Myocardial T2* Measurements in Iron Overloaded Thalassemia

Abstract

Reproducible and accurate myocardial T2* measurement is required for the quantification of iron in the tissue of heart in transfused thalassemia. The aim of this study was to determine the best method to measure the myocardial T2* from multi-gradient echo data acquired both with and without black blood preparation. Sixteen thalassemia patients from six centres were scanned twice locally, within 1 week, using an optimised bright blood T2* sequence and then subsequently scanned at the standardization centre in London within 4 weeks, using a T2* sequence both with and without black blood preparation. Different curve fitting models (Mono-exponential, Truncation, and Offset) were applied to the data and the results were compared by means of reproducibility. The T2* measurements using both the bright and black blood techniques were also compared. The black blood data was well fitted by the mono-exponential model, which suggests a more accurate measure of T2* by removing the main source of errors in the bright blood data. For bright blood data, the offset model appeared to underestimate T2* values substantially and was less reproducible; the truncation model gave rise to more reproducible T2* measurements, which were also closer to the values obtained from the black blood data.

Introduction

Myocardial iron measurement is important for assessing the risk of cardiac complications (1), and tailoring appropriate iron-chelating treatment in transfused thalassemia major (TM) (2,3). T2* values derived from Magnetic Resonance Imaging (MRI) are inversely related to tissue iron levels. In particular, myocardial T2* values less than 20 ms indicate cardiac iron overload, and this is considered severe when T2* is less than 10ms (2,4, 5). Most recorded cases of heart failure in thalassemia to date have occurred in patients with very low T2* values (2). Therefore, the accuracy and reproducibility of T2* measurements are of importance for the management of patients with iron overload in the heart.

Over recent years a bright blood T2* sequence has been developed and demonstrated to have relatively good reproducibility (6,7). However, this technique can be affected by noise, motion and blood artifacts which can complicate the analysis. This is a particular problem for heavily iron loaded hearts where the very short T2* (<10ms) results in a rapid decay in myocardial signal intensity (SI) leading to a plateau for the low signal to noise ratio (SNR) in the later echo time images (Figure 1). This signal plateau complicates the approach to curve fitting for evaluating the T2* and different approaches have been adopted to understand and model this. In our previous work, we have used a truncation model where the late “plateau” points are subjectively discarded and then the remaining signal is fitted with a mono exponential equation (6,7,8). An alternative more objective approach is to use the offset model where an exponential equation plus a constant offset has been used to tackle the problem (9). With this approach, the unknown constant is derived by use of an iterative process. For a typical example shown in figure 1, the truncation model has been applied to fit only the first four data points producing a measured T2* of 6.9ms. In comparison, the offset model fitting all the data points, gives a lower T2* measurement of 4.4ms. With this discrepancy of more than 50%, the open question is: which measurement is closer to the most expected T2* value? Additionally, bearing in mind the great importance of reproducibility in clinical measurement, the second question is which measurement is more reproducible?

Figure 1.

An example showing patient data fitted using both the offset and truncation models. The offset model fits all the data points well (R²=0.996, T2*=4.4ms) while the truncation model fitted the first 4 points only (R²=0.999, T2*=6.9ms). The mean noise as measured from a background region is also shown.

In a concurrent study, we have shown that T2* measurement can be optimised for an ex vivo heart by first minimising or correcting for the noise and then using a simple mono-exponential fit (10). For in vivo scanning however, although random noise can be reduced by optimising the protocol (Figure 1, line of mean noise), the measurements are further complicated by cardiac motion and blood artifacts. The recent development of a T2* sequence with black blood preparation provides more clearly defined endocardial borders and reduces artifacts from blood which affect the myocardial signal measurement and this has been shown to have improved reproducibility (11). We have also shown in the concurrent study (10) that the truncation model can produce more accurate and reproducible measurements for the ex vivo scans.

Through our previous work, we have formed the impression that the offset signal is largely due to blood signal and motion artifacts, and therefore the purpose of this study was to investigate whether with the black blood preparation and an optimised sequence to minimise noise, the T2* could be well modelled using a simple mono-exponential model fit. On the premise that blood signals and artifacts represent the major sources of error we would suggest that the optimised black blood approach with simple mono-exponential modelling would lead to a more accurate measure of T2*. Another purpose of this study was to investigate the most appropriate model to use for analysing and measuring T2* from the bright blood sequence images.

Material and Methods

Study Population and MR Protocols

Sites using 1.5T scanners in Cairo, Izmir, Kuala Lumpur (all Siemens Symphony Quantum), Hong Kong (Siemens Sonata), Beirut (GE Signa) and Athens (GE CVI) were involved in this study. Additionally the scanner at our site (the standardization site) in London was a Siemens Sonata. A single-breath-hold bright blood T2* sequence (6) was installed and validated using a T2* phantom on each local scanner. Both the bright and the black blood T2* sequences (6, 11) were used at our site in London. The pulse sequence details and general methods of the T2* calculation have been previously described (6,11). In brief, all sites used a cardiac phased array coil with ECG gating and a single mid-ventricular short axis slice was imaged at eight echo times. Owing to inherent differences in scanner performance and software at the various sites, the sequence parameters (particularly TE and TR) were not identical in all sites, but were kept similar to those described in the original publication (6). These variations would not be expected to significantly affect the T2* measurement, a fact that has been confirmed by a number of previous studies (7,8, 12, 13,14).

Sixteen TM patients (mean age 25±6 years) from the above mentioned sites with a T2*<10ms were studied. For interstudy reproducibility, the patients were scanned twice at their local site using the bright blood T2* sequence (6) with the repeat scan being within one week. All the patients were subsequently scanned at the standardization site in London, again using the bright blood sequence, within four weeks of their original scan, thus enabling a measure of interscanner reproducibility. Additionally at the London site only, these patients were also scanned using the black blood T2* technique (11). The TM patients had been regularly transfused since early childhood and were all receiving regular iron chelation at the time of the study. The study was approved by all the required ethics committees and the subjects gave their informed consent.

T2* Measurement

For T2* analysis, a homogeneous full-thickness region of interest (ROI) was chosen in the septum (6,11). The mean signal intensity of the ROI was measured for each of the images, and the data were plotted against the TE to form a decay curve. All curve fitting models have been described in (10). We were interested in applying the different models to see how they affect the in vivo T2* measurement in order to identify the model which produces optimal results.

All the bright blood data were analysed using the truncation model, where the first few points were truncated and then modelled using:

$$\mathit{SI}=P_0\cdot e^{-\mathit{TE}∕T{2}^{\ast }}$$ [1]

where P₀ represents a constant of magnetization, TE represents the echo time and SI represents the image signal intensity. The black blood data were also modelled using this equation but without truncation (mono-exponential model).

The offset model with a constant C:

$$\mathit{SI}=P_0\cdot e^{-\mathit{TE}∕T{2}^{\ast }}+C$$ [2]

was applied to both the bright and black blood data. The bi-exponential model in the form of:

$$\mathit{SI}=P_0\cdot e^{-\mathit{TE}∕T{2}_a^{\ast }}+P_1\cdot e^{-\mathit{TE}∕T{2}_b^{\ast }}$$ [3]

was investigated, but we found its first component was the same as that of the offset model and the second component (with a very long T2* value) simply represented a constant value.

Although background noise was trivial with regards to the analysis in this study, the bi-exponential model in the form of:

$$\mathit{SI}=P_0\cdot e^{-\mathit{TE}∕T{2}_a^{\ast }}+P_1\cdot e^{-\mathit{TE}∕T{2}_b^{\ast }}+C$$ [4]

was also investigated, as before we found its first component was nearly the same as that of the offset model and the sum of the other two components simply converged to a constant value. This resulted in the fact that the bi-exponential model reduced itself to be the same as the offset model in this study. For this reason, the results of the bi-exponential model are not presented.

The T2* analysis of the blinded bright blood data was carried out by two experienced observers independently using the truncation model. These same ROI data were subsequently analyzed by the first author using the offset model. The black blood data were analysed again by the first author using the mono-exponential and the offset models using the identical ROI’s. The Levenberg-Marquardt algorithm (15,16) of nonlinear estimation was employed throughout this study, where the results from a linear mono-exponential model without truncation were set as initial values to improve the fitting process (10). All the analysis was carried out on a PC using Matlab, Excel and Thalassemia-Tools software (a plug-in of CMRtools, Cardiovascular Imaging Solutions, London).

Statistics

All measured data were expressed with 95% confidence intervals. R² was used in this study to describe how well the exponential model fitted the empirical data. Bland—Altman analysis was used to plot the differences versus the average values between two data sets. Reproducibility (both inter-scanner and inter-study) is expressed as the coefficient of variation (CoV, standard deviation of the differences between the two separate measurements, divided by their mean). Wilcoxon’s signed-rank test was applied to paired and unpaired data from two samples. Group data from three or more samples were compared using Friedman’s test. A p value of <0.05 was considered statistically significant.

Results

Mono-Exponential Curve fitting of Black Blood Data

Figure 2 demonstrates representative bright and black blood images acquired at different echo times from a thalassemia patient. Images in the short-axis plane show good blood suppression and high contrast in the black blood images, which are consistent with our previous findings (11). Figure 3 shows one typical example of the mono-exponential model fitted to the bright and black blood data respectively. It demonstrates that the mono-exponential model did not fit the bright blood data well (Figure 3, left, R²=0.970) while the black blood data was much better fitted (Figure 3, right, R²=0.998). Taking all the patient data (n=16), the mean R² values were 0.966±0.085 and 0.997±0.005 respectively, where Wilcoxon’s signed-rank test indicated a significant difference (p<0.001). Although the mono-exponential fitting to the black blood data can be improved by further truncation, e.g. the last point can be discarded for a better fit in Figure 3 (right), this causes little change to T2* measurement (∼1%).

Figure 2.

Example short-axis midventricular images of the bright (top) and black (bottom) blood data obtained from the same TM patient, where different echo (2.60 ms, 4.62 ms, and 6.64 ms) images were shown respectively in left, middle and right columns.

Figure 3.

Example decay curves of the bright (left) and black (right) blood data obtained from the same TM patient.

T2* Measurements between Bright and Black Blood Data

Figures 4 and 5 show Bland-Altman plots and scatter plots of the myocardial T2* values obtained by using both the truncation and the offset models for both the bright and the black blood data. The mean difference for the truncation model was close to 0 (0.08ms) while it was obviously biased (-1.14ms ) for the offset model. The confidence intervals were correspondingly 1.54ms and 3.13ms for the two models, indicating increased variability of the latter model. The CoV was 5.3% for the truncation model but increased to 13.6% for the offset model. The bias using the offset model is also clearly demonstrated in the scatter plots (Figure 5, right), where the T2* from the bright blood was substantially lower (5.2±2.6 ms vs 6.4±3.4 ms, p=0.02).

Figure 4.

Bland-Altman plots for T2* comparison between bright and black blood data with the 95% confidence intervals shown as a dotted line. Left: Truncation model; Right: Offset model

Figure 5.

Scatter plots for T2* comparison between bright and black blood data. Left: Truncation model; Right: Offset model

Interstudy Reproducibility

Figures 6 and 7 shows Bland-Altman plots and scatter plots of the myocardial T2* values obtained by using both the truncation and the offset models for local interstudy scans using the bright blood data. The mean differences for both the truncation and the offset models were close to 0 (0.04ms and -0.01ms respectively). The confidence intervals, however, were correspondingly 1.38ms and 2.77ms for the two models, indicating increased variability of the latter model. The CoV was 4.7% for the truncation model but increased to 13.6% for the offset model.

Figure 6.

Bland-Altman plots for interstudy reproducibility with the 95% confidence intervals shown as a dotted line. Left: Truncation model; Right: Offset model

Figure 7.

Scatter plots for interstudy reproducibility. Left: Truncation model; Right: Offset model

Interscanner Reproducibility

Figures 8 and 9 demonstrates Bland-Altman plots and scatter plots of the myocardial T2* values obtained by using both the truncation and the offset models for interscanner repeated scans between local sites and London using the bright blood data. The mean difference for the truncation model was 0.04ms in comparison to -0.11ms for the offset model. The confidence intervals were correspondingly 1.55ms and 3.90ms for the two models, indicating increased variability of the late model. The CoV was 5.3% for the truncation model but increased to 19.0% for the offset model. We noticed (Figure 8, right) that there was an outlier using the offset model. The data was retrospectively checked and found that this was due to clear motion and blood artifacts in a local scan. Interestingly in this case, the same data when analysed using the truncation model produced no outlier (Figure 8, left). The results for the offset model are slightly improved if the outlier is removed, with the mean difference decreased to 0.09ms, the confidence interval to 2.4ms and CoV to 11.6% respectively.

Figure 8.

Bland-Altman plots for interscanner reproducibility with the 95% confidence intervals shown as a dotted line. Left: Truncation model; Right: Offset model

Figure 9.

Scatter plots for interscanner reproducibility. Left: Truncation model; Right: Offset model

T2* comparison between the truncation and the offset models

Friedman’s test showed no significant difference (p=0.962) between the bright blood T2* measurements (2 local and one London scan) using the truncation model and the black blood T2* measurement. On the other hand a significant difference (p<0.001, n=16) was found between the bright and the black blood T2* measurements using the offset model. For a comparison, all T2* measurements in this study (both the bright and the black blood data, n=48) were calculated using the truncation and the offset models respectively. Figure 10 shows the Bland-Altman plot and scatter plot with line of identity. It was evident that the offset model produced shorter T2* measurements (mean difference of 1.90±1.70ms, p<0.001, n=48) than using the truncation model. From Figure 10, it can be also observed that for black blood data (black dots) the discrepancies were significantly reduced (p<0.001) (n=16, mean difference of 0.99±1.4ms) compared to those from the bright blood data (white circles) (n=48, 2.23±1.3ms).

Figure 10.

Bland-Altman (left) and scatter plots (right) plots for T2* comparison between the truncation and the offset models with the 95% confidence intervals shown as a dotted line. The white circles represented T2* comparison obtained using the bright blood technique while the larger black dots for those obtained using the black blood one.

Discussion

Theoretically, T2* is inherently related to T2 due to spin-spin relaxation, but is also affected by local field irregularities that are increased by iron deposition. T2* may also potentially be affected by magnet and sequence parameters such as shimming and voxel sizes as well as small scale inhomogeneities due to susceptibility changes at tissue interfaces. Myocardial T2* is more challenging because it is further complicated by cardiac motion and blood artifacts. For thalassemia patients, heavy iron overload can induce further problems. The contribution of these influences in clinical practice is not well characterized.

Myocardial T2* measurement to assess iron overload has been developed in our unit since 1999. This T2* technique was originally developed (4,6) with the aim of minimizing imaging artifacts, e.g. flow compensation was used. With continuing advances of MRI technology, main field non-uniformity is mainly patient-induced and the T2* is subject to localized susceptibility artifacts (17,18,19). Usefully, the susceptibility artifacts can be minimized by confining the measured ROI to the septum and as such, this technique has demonstrated good reproducibility (6,7,8) and recently has been further optimized in terms of trigger delay, echo-spacing, acquisition window and other parameters (20). A more recent advance is the inclusion of black blood preparation, which provides suppressed blood signal, clearly defined borders and improved interobserver reproducibility (11). In the current study, the black blood preparation has allowed the removal of what appears to be the major component of the myocardial signal offset. When using this sequence we have found that the signals can be well fitted by a simple mono-exponential model and we believe that from this we can measure the T2* most accurately. This finding also fits with the result from our concurrent ex vivo study (10) where noise could be minimized or corrected to optimise accuracy of T2* measurement by use of a simple mono-exponential model. On the basis of these findings, the black blood protocol was taken as the “standard” for the remainder of this study.

The exact manner in which the blood signal interferes with the myocardial signal in the series of T2* images is complex (11). Because of the relatively long T2* of left ventricular blood, the offset or the bi-exponential models have quite reasonably been proposed to address the issue of blood signal contamination of the myocardium (9). In this study, however, we have shown that these models were not as reproducible for the bright blood data, a fact that we believe can be explained by the noise sensitivity of the offset model and the particular behaviour of the blood and artifact signals (see below).

This study also showed that the offset model gave rise to a larger discrepancy than the truncation model in T2* measurements between the bright and black blood data. This suggests that the offset is at least in part due to the blood signal but that it is affecting the signal in a manner that may not have been expected. Interestingly, again in comparison to the ex vivo data, the effect of the blood signal on the decay curve appears to be similar to the effect of Rician noise, i.e. it is not a constant contribution to all echo time points. The blood signal appears to result in an underestimation of the T2* by the offset model, a finding that mirrors the result of the ex vivo study, where we showed that the offset model underestimated T2* and that the scale of the underestimation reduced with less noise. The fact that the blood signal and artifact appears to affect the signal decay similarly to noise could explain the greater variability in the offset model measurements. The quality of patient scans can be affected by many factors and different noise and artifacts levels can often be found even for repeat scans with the same scanner. Although the truncation model is subjective, it proved much more reproducible and produced a close value to what we believe to be the more accurate black blood T2* measurement. The subjective nature of the truncation model where the observer analysing the data makes a decision on which points to discard has the potential to introduce a measurement error. In our experience, however, the decision is reasonably clear and the measurement difference between inclusion and exclusion of an uncertain point is also not large. In this work two trained observers analysed the data as they do routinely for clinical studies. This approach works well; however, we believe it would be entirely feasible to devise an automated method of deciding which points to exclude based on the curve fitting parameters or a noise measurement.

The analysis was restricted to the mid-ventricular septum to avoid susceptibility artifacts. Interestingly, a recent study (21) showed that the mid-ventricular septum T2* correlated well with the global T2* value (r = 0.95, P < 0.0001). The same group (19) also developed a correction map to compensate for the variations in segmental T2* analysis. However, it demonstrated that this correction map failed to compensate the segmental T2* variations in patients with mild iron overload. Furthermore, the authors concluded in the same study that the effect of the correction map was completely negligible in patients with severe iron overload. These findings, therefore, suggest that the correction map developed from normal subjects may not be directly transferred to patients with iron overload. Overall, although results from these and other studies suggest that the septum T2* can be a good indicator of the whole myocardial iron, such data is limited to date and a more robust methodology is needed to allow the evaluation of the whole myocardium in a reproducible way.

The results of a complementary ex vivo study are referred to throughout this manuscript purely to indicate the similarity between the effects of Rician noise (ex vivo) and the blood signal and artifacts (in vivo). A previous study reported by Ghugre et al. (22) has suggested a difference between the in vivo and ex vivo measurement of T2*. The reasons for the difference may or may not be clear, however, the conclusions in this study relating to the mathematical model and quality of curve-fitting are probably not affected. Another limitation of the current study is that only ROI based analysis was investigated. As discussed in our ex vivo study (10), although a comparison between the ROI based and pixel by pixel analyses would be of particular interest, this comparison is beyond the scope of the current study. In addition, the pixel by pixel analysis is actually similar to a smaller scale ROI based analysis and further study is required to fully assess the relationship between these two methods.

The current study has only included patients with a short myocardial T2* (<10ms). There would in fact be no barrier to expanding the study to the whole range of T2* values, however, we intentionally studied patients with short T2*’s because we were particularly interested in the signal contribution from blood and blood artifacts at the longer echo times after the myocardial signal had decayed away. Only in this scenario, the performances of different models can be effectively evaluated. It is known that data from the same site are more correlated than data from different sites. However, we do not view this as a drawback for the current study. Our results show that the truncation model can produce good reproducibility even for uncorrelated data from different sites and different patient populations. This finding is not expected to change if the patients are from a single site. From the point view of technical development, the model is usually expected to be tested on data with less coherence.

It should be pointed out that although we believe that the described methods lead to more accurate and reproducible measurement of T2*, this may not mean a more accurate measure of tissue iron. Although a correlation between iron loading and T2* has been demonstrated in an animal study (23) and a post mortem study (22), we believe that further studies are required to fully understand this relationship.

In conclusion, we believe the T2* measurement can be both reproducible and accurate. For black blood data, the mono-exponential model generated a good fit to the data and produced reproducible T2* values. For bright blood data, however, the mono-exponential model gave rise to a poor curve fitting and generally overestimated the T2* measurement if including all the data points. Although the offset model fitted the curve well in this situation, it appeared to underestimate T2* values and lower reproducibility, compared to the black-blood “standard”. The truncation model, although subjective, gave more reproducible T2* measurements, which are also close to those obtained from the black blood data. Overall, we believe that the black blood technique optimises T2* measurement by avoiding the blood and motion artifacts. Nevertheless, the black blood technique is not as widely available in which case we believe the truncation model should be used with the bright blood technique in order to obtain more reproducible and more accurate T2* measurements.