximity to cardiac veins are responsible. Septum R2* values were within 20% of the global mean. Figure 7b demonstrates endocardial to epicardial changes in R2* within the interventricular septum with mean values of: SCD = 24.49 Hz, TM (T2* > 20 msec) = 25.93 Hz, and TM (T2* < 20 msec) = 145.39 Hz. A large gradient (greater than two-fold) can be observed for patients with high values, consistent with preferential epicardial iron deposition described in autopsy studies (13–15).
While cardiac R2* measurements demonstrate promise for the preclinical detection of cardiac iron, there has been little critical assessment of the technical factors limiting accuracy of this technique. In this paper we demonstrate the critical importance of using offset-correction and short minimum echo times, validate the use of an exponential signal decay function and ROI-based measurements, and highlight the pitfalls of R2* measurements outside the interventricular septum.
The source of the signal “offset” is probably multifactorial, containing potential contributions from noise bias, heterogenous iron distribution in myocytes, analog-to-digital signal offsets, or contributions from iron poor tissue such as fibroblasts, or myocardial blood volume. A similar offset has been described in liver iron studies as well (24,25). Regardless of the etiology, neglecting this component leads to large underestimation errors. We chose to fit the second component with a constant, rather than a second exponential, because it provided more robust fitting (fewer degrees of freedom). Our residual error from the patient data (Fig. 6b) demonstrated no systematic “drift,” suggesting that this approximation is valid. Nonetheless, constrained biexponential fitting should work in this scenario and has been successfully applied to this problem in liver R2 estimation (26–28). Alternatively, one can use an exponential fit but “discard” later echoes where the plateau becomes evident. This approach has been used by some but introduces unnecessary subjectivity.
Although the signal decay function is bicomponent, we demonstrate that the dominant component of relaxation had a Lorentzian lineshape (exponential decay, Fig. 6a). Since the decay contains contributions from both microstructural and long-range magnetic inhomogeneities, the behavior had the potential to be quite complex, particularly when T2* was in the normal range.
Figure 6a explains the relatively poor robustness of T2* methods for estimating “normal” values (T2* > 20 msec) (10). Even with a maximum TE of 18 msec, there is little dynamic range to accurately estimate the relaxation constants of more slowly decaying signal (groups 1, 2, and 3 in Fig. 6a). Given such a small range of TEs, small oscillations (Fig. 6b) can have profound effects in estimation of long T2* values. Other modulators of T2*, such as cardiac motion, flow artifacts, and blood oxygenation level dependent (BOLD) effect also become relatively more important at longer values. BOLD effects within the myocardial capillary network (29) can contribute a 17% increase in T2* in normal subjects (30). However BOLD effect changes in R2* (~7 Hz) are quite small given the dynamic range of R2* observed in iron overloaded thalassemia patients (up to 500 Hz) and are hence unlikely to contribute in the clinically relevant range (T2* < 20 msec).
Deviations from exponential relaxation were small (1% of S0) and are in accordance with the expected “beating” phenomenon between fat and water protons. Myocardium is relatively lipid poor so the observed magnitude is not surprising. Although the fat signal modulation contributes only 1% of the average total signal, it introduces echo time dependence into the R2* measurement. The first out-of-phase and in-phase points are at TE = 2.3 msec and 4.6 msec, respectively. To minimize this effect, one could collect the first echo near a relative null (1.15 msec) and sample at 2.3-msec intervals; single breath-hold echo planar readout techniques might be particularly suitable for this sampling scheme. However, given the relative small contribution of this effect, a more practical approach is to space the echoes such that these phase effects are well-balanced and to always use every echo in the fitting process.
Figure 5 demonstrates the critical importance of using a short minimum TE to recall rapidly dephasing spins. Ideally, one should have a minimum TE of 1 msec to ensure accurate characterization of all clinically relevant cardiac R2* values (we have observed one patient with a measured value of 500 Hz). In practice, a minimum TE of 2 msec (achievable on most machines) will provide unbiased estimates down to a T2* of 3 msec, covering over 95% of the patient range. Original T2* studies in thalassemia patients (10) employed minimum TE of 5.6 msec and exhibited a conspicuous absence of patients having T2* < 5 msec (R2* > 200 Hz). Although one might question the clinical relevance of this finding, more recent work (using minimum TEs of 2 msec) has demonstrated that patients with a T2* of 4 msec have nearly 50% greater relative risk of cardiac dysfunction than patients having T2* of 5 msec (20).
The circumferential asymmetry that we observed in non iron overloaded patients mirrors prior studies in normal volunteers (12). Higher R2* values located in these studies were attributed to susceptibility artifacts arising from deoxygenated hemoglobin in the great cardiac vein, posterior vein of LV, and middle cardiac vein.
However, susceptibility of epicardial venous blood cannot explain the large circumferential variation observed in the iron overloaded patients; the absolute circumferential fluctuations in R2* are three-fold larger than for the non iron-loaded patients. These fluctuations could represent true variation in cardiac iron density or artifactual variation from cardiovisceral geometry. Although some patients have demonstrated large circumferential variation in iron deposition, this has not been systematically described (13). We believe that these circumferential R2* variations should be considered artifactual until proven otherwise, and recommend limiting measurements to the interventricular septum. In contrast, endocardial-to-epicardial R2* gradient (Fig. 7b) is consistent with post-mortem studies (13–15). This pattern was not observed in any of the sickle cell disease patients. Thalassemia patients with T2* > 20 msec demonstrated a slight but significant endocardial to epicardial R2* gradient, raising the question whether this may be an early indicator of cardiac iron deposition.
The cardiac T2*-weighted images in this study were acquired using an older technique requiring one breath-hold for each TE. Despite image registration, there is likely to be residual image misalignment that would blur pixel-wise fitting and spatial variability of R2*. Newer echoplanar techniques allow acquisition of one or more slices in a single breath-hold. These techniques could potentially improve the sensitivity and robustness of pixel-wise techniques in patients. Nonetheless, given the agreement between our modeled data, which was perfectly registered, and our observed results, we do not believe that the newer techniques will fundamentally change the conclusions of this paper.
In conclusion, myocardial R2* mapping demonstrates that nonspecific artifacts are large, even in iron overloaded patients, but can be minimized by restricting analysis to the interventricular septum. Thalassemia patients exhibit a R2* gradient from endocardium to epicardium, not found in patients with sickle cell disease. ROI analysis techniques are as accurate as pixel-wise mapping methods. Acquired signal exhibits exponential relaxation; however, offset-correction or biexponential curve fitting is essential for accurate measurements. The reciprocal of minimum TE determines the range of R2* that can be quantitated. Contributions from fat-water oscillations are small. In short, myocardial R2* measurements can be made accurately with simple attention to technical detail. Although absolute calibration between cardiac R2* and cardiac iron is currently lacking in humans, the present work supports the continued use of cardiac R2* as a biomarker for iron cardiomyopathy.
The authors wish to acknowledge Susan Carson and Debbie Harris for recruiting patients.
Contract grant sponsor: National Institutes of Health; Contract grant numbers: RR00043-43; 1 R01 HL75592-01A1; Contract grant sponsor: Novartis Pharma AG.