Quantification of [¹¹C]-meta-hydroxyephedrine uptake in human myocardium

Recently, non-invasive imaging of sympathetic innervation of the myocardium using PET [1]-[5] or SPECT [6]-[8] has gained interest based on its ability to predict life-threatening ventricular arrhythmias [9]-[11] and to assess whether implantable cardioverter defibrillator (ICD) therapy is appropriate [12],[13]. It has been shown that sympathetic nerve terminals in the myocardium are more sensitive to ischemic damage than cardiomyocytes [9],[14]-[21]. In addition, the area of innervation defects often exceeds the area of non-viable scar tissue [9],[11],[16],[17]. More importantly, it has been shown that areas of viable myocardium, but with impaired innervation, are related to inducible ventricular tachycardias originating from these areas [22]. Therefore, non-invasive imaging of sympathetic innervation may play a key role in risk stratification and treatment planning in ischemic cardiomyopathy.

Sympathetic innervation can be measured using the PET tracers [¹⁸ F]-6-fluorodopamine [2], [¹¹C] epinephrine [1],[3] or [¹¹C]-meta-hydroxyephedrine ([¹¹C]HED) [4],[5], with [¹¹C]HED being used most often. Analysis of [¹¹C]HED and [¹¹C] epinephrine data often has been performed using the retention index (RI) [5], a semi-quantitative parameter that can be derived relatively easy. In general, it is obtained by normalizing late activity concentrations to the integral of the blood time-activity curve. However, RI can be sensitive to cardiac motion, partial volume effects and contribution of intravascular activity and spill-over and therefore may yield inaccurate results. More quantitative results can be obtained by using compartment models, in which correction factors for blood volume and spill-over of activity from the blood can be applied. Interestingly, RI has never been validated using a direct comparison with quantitative results obtained from a full tracer kinetic analysis. Furthermore, blood time-activity curves both with [3],[22] and without [23]-[25] corrections for radioactive metabolites of [¹¹C]HED have been used for RI, making it difficult to compare studies. On the other hand, another semi-quantitative measure, standardized uptake value (SUV), could provide a further simplification relative to RI, as it does not require measurement of blood time-activity curves, nor dynamic scanning.

Only few studies [26],[27] have performed quantitative analysis of [¹¹C]HED data using a single-tissue compartment model, but no systematic assessment of the best tracer kinetic model was performed. Therefore, the aims of this study were to determine the optimal tracer kinetic model for analysing [¹¹C]HED data and to assess the validity of several simplifications, such as RI, SUV and a population-averaged metabolite correction.

In the present study, the optimal tracer kinetic model for kinetic analysis of [¹¹C]HED scans was identified and several commonly used simplifications were evaluated. First, Akaike and Schwarz criteria clearly indicated that, for both clinical and simulated (Table 2) data, the reversible two-tissue model was the preferred model for describing [¹¹C]HED kinetics. Similar results were obtained when using the F-test, which is insensitive to any potential bias introduced in AIC and SC due to the scaling factors as presented in Equations 3 and 4. This confirms that, based on goodness of fit, the reversible two-tissue model was the optimal model. Simulations, on the other hand, also indicated that with increasing noise levels (starting at 5% noise, corresponding with noise at the level of the coronary territory), precision of all parameter estimates using this model rapidly decreased (Table 3) and that, for higher noise levels, model preference shifted towards models with fewer parameters (Table 2). In addition, for clinical data, the number of reliable estimates was lower for the 2T4k parameters V_T,2t, V_S and BP_ND (77.1, 77.1 and 23.3%, respectively) than for K _i derived using the 2T3k model (86.0%) and V_T,1t obtained using the 1T2k model (94.6%). The large number of estimates with high uncertainty indicates that the use of the reversible two-tissue model may not be feasible in routine clinical practice, as it would mean discarding over 20% of all data. In addition, most estimates with high uncertainty in 2T4k parameters were in regions with reduced [¹¹C]HED uptake, indicating that the 2T4k model is especially vulnerable in diseased myocardium, which significantly limits the applicability of this model.

For the reversible two-tissue model, the total volume of distribution can be separated in specific and non-specific volumes of distribution. In this study, the total volume of distribution approximated the specific volume of distribution and enabled the use of the simpler single-tissue model, as the contribution of the non-specific compartment was very small, if not negligible. As shown in Figure 2, V_T,1t yields similar results to V_T,2t, indicating that the single-tissue model yields similar information to the reversible two-tissue model. In addition, precision was much higher for V_T,1t than for V_T,2t (Tables 3 and 4) and noise-induced biases were lower for V_T,1t. Since the single-tissue model is simpler, more stable and less sensitive to noise compared with the reversible two-tissue model, it is preferred for further use, especially for smaller regions or regions with relatively low uptake.

In this study, arterial activity was assumed to be spill-over from the arterial blood pool, rather than activity originating from arterial blood within the tissue. Both assumptions are physiologically not fully correct and ideally, a combination should be used. However, differentiating between contributions from spill-over and myocardial blood volume is impossible, as kinetics of both contributions are identical. Previously, it has been shown that, in a comparison of MBF as measured with [¹⁵O] water against labelled microspheres, assuming arterial activity to be spill-over yields more accurate results [37], at least for [¹⁵O] water. A simulation study (see Figure 6 in Appendix) showed that the same is true for HED, although a bias was found in both implementations. Bias was smallest when arterial activity was assumed to be due to spill-over. In addition, arterial blood volume fraction within myocardial tissue is relatively constant in the myocardium [41], and therefore, it is expected that bias in the spill-over implementation, which is dependent on the magnitude of this blood volume fraction, is more consistent than bias in the tissue blood volume implementation, which depends on the recovery of the scanner and segment size.

The use of HED as an innervation tracer has recently been debated [42], where it was suggested that HED is a flow-limited rather than innervation-limited tracer. This was suggested to lead to the impossibility to fit HED data using standard kinetic modelling techniques and make HED uptake insensitive to early changes in innervation. However, when performing tracer kinetic analyses, parameters representing innervation (i.e. K _i , V_T, V_S and BP_ND) can be obtained together with parameters representing flow and extraction (i.e. K₁). This enables separating flow effects from innervation effects by studying both K₁ and, when the 1T2k model is used, V_T. Furthermore, in this study we show that, utilizing plasma input functions fully corrected for radiolabelled metabolites and plasma-to-blood concentration ratios, HED could reliably be fitted using a 2T4k model or, with slightly poorer fits, using a 1T2k model, which is in contrast to the suggestion in [42].

The non-linear relation between RI and V_T, however, does indicate that for RI it might be the case that early changes are not readily detected. Indeed, when a sensitivity analysis is performed for changes in k₃, which represents HED uptake through the norepinephrine transporter, using K₁, k₂ and k₄ of 0.6 mL • cm^-3 • min^-1, 0.15 min^-1 and 0.02 min^-1, respectively, and gradually decreasing k₃, RI showed a lower and non-linear sensitivity to changes in innervation (see Additional file 1). On the other hand, V_T,1t showed good sensitivity to changes in innervation, where decreases in V_T,1t were similar to decreases in k₃, confirming that V_T,1t is a more reliable measure of innervation than RI. Nevertheless, sensitivity of RI to changes in k₃ may still be sufficient for routine clinical use as large reductions in k₃ are readily detected by RI. It is also important to note that another recent study [43] showed that HED was able to detect early changes in innervation, confirming that HED is likely to represent innervation even when using RI.

Potentially, a population-averaged metabolite correction could, in combination with an image-derived input function, make arterial cannulation obsolete. There was, however, substantial variation in individual metabolism (Figure 3), resulting in large errors of up to 30% in fitted parameters when using a population-averaged metabolite correction (Figure 4) for approximately 20% of the patients. Clearly, for absolute quantification, a population-averaged metabolite correction is not feasible, at least not in the case of [¹¹C]HED, as a method that produces inaccurate results in as much as 20% of the cases should be discarded. However, when relative differences within a single patient are of interest, e.g. when assessing defect areas (percentage of LV with V_T below a reference value), population-averaged metabolite corrections can be considered. Using venous samples for metabolite correction might be an alternative for both relative analysis and absolute quantification, but this requires further validation.

In addition, the use of an image-derived input function would be preferred over the use of online blood sampling. However, image-derived input functions require validation for each individual radiotracer and preliminary results for [¹¹C]HED [44] have shown that late activity concentrations were overestimated in the image-derived input functions, which is in line with observations in other studies that made use of a combination of image-derived input function and blood sampler data [26],[27]. Given the fact that an image-derived input function can be used for other tracers such as [¹⁵O] water [45] and [¹⁸ F] FDG [33], the particular biodistribution of HED with its high contrast between the myocardium and the blood at later time points, together with high liver uptake, will require improved reconstruction methods and/or scatter correction algorithms before an image-derived input function can be used for [¹¹C]HED.

A reasonable correlation was found between V_T,1t and either RI_P, RI_WB or SUV (r² = 0.77, 0.76 and 0.62, respectively). These correlations suggest that RI_P, RI_WB and SUV may be used to distinguish innervated from denervated regions of the myocardium. However, it should be noted that relationships were non-linear (Figure 5). Consequently, these simplified measures will be less sensitive in detecting early changes in regions with high innervation, e.g. in early disease, or for monitoring response during treatment. In these cases, the use of quantitative measures is recommended. On the other hand, for routine clinical (diagnostic) use, RI might be sufficient, as large reductions in HED uptake are reflected in large reductions in RI. In addition, simplified methods may lead to biased estimates of defect sizes, as these usually are derived using cut-off values, which are based on a percentage of uptake in a reference region within the heart. Due to the non-linear relationships, cut-off values for RI_P, RI_WB and SUV will be different from those for V_T,1t and, consequently, defect areas may differ significantly.

In the present study, RI was based on uptake between 50 and 60 min post injection, whilst 30 to 40 min is more commonly used. When RI was calculated using uptake from 30 to 40 min post injection, similar results were obtained (r² = 0.72 and r² = 0.73 for RI_P and RI_WB, respectively). For RI_P, this correlation was slightly, but significantly, lower than that for the 50 to 60 min uptake period (p = 0.043). For RI_WB, this difference was not statistically significant (p = 0.071). Although correlation with V_T,1t was still reasonable, based on these findings, it is recommended to use later times for measuring RI due to the slow kinetics of HED.

Finally, it is important to note that the aforementioned results and interpretations are made from a modelling point of view, i.e. which parameter describes the measured HED activity best. This might be different from which parameter actually describes myocardial presynaptic activity best. Ideally, a blocking study or animal study should be performed to compare the obtained parameters with a gold standard measurement. In addition, the performance of the quantitative parameters used in the present study needs to be assessed in clinical studies, as the added clinical value of V_T over simplified measures such as RI_P still needs to be defined. Whilst simulations (see Additional file 1) show that quantitative measures are more sensitive to changes in k₃, the rate constant reflecting norepinephrine transport, RI also showed significant correlation with k₃, suggesting that RI may be sufficiently sensitive for measuring (changes in) innervation. Further clinical validation studies using RI_P, RI_WB, SUV and V_T are warranted.

Although a reversible two-tissue compartment model best describes [¹¹C]HED kinetics, a single-tissue compartment model is preferred for routine clinical studies, as it is more robust at clinically relevant noise levels and, at the same time, provides V_T results that are highly correlated with those obtained with the two-tissue compartment model. Simplified measures, such as RI and SUV, showed good correlation with fully quantitative results and may be used to detect regions of denervation, although the non-linear relationship with V_T might limit their application in, for instance, monitoring response to treatment.

HJH performed the analyses and drafted the manuscript. SdH and MTR included and scanned the patients. PK and CPA participated in the design of the study. RCS was responsible for blood sampling and metabolite analysis. ADW was responsible for tracer synthesis. AAL, MCH and ML helped with analyses, discussions and drafting of the manuscript. All authors read and approved the final manuscript.

Contribution of arterial activity measured within a myocardial region is typically the combination of two different effects: a physical blood volume fraction within the myocardium and spill-over of activity from the left ventricular cavity into the myocardium. This can be represented as follows:

in which C_T(t) represents the HED concentrations in the myocardial tissue itself whilst C_PET(t), C_A(t) and C_RV(t) represent the measured regional, arterial and right ventricular blood activity concentrations, respectively. V_A represents the actual arterial blood volume fraction in the tissue, and V_LV and V_RV represent the spill-over fractions from the left and the right ventricular cavity, respectively. During compartment modelling, it is not possible to distinguish V_A from V_LV, and therefore, the estimated arterial activity measured within a myocardial region can be interpreted in two different ways:

Interpretation SO (spill-over, Equation A2) assumes all measured arterial activity in the tissue regions to be due to spill-over from the left ventricular cavity, and a correction factor V_A • C_A(t) is used to correct for this spill-over fraction. In contrast, interpretation BV (blood volume, Equation A3) assumes all arterial activity in the tissue regions to be due to intravascular activity originating from the arterial blood volume within the myocardium itself. Since measured K₁ represents K₁ in the myocardial tissue only, a correction factor for tissue fraction has to be applied in model BV, in addition to the correction factor V_A • C_A(t). This tissue fraction is estimated as 1 - V_A. In reality, measured activity originates from both spill-over from the left ventricular cavity and myocardial arterial blood, but it is impossible to distinguish between both contributions and, in practice, only a single correction can be applied. It is important to note that the actual fits obtained using Equations A2 and A3 are identical and only the interpretation of K₁ and consequently V_T is different.

To define the optimal model implementation, a simulation study was performed for both model interpretations. Three tissue curves (i.e. healthy tissue, scar tissue and viable but denervated tissue) were defined, using the same parameters as for the noise simulations. V_RV was 0.1 for all simulated tissue curves. For each of the three tissue curves, arterial activity was simulated to be due to both spill-over from the left ventricular cavity (V_LV) and arterial blood volume in the myocardium (V_A) using Equation A1.

The total contribution of arterial blood activity was set to 0.3, based on clinical data (average V_A for all segments 0.30 ± 0.10, 0.28 ± 0.10 and 0.25 ± 0.11 for 1T2k, 2T3k and 2T4k models, respectively), and V_RV was set to 0.1, again based on clinical data (average V_RV for all segments 0.07 ± 0.08, 0.06 ± 0.07 and 0.06 ± 0.07 for 1T2k, 2T3k and 2T4k models, respectively, and for septal segments alone 0.14 ± 0.09, 0.13 ± 0.09 and 0.13 ± 0.09 for 1T2k, 2T3k and 2T4k models, respectively). For each tissue curve, a simulation was performed with V_A set to 0.1, 0.15 and 0.2 and V_LV to 0.2, 0.15 and 0.1, respectively. In addition, for each tissue curve, V_A, V_LV and V_RV were set to 0 and a fit was performed to obtain estimates without any confounding effects of V_A, V_LV and V_RV, and these values were used as reference values. Then, the obtained time-activity curves were fitted using both Equations A2 and A3 and the reversible one- and two-tissue compartment models, and the obtained K₁ and V_T were compared with the reference values. For healthy tissue, bias in the obtained results is shown in Figure 6. Similar results were obtained for scar and viable but denervated tissue. As expected, bias was observed for both BV and SO. In general, magnitude and variation in bias were larger for BV than for SO, indicating that estimates obtained with SO are more stable and closer to the true simulated tissue values. For these reasons, SO was considered to be the most robust model interpretation. In addition, V_A is relatively constant in the myocardium (around 10% [41]), and therefore, it is likely that spill-over has a larger and more variable contribution to V_A. Therefore, it is expected that bias in SO, which is dependent on the magnitude of V_A, is also relatively constant. In contrast, bias in BV is dependent on the magnitude of V_LV, which depends on scanner resolution, segment size and thickness of the myocardial wall, and therefore, it is expected to vary between patients and scans. For these reasons, SO was used in the present study.