Automatic Extraction of Myocardial Mass and Volume Using Parametric Images from Dynamic Nongated PET

DISCUSSION

This study showed the feasibility of calculating m_LV, wall thickness, volume, and LVEF using parametric PET images and a high degree of automation. The method does not use electrocardiogram gating; instead, it uses the same dynamic images as for routine quantification of myocardial blood flow (reflected in K₁) and myocardial oxygen consumption (quantified using k₂). Myocardial oxygen consumption is commonly used to calculate the efficiency with which the myocardium converts energy into effective work (22). Use of simultaneous assessment of myocardial blood flow and oxygen consumption from ¹¹C-acetate is limited, although it has been shown to be a valuable tool in studying heart failure (23). Global myocardial blood flow and oxygen consumption cannot be quantified without segmentation of the LV wall, and in this study the segmented LV wall was used for additional analysis of LV dimensions and function. The method is highly automated and can therefore be implemented in routine analyses that require segmentation of the left ventricle for calculation of K₁ and k₂, regardless of whether LV mass and volume are of interest. Because any tracer described by a single-tissue-compartment model can yield parametric images of K₁ and k₂, the method is in theory applicable to other tracers, although this possibility requires further validation. Similarly, it is expected that regions with scarring (i.e., low K₁) will be segmented correctly because the entire K₁ profile (Fig. 1D) is lowered, resulting in proportionally lowered relative cutoffs and correct placement of the endo- and epicardial contours. However, this expectation, too, remains to be validated.

Correlations and agreements for all studied variables were high. As indicated by the analyses of repeated measurements, the accuracy of the new PET method is comparable to that of CMR, and intra- and interobserver variability was minimal, with correlation coefficients close to 1. Combining parameters (e.g., for stroke volume and LVEF) propagates and enlarges small measurement inconsistencies in original parameters (i.e., EDV and ESV), resulting in reduced correlations. However, each of the primary measurements—m_LV (3,4), ESV (1), and EDV (24)—holds prognostic information on its own, and routine calculation of these parameters is relevant in most clinical settings. Furthermore, the repeated study analysis indicates that the new PET method may equal CMR for precision in serial studies of cardiac geometry.

m_LV was slightly but significantly overestimated in patients and slightly but significantly underestimated in controls (Table 1), resulting in a significant group difference. This difference was correlated to the actual thickness only, and it is therefore likely that partial-volume effects, more prominent in hearts with normal thickness, resulted in an underestimation of m_LV in controls. CMR thickness was significantly higher in patients than in controls (P = 0.01), as was expected because hypertrophy is a hallmark of aortic stenosis. PET thickness showed an even more significant group difference (P = 0.0001), a consequence of the partial-volume effect. In nongated PET, thickness is averaged through the cardiac cycle, whereas in CMR, thickness is measured at end-diastole. Thickness should therefore be higher with PET than with CMR in all subjects but was found to be so in patients only. A likely explanation is that the overestimation is counteracted by setting K₁ to 0 in voxels with V_A or V_V > 0.6, which could remove part of the edges of the LV wall from the contours. Since the edges have a relatively larger contribution in thinner hearts than in thicker hearts, this effect is more prominent in controls and may explain why thickness and m_LV were overestimated in patients but not in controls.

In controls, the mean thickness with each modality was identical, suggesting that partial-volume effects counter the averaging effect of measuring thickness from nongated PET images.

PET estimates were accurate for ESV and EDV, and only a small difference was found for m_LV. For both stroke volume and LVEF, both fixed and proportional biases in PET-derived values were found. Therefore, care has to be taken in interpretation of LVEF and stroke volume values. Indicator-dilution methods can be applied to dynamic PET data for assessing forward stroke volume (15), which may serve as an alternative.

The high degree of automation resulted in excellent inter- and intraobserver reproducibility. The only manual step in data processing was determining which short-axis slices to include. However, the use of only slices that showed a high K₁ in at least 50% of the circumference provided sufficient standardization, and after a short training interval, consistent results could be obtained. Similarly, the reproducibility of repeated measurements was high for PET. The intraclass correlation coefficient was minimally but significantly higher for CMR m_LV, whereas all other parameters were reproduced with insignificant differences.

Segmenting the LV wall was done with K₁ images, which are associated with higher contrast and less noise than electrocardiogram-gated images. This advantage provides an inherent filtering step in the analysis, potentially resulting in higher reliability, accuracy, and precision of the estimates than is possible with electrocardiogram-gated images. The cutoffs used for removal of blood regions from parametric images were based on previous studies (20). The two-thirds cutoff used during the radial profile step of segmenting the LV wall, and the 3-cm maximum distance of LV tissue from the blood, were chosen empirically because these values showed the most consistent results and required minimal manual intervention in a pilot subset of 10 patients. Using these cutoffs, only the outflow tract had to be opened manually in some cases by dragging endo- and epicardial contours to the same location. No other user intervention was required. Calculation of ESV and EDV was based on segmentation of parametric V_A images, and the optimal cutoffs were derived empirically using the same subset of 10 patients. The cutoff of the arterial blood fraction V_A for ESV was varied between 0.5 and 0.9 in steps of 0.05, and a value of 0.70 showed the highest correlation and the smallest differences from CMR data (Supplemental Table 1; supplemental materials are available at http://jnm.snmjournals.org). For EDV, the cutoff was also varied, between 0 and 0.4 in steps of 0.025, and 0.175 was found to give the highest agreement with CMR values (Supplemental Table 2). When a different subset of 10 patients was randomly chosen, the same cutoffs were obtained in 72% and 50% of the subsets for ESV and EDV, respectively, illustrating the consistency of the results.

The lack of electrocardiogram gating did not seem to reduce the accuracy of the method. The correlation coefficient for m_LV was higher than those found in two previously published studies (r = 0.82 (7) and 0.7 (9)), with a smaller difference and narrower range (2.3 ± 18.1 g in our study as compared with 4 ± 32 g in one of those previously published studies (9)). However, both of those studies (7,9) included patients with large defects whereas the current study did not. Areas with low tracer uptake may be more challenging to accurately delineate, explaining the difference in correlation between those two studies and ours. However, when relative cutoffs based on radial profiles are used, as in the current study, it is likely that the presence of defects will not significantly affect the accuracy of m_LV, although this possibility requires further validation. For thickness, the correlation coefficient was in line with that described in a direct comparison of echocardiography and MRI (25) (r = 0.76), suggesting that thickness derived using parametric PET images is precise enough for widespread use.

For EDV and ESV, correlation coefficients were in line with the 0.90–0.93 for EDV and 0.90–0.99 for ESV found in the literature for studies that used electrocardiogram-gated methods with a similar width of the 95% confidence interval during Bland–Altman analysis (ESV: ±19.7 mL in our study and 8–88 mL in the literature; EDV: ±37.9 mL in our study and 26–92 mL in the literature). Finally, the correlation coefficients for LVEF were also in line with the literature (0.76–0.96 (6–11)), with similar width of the 95% confidence interval (±11.6% in our study and 8%–20% in the literature). In the current cohort, only 5 of 35 patients had an LVEF below 50%, of which 4 had an LVEF in the range of 45%–50%. A more comprehensive assessment in measuring LVEF is required in patients with low LVEFs.

The fact that the method presented in this study relies on image segmentation and empirically derived cutoff values suggests that factors influencing the final image resolution can affect all studied parameters by influencing the shape of the profiles (Fig. 1D) and of the V_A images. More studies are warranted to evaluate the impact of scanner type, reconstruction protocol, and input function definition on the new method and to study the generic applicability of the cutoff values used here. In addition, as in all PET measurements, subject motion may play a role. However, estimation of V_A and V_V is based mainly on activity during the first pass of the bolus through the heart. This phase is typically less than 30 s, reducing the probability that subject motion will affect V_A and V_V. For m_LV estimates, only severe subject motion will significantly affect K₁ images. Such motion may cause, for instance, misalignment between V_A and K₁ images, resulting in gaps in K₁ images due to the V_A > 0.6 threshold used in generating K₁ images, or artificially high K₁ values in some regions. Both artifacts are obvious and, when identified, can compromise not only m_LV estimates but any regional quantitation and the scan should be discarded for any purpose.

Accurate and highly reproducible measurements of cardiac geometry obtained in an almost fully automated fashion are desirable with any clinical modality. Although there is a growing interest in the use of quantitative PET in cardiology, cardiac PET is still used mainly in academic centers and does not rival echocardiography or CMR in accessibility. The data from this study support the use of the method for research studies involving PET in healthy volunteers and early phases of cardiac disease where information on LV remodeling is required. The nongated approach may be of particular relevance in patients with arrhythmias. Expanding the use of the novel methods into routine diagnostic work requires validating studies with other tracers and other scanners in more patient categories and larger reference databases. Specifically, studies with standard perfusion tracers, both during rest and during stress, and ¹⁸F-FDG in patients with chronic ischemic heart disease are needed to assess the relevance of the methodologic assumptions.