Value of PET ECG gating in a cross-validation study of cardiac function assessment by PET/MR imaging

This study was performed in a cohort of 30 patients (demographic information in Table 1) with known chronic total occlusion previous to revascularization procedure to assess myocardial viability using an integrated PET/MRI. All subjects gave their written informed consent in accordance with the guidelines of the local ethics board that approved the study 169/16 S.

Imaging was performed using a simultaneous PET/MR system (Biograph mMR, Siemens Healthcare, Erlangen, Germany). ECG signal was recorded with MR-compatible 3-lead electrodes and used for both cardiac MRI acquisition and ECG-gated PET reconstructions.

Conventional multi-slice 2-dimensional short axis CINE sequences were used to obtain the reference values of the volumes and the ejection fraction of the left ventricle at the beginning of the PET/MR exam. The images have a reconstructed matrix size: 256 × 208, number of slices: 10-13, voxel size: 1.33 × 1.33 × 8 mm³, spacing between slices: 8 mm, and temporal resolution: 25 cardiac phases.

To optimize glucose uptake in the heart and to standardize the metabolic environment in all patients, a hyperinsulinaemic-euglycaemic clamp procedure was performed. After stabilization of the plasma glucose level for approximately 60 minutes, 330 ± 32 MBq of fluorine-18 fluorodeoxyglucose (FDG) (4 MBq prescribed per kg body weight) was administered intravenously. The electrocardiographic-gated (ECG gated) list-mode PET was started approximately 60 minutes after the intravenous injection of FDG. PET data were acquired during the whole PET/MR exam (average duration of 42 minutes with a range [40-50] minutes). PET images were reconstructed using e7tools framework (Siemens Healthcare, Knoxville, TN) with an ordinary Poisson ordered-subset-expectation maximization iterative reconstruction algorithm (OP-OSEM) with 3 iterations—21 subset, matrix size 344 × 344, zoom: 1, reconstructed voxel size: 2.08 × 2.08 × 2.03 mm³, and 8 cardiac phases. Attenuation correction maps were generated from a Dixon-based MRI sequence under breath-hold at end-expiration. Additionally, the maximum-likelihood reconstruction of attenuation and activity (MLAA) algorithm was used to correct for arm truncations in the Dixon attenuation maps occurring due to the limited MRI field-of-view.⁸

ECG-based cardiac gating with 8 phases was performed on PET list-mode data. Cardiac cycles (hereafter R-R intervals) were considered as the time between two consecutive R peaks detected by the PET/MR system. Each R-R interval was divided into gates and the PET counts acquired in each one were grouped to reconstruct an image per gate.

Three cardiac gating methods were studied (Figure 1):

The cardiac gating methods STD and STD-BR correspond to the default approaches available in the PET/MR, while FW addresses the physiological finding that the systolic phase remains constant among different cardiac cycles.⁹

Volume measurements and analysis workflow are presented in Figure 3. End-diastolic and end-systolic volumes (EDV and ESV) as well as the corresponding left ventricular (LV) ejection fraction (EF) were assessed in both modalities using the software MunichHeart (MH).¹⁰ For 2-dimensional CINE images, the LV chamber was manually delineated by a cardiologist following standard guidelines.¹ EDV and ESV were defined as the maximum and minimum values of ventricular volumes measured from phases that were visually defined as end systole and end diastole. Papillary muscle volumes were considered part of the LV chamber.

To analyze ECG-gated FDG-PET images, MH computed EDV and ESV using a geometric model based on regional uptake threshold of the myocardium using a manual definition of the long axis and the base plane of the LV. End-diastolic and end-systolic phases were defined automatically based on the PET counts in the cardiac cycle.¹⁰

The multi-modality comparison of EDV, ESV and EF from MRI and PET was performed via correlation analysis using Pearson’s r coefficient and linear regression [Deming’s method, considering the errors for the values of both modalities]), and via agreement analysis with a paired Wilcoxon test for mean comparison, and bias and limits of agreement of Bland-Altman was calculated with stats, deming, and blandr R packages (version 4.0.4, The R Foundation for Statistical Computing, Vienna, Austria). Differences were considered as statistically significant when corresponding statistic test presented a P value < .05.

MRI- and PET-derived volumes as well as EF values presented overall high correlations (Figure 4). In the case of EDV, the slopes were 1.23 ± 0.18 (Pearson’s r = 0.75) for STD, 1.19 ± 0.15 (r = 0.81) for STD-BR, and 1.29 ± 0.16 (r = 0.80) for FW. In ESV, the slopes were 0.99 ± 0.08 (r = 0.92) for STD, 1.05 ± 0.08 (r = 0.92) for STD-BR, and 1.10 ± 0.09 (r = 0.92) for FW. Furthermore, in the case of EF, the slopes were 0.68 ± 0.09 (r = 0.79) for STD, 0.81 ± 0.07 (r = 0.91) for STD-BR, and 0.72 ± 0.07 (r = 0.87) for FW. Hence, PET ESVs correlated better with MRI than PET EDVs, and in particular, the STD-BR method yielded the best while STD yielded the poorest correlation.

The comparison of mean values against MRI (Table 2) showed that PET imaging underestimated all the parameters, with statistically significant differences in EDV for all gating methods, and for EF also STD and FW.

Comparison of Bland-Altman plots of volumes and ejection fraction obtained with MRI and PET are presented in Figure 5. The mean relative biases for the three methods in EDV were in the range (9-13)% with wide relative limits of agreement (60-68)%. For ESV, relative biases were much less pronounced (− 1-3)%, with similar limits of agreements (65-70)%. The range of EF relative biases was (5-14)% with limits of agreement (49-71)%. Here, there were no marked differences between the STD-BR and FW methods. However, STD showed the strongest disagreement.

Results from the STD gating approach showed the importance of the beat acceptance window. The inclusion of irregular R-R intervals in the gating process drives to merging PET counts from dissimilar cardiac phases, leading to an effective smoothing of the myocardium and subsequently affecting the estimation of left ventricular volumes. The similar performance of BR-STD and FW in this study indicated that the effects associated with the selection of the gate width could not be clearly discerned, potentially due to the higher impact of differences linked to the multimodal comparison PET-MRI, as discussed below.

Even though the STD-BR method avoids abnormal R-R intervals, the rejection of PET counts increases image noise, affecting the accuracy of the measurement. For instance, in one of the subjects, only 12 out of 40 minutes was used to reconstruct the PET image due to 40% of abnormal R-R intervals (equivalent to rejection of 66% of the total time)—an effect which would be even more detrimental in shorter exams. Extra efforts are needed to reduce this loss in counts: exhaustive quality controls to the ECG signal and R peak detected during acquisition, improvements of R peak detection algorithms (in particular, incrementing robustness to MR artefacts^11,12), generalization of ECG gating taking thoughtfully into account arrhythmic heartbeats as well.

In particular, it is important to emphasize the need for a strict quality control of the ECG signal and its processing in the clinical environment, since these artefacts affect not only the cardiac PET gating method, but also the acquisition of the cardiac MRI sequences that use this signal as a trigger.¹² Further investigations for alternatives to ECG-based cardiac gating are also encouraged.^13,14

Comparable works^15‐20 studied FDG-PET and MRI-based values of EDV, ESV and EF obtained using PET/MR, PET/CT, PET and MRI systems. Table 3 summarizes the main results. In general, we can see high correlations, but varied biases: (− 1.1-28.4)% for EDV, (− 5.9-29.6)% for ESV, and (− 12.0-13.6)% for EF, and principally wide limits of agreement: (44.3-95.4)% for EDV, (25.5-153.9)% for ESV, and (26.1-142.5)% for EF. Another comparative study²¹ between nitrogen-13 ammonia PET and MRI-based values reported higher correlations and lower biases, but the limits of agreements are still wide (33.3% for EDV, 48.2% for ESV, and 29.7% for EF). These comparisons support the evidence that cardiac function measurements are not interchangeable between modalities.

Although such a setting should ideally be the perfect scenario for multimodal comparisons with comparable physiological conditions,²² several multiple factors contribute to the observed variability. In addition to potential intra-/inter-observer variability, the discrepancies between the two modalities might be also explained by temporal and spatial resolutions, differences in the geometrical model of the heart between MRI and PET, and volume variation with heart rate changes.

Since the effective number of PET gates is three times less than in MRI, the measured PET-derived LV volumes are potentially undersampled (smoothing effect), affecting the accuracy of the parameters. Studies compared the volumes obtained with different numbers of gates (8, 16, and 32) in SPECT and PET images^23‐25: when using only 8 gates, smaller EDV, larger ESV, and a corresponding lower EF was observed, with the highest impact found in ESV measurement (median changes among those studies of 3% vs − 10%, in EDV and ESV, respectively). Nevertheless, temporal resolution issues might not be sufficient to explain the discrepancies in our cohort since the highest differences were found in EDV values.

The delineation of short and long axes, of base and apex of the heart in both modalities, and the lack of compatibility between them may also increase variability in the definition of the geometric model used to measure LV volume. In MRI, the axes are defined prospectively before the acquisition of the cine images, while in PET, the task is retrospectively carried out by physicians as a part of the data post-processing, adding to the inter-observer variability from two fully different directions. Another plausible factor of the discrepancies is an inaccurate FDG uptake-based PET contouring due to the severity myocardial damage, however, most of the LV myocardial segments of this cohort were classified as viable by FDG-PET assessment.

The lack of a perfect time matching (different time and duration) between PET and MRI acquisition in the exam (total scan time in PET vs several minutes in MRI) potentially leads to parameters that may not fully coincide. Two studies^26,27 analyzed the heart rate dependency of LV volumes of cardiac phantoms in SPECT/CT and CT images and a range 40-100 beats per minute. They found differences in EF of up to 2.5% in the CT images with low temporal resolution (175 ms per gate) between 60 and 80 beats per minute, but on average of only 1% with higher temporal resolutions (75 ms in CT and 38 ms in SPECT per gate). Regarding our study, even though PET temporal resolution was on average 120 ms (62 beats per minute), the spatial resolution was lower than with CT images, and thus intra-scan heart rate variation might be considered as part of the source for the discrepancies.

Even though the current clinical CINE MRI sequences provide images with the highest in-plane spatial and temporal resolutions, clinical 2-dimensional acquisition schemes limit the accuracy of inter-plane information. Due to the sequential, slice-by-slice acquisition, 2-dimensional images are more susceptible to motion, and the thickness of slices is acquired with worse spatial resolution than in the in-plane situation to cover the entire LV chamber in a feasible clinical acquisition time. Consequently, the use of cine MR values as the reference should be re-evaluated. In this sense, 3-dimensional acquisition for cine images seems to be the next step. However, clinically suitable schemes for its implementation are still in development and might hold other limitations in store.²⁸