Quantitative three-dimensional cardiovascular magnetic resonance myocardial perfusion imaging in systole and diastole

Forty consecutive patients with known or suspected CAD were recruited. All patients were imaged within 30 days of clinically scheduled diagnostic coronary angiography. No revascularization or clinical events occurred between angiography and CMR. Exclusion criteria were contraindications to CMR, adenosine, or gadolinium contrast agents, recent myocardial infarction (MI) or unstable angina (within 6 months), or poorly controlled arrhythmias. Patients were instructed to refrain from caffeine for 24 hours before their CMR study but continue cardiac medications as normal. The study was approved by the regional ethics committee and all patients gave written consent.

All studies were performed on a 3.0 T scanner (Achieva TX, Philips Healthcare, Best, The Netherlands) equipped with dual-source parallel radiofrequency transmission technology and a 32-channel cardiac coil. For perfusion imaging, a 3D spoiled turbo gradient-echo sequence was used (TR/TE/flip angle 1.8 ms/0.7 ms/15°; saturation prepulse delay 150 ms; linear k-space encoding; 70% partial Fourier acquisition in two dimensions; typical field of view 350 × 350 mm; 10 fold k-t acceleration and 11 training profiles leading to a net acceleration of 7.0; typical acquisition duration 192 ms, k-t principal component analysis (PCA) reconstruction; reconstructed to 12 contiguous slices with voxel size 2.3×2.3×5 mm³)[2, 7].

Two k-t undersampled 3D data sets were acquired in each R-R interval, each preceded by a non-selective saturation prepulse. Vertical and horizontal long-axis cine images were used to identify appropriate trigger delays for systolic and diastolic acquisition[8‐10]. Additionally, because of the longitudinal lengthening of the heart from systole to diastole, the position of the end-systolic and mid-diastolic perfusion stacks (12 slices each) were individually planned from the chosen systolic and diastolic cine frames (Figure 1)[8, 9]. The same trigger delays were used for stress and rest acquisitions.

Stress perfusion images were acquired during intravenous adenosine-induced hyperemia administered for 4 min at 140mcg/kg/min. Consistent with previous 3D-perfusion CMR studies, an intravenous bolus of 0.075 mmol/kg of gadobutrol (Gadovist, Bayer Schering Pharma, Berlin, Germany) was administered at a rate of 4.0 ml/s followed by a 20 ml saline flush. Stress perfusion CMR was followed by cine imaging covering the left ventricle in 10-12 short-axis sections. Rest perfusion CMR was performed 15 min after stress, using identical imaging parameters. Late gadolinium enhancement (LGE) imaging was acquired in the same short-axis geometry as perfusion imaging after an additional 10-15 min using conventional 2D methods (T1 weighted segmented inversion recovery gradient echo; TR/TE/flip angle 4.9 ms/1.9 ms/15°; inversion time individually adjusted according to Look-Locker scan; spatial resolution 1.35 × 1.35 × 10 mm).

Systolic and diastolic perfusion images were analyzed in separate reporting sessions in random order (MM, 2 years experience of perfusion CMR). Overall image quality was scored as follows: 0 = non-diagnostic, 1 = poor, 2 = adequate and 3 = excellent. The occurrence of artifact related to respiratory-motion, k-t reconstruction or dark-rim artifact was scored as follows: 0 = none, 1 = mild, 2 = moderate and 3 = severe.

Perfusion images were processed offline using previously validated in-house software (PMI 0.4; written in IDL 6.4 (ITT Visual Information Systems, Boulder, CO)[11]. All short-axis slices with clearly identifiable LV cavity enhancement during first-pass perfusion and with >75% circumferential LV myocardium were included in the analysis[5, 6].

The above analysis was repeated on a per territory basis using the 17-segment AHA model adjusted for coronary dominance[14]. For this, all slices from an individual patient were first visually allocated to basal, mid or apical sections of the model. For each perfusion territory, a myocardial ROI was then outlined including all segments pertaining to that territory across all slices according to the 17-segment AHA model. MBF and MPR estimates were obtained using the same algorithm as for the whole-heart ROI.

Thirty random territories were re-analyzed 1 month later by the same observer (M.M.) and by a second observer A.K. (2 yrs and 1 yr experience respectively). A.K. was blinded to the results of all previous analyses.

Quantitative coronary angiography was performed (QCAPlus, Sanders Data Systems, Palo Alto, CA, USA) on anonymised X-ray angiography images (M.M. 6 years experience in coronary angiography). Significant CAD was defined as luminal stenosis ≥70% diameter in any of the main epicardial coronary arteries or their branches with a diameter of ≥2 mm.

Analysis was performed using SPSS 17.0 (SPSS, Chicago, IL). Data are presented as mean ± SD. Group means were compared using paired or unpaired Student t-tests; or within-subjects analysis of variance with Greenhouse-Geisser correction for multi-sample sphericity, as appropriate. Ordinal data were compared using the Wilcoxon signed-rank test. Receiver-operating characteristic (ROC) analysis was performed on a per territory basis, to determine the diagnostic accuracy of MPR to detect significant CAD. Diagnostic accuracies are presented as area under the ROC curve (AUC); and were compared between systole and diastole using methods described by Delong and Delong. Optimal MPR cut-off values, for both cardiac phases, were defined as values that maximised the sum of sensitivity and specificity. A secondary ROC analysis was performed to evaluate the diagnostic accuracy of stress MBF diastolic/systolic ratio. To assess reproducibility, the coefficients of variation (CoV) for intra- and inter-observer measurements were calculated. Because three coronary territories were examined per patient, the intra-cluster correlation coefficient (ICC) was calculated for MPR estimates to determine the design effect and the need to adjust data for clustering. All statistical tests were 2-tailed and a p value <0.05 was considered significant.

Five of the 40 recruited patients were excluded: 3 because of claustrophobia and 2 owing to technical problems (1 mistimed contrast injection; 1 significantly altered patient position between stress and rest scans). A total of 35 patients (105 coronary territories) were therefore available for analysis. Table 1 shows the baseline patient characteristics. QCA confirmed significant CAD in 20 patients (57%) and 38 coronary territories (36%). Only 3 patients had evidence of MI on LGE imaging (the same 3 patients with a clinical history of MI), and this involved only 3 of the 105 territories analysed. Figure 2 shows an example of the stress perfusion images acquired in a patient with significant CAD.

Overall image quality was better in systole than in diastole (median image quality score: 3 vs. 2 respectively; p = 0.002). In diastole, there was a greater frequency of dark-rim artifact (19 patients [54%] vs. 9 patients [26%] and a higher overall artifact score compared to systole (median scores: 1 vs. 0 respectively; p < 0.0001).

In 5 patients (14%), perfusion images (both cardiac phases) were affected by k-t reconstruction artifacts at stress and/or rest due to respiratory motion, but all of these artifacts occurred at the end of the breath-hold and did not affect analysis of the first-pass perfusion images.

Estimates of MBF and MPR for both cardiac phases are seen in Tables 2,3 and4. On per patient (n = 35) and per territory analysis (n = 105), mean resting MBF was similar in both cardiac phases (all p values > 0.05); but mean stress MBF and MPR were significantly greater in diastole than systole (all p values <0.001). These relationships existed in normal and CAD subgroups, as well as overall (all p values <0.01) (Tables 2,3,4). In both cardiac phases, stress MBF and MPR were significantly lower in the presence of CAD than in normal patients (all p values < 0.01) or normal territories (all p values < 0.0001) (Tables 2,3,4).

Analysis of the normal patient group (n = 15) found no significant regional differences in stress MBF, rest MBF or MPR between the LAD, LCX or RCA perfusion territories in both phases (all p values <0.05) (Table 5).

Based on MPR estimates, the ICC for this study was low (r1 =0.09 [95% CI: -0.25 to 0.41]) with a design effect of 1.18 where cluster size = 3. This shows that the study design of using 3 ‘related’ coronary territories per patient (n = 35) to derive a sample size of 105 coronary territories does not significantly diminish statistical power.

On per territory analysis (n = 105), MPR also had a high overall diagnostic accuracy for the detection of significant CAD, and this was similar in both cardiac phases (AUC, systole: 0.92 vs. diastole: 0.94; p = 0.41) (Figure 3). The optimal MPR cut-off value was 1.75 for systole and 2.02 for diastole (Figure 4). At these thresholds, the sensitivity and specificity were 82% and 93% respectively for systole; and 87% and 94% for diastole. The diagnostic accuracy of MPR to detect CAD in each of the 3 coronary territories is shown in Table 6 and no significant differences were seen between cardiac phases.

The diastolic/systolic stress MBF ratio was significantly lower for territories with CAD than in normal territories (1.07 ± 0.06 vs. 1.17 ± 0.11; p < 0.0001). On ROC analysis, the diagnostic accuracy (AUC) of this ratio to detect significant CAD was 0.79. The optimal cut-off value for the ratio was 1.10 which gave a sensitivity of 82% and specificity of 76% (Figure 5).

The mean absolute difference between intra-observer measurements of MPR was similar in systole and diastole (0.30 ± 0.15 vs. 0.36 ± 0.13; p = 0.09); and the corresponding CoVs were both 18%. The mean absolute difference between inter-observer measurements of MPR was also similar in systole and diastole (0.35 ± 0.17 vs. 0.41 ± 0.15; p = 0.07) with corresponding CoVs of 20% and 21% respectively.

The spatio-temporal undersampling methods required for 3D data acquisition are sensitive to respiratory motion, cardiac arrhythmia and low-pass temporal filtering - all of which pose additional challenges to quantitative assessment. Low-pass temporal filtering in particular may lead to underestimation of MBF. We reduced these limitations by use of the constrained k-t PCA framework for image reconstruction, which has been shown to improve temporal fidelity, permitting robust measurements of MBF at very high acceleration factors[32]. The latter is also less prone to respiratory artifact as temporal basis functions are derived based on the low-resolution training data acquired in every heartbeat[2].

Although MPR performed well in our study, perfusion imaging is a measurement of the hemodynamic consequences of a stenosis rather than its anatomy, and therefore our use of QCA is an imperfect reference standard. Following this initial feasibility study, future validation against fractional flow reserve is planned in a larger clinical population. Another limitation, common to many previous studies, is the potential effect of data clustering as three perfusion territories are examined per patient[20, 27]. However, the design effect of this was low (1.18) owing to a small ICC and cluster size.

Finally, the model used for estimating MBF assumes a linear relation between signal and contrast agent concentration i.e. ignoring saturation effects in the LV blood pool which can lead to underestimation of MBF[33]. This is particularly relevant as we used a relatively high contrast agent dose of 0.075 mmol/kg bodyweight to be consistent with previous 3D-perfusion CMR studies. Proposed solutions include the use of a non-linear signal model combined with precontrast T1-mapping and/or the use of a small pre-bolus to measure the AIF. There is currently no evidence that either of these potential solutions actually leads to improved diagnostic accuracy for the detection of CAD in the clinical setting. In fact the only study directly addressing this question came to the opposite conclusion i.e. the use of a pre-bolus AIF was found to reduce diagnostic accuracy compared to a single-bolus approach[34].