Author + information
- Received February 23, 2017
- Revision received January 26, 2018
- Accepted February 22, 2018
- Published online May 7, 2018.
- John D. Biglands, BSc, MSc, PhDa,b,∗ (, )
- Montasir Ibraheem, MSca,
- Derek R. Magee, BSc, PhDc,
- Aleksandra Radjenovic, BSc, MSc, PhDd,
- Sven Plein, MD, PhDa and
- John P. Greenwood, MBChB, PhDa
- aDivision of Biomedical Imaging, Leeds Institute of Cardiovascular and Metabolic Medicine, University of Leeds, Leeds, United Kingdom
- bDepartment of Medical Physics and Engineering, Leeds Teaching Hospitals NHS Trust, Leeds, United Kingdom
- cSchool of Computing, University of Leeds, Leeds, United Kingdom
- dInstitute of Cardiovascular and Medical Sciences, College of Medical, Veterinary and Life Sciences, University of Glasgow, Glasgow, United Kingdom
- ↵∗Address for correspondence:
Dr. John D. Biglands, Room 8.6, Division of Medical Physics, Worsley Building, University of Leeds, Leeds, West Yorkshire LS2 9JT, United Kingdom.
Objectives This study sought to compare the diagnostic accuracy of visual and quantitative analyses of myocardial perfusion cardiovascular magnetic resonance against a reference standard of quantitative coronary angiography.
Background Visual analysis of perfusion cardiovascular magnetic resonance studies for assessing myocardial perfusion has been shown to have high diagnostic accuracy for coronary artery disease. However, only a few small studies have assessed the diagnostic accuracy of quantitative myocardial perfusion.
Methods This retrospective study included 128 patients randomly selected from the CE-MARC (Clinical Evaluation of Magnetic Resonance Imaging in Coronary Heart Disease) study population such that the distribution of risk factors and disease status was proportionate to the full population. Visual analysis results of cardiovascular magnetic resonance perfusion images, by consensus of 2 expert readers, were taken from the original study reports. Quantitative myocardial blood flow estimates were obtained using Fermi-constrained deconvolution. The reference standard for myocardial ischemia was a quantitative coronary x-ray angiogram stenosis severity of ≥70% diameter in any coronary artery of >2 mm diameter, or ≥50% in the left main stem. Diagnostic performance was calculated using receiver-operating characteristic curve analysis.
Results The area under the curve for visual analysis was 0.88 (95% confidence interval: 0.81 to 0.95) with a sensitivity of 81.0% (95% confidence interval: 69.1% to 92.8%) and specificity of 86.0% (95% confidence interval: 78.7% to 93.4%). For quantitative stress myocardial blood flow the area under the curve was 0.89 (95% confidence interval: 0.83 to 0.96) with a sensitivity of 87.5% (95% confidence interval: 77.3% to 97.7%) and specificity of 84.5% (95% confidence interval: 76.8% to 92.3%). There was no statistically significant difference between the diagnostic performance of quantitative and visual analyses (p = 0.72). Incorporating rest myocardial blood flow values to generate a myocardial perfusion reserve did not significantly increase the quantitative analysis area under the curve (p = 0.79).
Conclusions Quantitative perfusion has a high diagnostic accuracy for detecting coronary artery disease but is not superior to visual analysis. The incorporation of rest perfusion imaging does not improve diagnostic accuracy in quantitative perfusion analysis.
- cardiovascular magnetic resonance
- diagnostic accuracy
- myocardial ischemia
- quantitative myocardial perfusion
Cardiovascular magnetic resonance (CMR) is a well-established technique for the assessment of patients with coronary artery disease (CAD), being diagnostically superior (1,2), cost effective (3,4), and a better predictor of cardiovascular events (5) than myocardial perfusion scintigraphy by single-photon emission computed tomography. CMR compares favorably with positron emission tomography (6); has higher image resolution; is more widely available; does not use ionizing radiation; and can evaluate function, perfusion, and viability in the same investigation. Perfusion CMR requires the passage of a contrast agent bolus through the heart to be visualized over time. Typically a saturation-prepared single-shot readout sequence is used to achieve adequate coverage and spatial and temporal resolution (7,8). Post-processing of CMR perfusion images can generate estimates of absolute myocardial blood flow (MBF). Absolute MBFs provide an objective measure of perfusion that does not require a healthy region of myocardium for visual comparison. They have been used to show diffuse perfusion changes caused by smoking (9) and type 2 diabetes mellitus (10), and there is evidence to suggest that these measurements may bring improvements in diagnostic performance (11). However, assessments of the diagnostic accuracy of MBF estimates have been limited to small studies (typically <50 patients) (11–15). Perfusion is often expressed as myocardial perfusion reserve (MPR = stress MBF/rest MBF). However, it is unknown whether the use of MPR values improves diagnostic performance over stress perfusion measurements alone. If not, the time-consuming rest perfusion scan could potentially be removed from the acquisition protocol without reducing the performance of the test (16,17).
The primary objective of this study was to compare the sensitivity, specificity, and diagnostic accuracy of expert visual analysis and MBF estimates against a reference standard of quantitative coronary angiography (QCA). This was done using a large representative subsample of the CE-MARC (Clinical Evaluation of Magnetic Resonance Imaging in Coronary Heart Disease) study (2). We hypothesized that quantitative CMR would have a higher diagnostic accuracy than visual analysis for identifying significant coronary artery stenosis. A secondary objective was to compare the diagnostic accuracy of MPR measurements, which use both rest and stress MBF data, with stress MBF measurements only.
The study protocol was approved by the national research ethics service. CE-MARC recruited patients with suspected angina pectoris, of which 676 had assessable CMR and angiography (2,18). For this substudy, 128 cases were randomly selected by an independent statistician from the CE-MARC population, such that the distribution of risk factors (hypertension, diabetes, smoking, age) and disease status (normal, single-, double-, or triple-vessel disease) was proportionate to those in the full population. This subsample contained 50 patients that have been included in a previous study (16).
Myocardial perfusion CMR and QCA data were acquired from each patient as previously described (2,18). All patients underwent invasive QCA within 32 days of their CMR examination. Adenosine (140 μg/kg/min) induced stress imaging was performed at least 15 min before rest imaging. Myocardial perfusion CMR was performed using a bolus intravenous injection of 0.05 mmol/kg dimeglumine gadopentetate (Magnevist, Schering AG, West Sussex, United Kingdom) through an arm vein at an injection rate of 5 ml/s. CMR imaging was carried out on 1.5-T Philips Intera (Best, the Netherlands) equipped with Master gradients (30 mT/m peak gradients and 150 mT/m/ms slew rate) using a 5-element cardiac phased-array coil and triggering performed by the vectorcardiographic method. Three short-axis images were acquired using a T1-weighted saturation recovery turbo field echo imaging sequence. A shared (nonslice selective) saturation pulse was used giving pre-pulse delay times to the center of k-space of 126 ms, 272 ms, and 418 ms for the basal, middle, and apical slices, respectively. The image acquisition parameters were as follows: echo time 1.0 ms, repetition time 2.7 ms, flip angle 15°, SENSE factor 2, matrix 144 × 144, field of view 320 to 460 mm, pixel size 2.2 to 3.2 mm, slice thickness 10 mm, and partial Fourier 0.67, giving a readout window of 130.2 ms per slice. Imaging continued until the first pass had been observed to pass through the myocardium. The average number of frames in the perfusion series was 56 (range: 26 to 78).
Late gadolinium enhanced (LGE) CMR was performed between 10 and 15 min after the rest perfusion study with a T1-weighted, segmented inversion-recovery gradient echo sequence. Pulse sequence parameters were as follows: echo time 1.9 ms, repetition time 4.9 ms, flip angle 15°, 10 to 12 short-axis slices, single slice per breath-hold, matrix 240 × 240, and field of view 320 to 460 mm as per patient size. The optimal inversion time to null signal from normal myocardium was determined before the scan using a Look-Locker approach (19).
Quantitative CMR analysis was performed blinded to the results of all other investigations. Contours describing the myocardium and a region within the left ventricular blood pool, avoiding papillary muscles, were drawn using dedicated cardiac image analysis software (Mass 7.0, Medis, Leiden University, Leiden, the Netherlands). Contours were copied to all time frames and manually adjusted for motion. Adjustments were limited to rigid translations only. Manual contouring took around 1 h per patient. The myocardium was subdivided into 6 circumferentially equidistant regions in the basal and mid slices and 4 in the apical slice according to the American Heart Association (AHA) standard (20). Individual perfusion datasets exhibiting excessive (more than 1 frame) through plane motion (typically caused by electrocardiographic gating failure) were visually identified and excluded before MBF quantitation. Signal versus time curves from the myocardium and blood pool were converted to contrast agent concentration curves assuming a linear signal response to contrast agent as described previously (16). All pre-contrast signal estimates were taken from the stress study. Values of 1,435 ms and 4.3 s-1.mM-1 were used for the blood T1 and contrast agent relaxivity, respectively. To avoid remnant contrast agent from the stress perfusion scan affecting the rest perfusion analysis the pre-contrast signal intensity was subtracted from the rest perfusion curves before analysis. MBF values were estimated using Fermi-constrained deconvolution (16,21). The arterial input function (AIF) was taken from the basal slice. The pre-contrast baseline signal, end of first-pass time point, and the bolus arrival time delay between the blood pool and myocardial curves were calculated using previously described automated methods (16,22).
Visual CMR perfusion images were jointly reported by 2 cardiologists (J.P.G., S.P.) with >6 years’ experience in CMR at the time, and who were blind to the results of all other investigations. This was a perfusion-only assessment that did not take into account cine, LGE, or angiography images sets. Scores for hypoperfusion (ischemia) of 0 (normal), 1 (equivocal), 2 (subendocardial ischemia), or 3 (transmural ischemia) were given by visual comparison of stress and rest CMR perfusion scans (16 segments of the 17 segment AHA model, excluding the apical cap segment). To generate the receiver-operating characteristic (ROC) curve the summed scores over all AHA segments were used. Diagnostic performance was ascertained from the ROC curve as the area under the curve (AUC) value. The cutoff value that generated the optimal sensitivity and specificity for the test was determined by maximizing the Youden index (23). A separate assessment of LGE was performed with a score of 0 (none), 1 (1% to 25%), 2 (26% to 50%), 3 (51% to 75%), or 4 (>75%) allocated to each segment of the AHA model.
All x-ray angiograms were performed after CMR. QCA analysis was performed off-line by a cardiologist blinded to the CMR results using QCAPlus software (Sanders Data Systems, Palo Alto, California). Significant CAD was defined as ≥70% diameter stenosis of a first-order coronary artery measuring ≥2 mm in diameter, or left main stem stenosis ≥50%. Single-, double-, and triple-vessel disease was defined as significant stenosis affecting 1, 2, or 3 vessels, respectively. Both visual CMR perfusion and QCA scores were taken from the original CE-MARC reports and were not reanalyzed for this substudy.
All perfusion results were compared with QCA on a per-patient basis. MPR values were calculated as the stress MBF estimate divided by the resting MBF estimate. To generate the ROC curve the AHA segment with the lowest perfusion measure (MPR or stress MBF) was used as the quantitative measure.
Diagnostic performance was evaluated using ROC curve analysis taking the QCA diagnosis as the reference standard. Diagnostic performance was first assessed in terms of the ability of the perfusion index to detect disease in any coronary artery. A separate assessment of the diagnostic performance for detecting disease in each individual coronary artery was performed using the AHA segmentation recommendations to map myocardial segments to individual coronary arteries. The number of detected perfusion defects that correctly corresponded to disease in the coronary artery specified by the AHA mapping was then assessed.
Categorical variables are expressed as numbers and percentages. Continuous variables are expressed as mean ± SD unless otherwise stated. With a sample size of 128 and using a correlation between the scores of r = 0.45, the study was powered to detect a difference of 0.15 in the AUC values between ROC curves with a power of 80% at the 5% significance level (24). ROC curves were generated using Analyse-it (Analyse-it Software Ltd., Leeds, United Kingdom). All other statistical analysis was carried out using SPSS version 21.0 (IBM, Armonk, New York). Comparison of ROC curves was performed using the DeLong method (25). There was no correction for multiple comparisons of AUC curves. Normally distributed data were compared using Student’s t test.
Baseline patient characteristics are summarized in Table 1. The study consisted of 128 patients (mean age 61 years; age range 37 to 77 years). A total of 77 (60%) were men (mean age 61 years; age range 45 to 76 years) and 51 were women (mean age 60 years; age range 37 to 77 years). There was no significant age difference between male and female groups (p = 0.33). A total of 42 patients had significant CAD as assessed by QCA and 86 did not. Four whole patient perfusion datasets (3%) were excluded from the study because of severe through-plane motion caused by electrocardiographic triggering failures (3 stress scans and 1 rest scan). These consisted of 1 patient with single-vessel disease, 1 with double-vessel disease, and 2 healthy patients as assessed by QCA. Post-exclusion, 40 patients with significant coronary heart disease and 84 without remained for analysis. Analysis of LGE images showed 33 patients had evidence of myocardial scaring (infarct pattern).
Mean global (i.e., mean MBF per slice averaged over all 3 slices) MBF values over all 3 slices are shown in Table 2. Mean MBFs from healthy patients for each slice are shown in Table 3. Perfusion was significantly lower in patients with ischemia than in normal: stress MBF 2.16 ± 0.70 ml/min/g versus 3.00 ± 0.81 ml/min/g (p < 0.001), and MPR 1.86 ± 0.57 versus 2.31 ± 0.67 (p < 0.001). ROC curves for visual and quantitative perfusion analysis are shown in Figure 1. The sensitivity, specificity, AUC, and optimal cutoff values are shown in Table 4, and Table 5 shows the respective contingency tables. The highest diagnostic accuracy was achieved using MPR measurements. There was no statistically significant difference in diagnostic performance between visual (AUC: 0.88, cutoff 2.0) and quantitative analysis for MPR (AUC: 0.89, cutoff 1.11; p = 0.72) or stress MBF (AUC: 0.87, cutoff 1.27 ml/min/g; p = 0.54). There was no significant difference in diagnostic accuracy between MPR and stress MBF quantitative ROC curves (p = 0.79).
Separate assessments for patients with single-vessel, double-vessel, and multivessel (double or triple) disease are shown in Table 6. There was no significant difference between the diagnostic accuracy of visual and quantitative analysis in single-vessel, double-vessel, or multivessel disease groups. In 28 (70%) of 40 cases the minimum quantitative perfusion score mapped correctly to a coronary artery territory that contained a significant stenosis according to the AHA segmentation model. Eight of 9 (89%) defects correctly corresponded to a stenosis in the left circumflex, 12 of 19 (63%) correctly corresponded to a stenosis in the left anterior descending, and 8 of 12 (70%) correctly corresponded to a stenosis in the right coronary artery. Separate assessments for the individual coronary arteries are shown in Table 7. Quantitative measures (stress MBF or MPR) did not perform significantly better than visual analysis for any of the coronary arteries.
The primary finding of this study is that quantitative myocardial perfusion analysis has a high diagnostic accuracy but does not out perform expert visual analysis. In addition, diagnostic performance of quantitative perfusion was not significantly improved by including rest perfusion measurements. This suggests that the rest perfusion acquisition may not be necessary for quantitative analysis, potentially saving time, expense (less contrast), and patient inconvenience. To our knowledge this is the largest investigation into the diagnostic performance of quantitative CMR perfusion to date, around twice as large as the previous largest study, with n = 67 (11).
The presence of myocardial infarction can make visual diagnosis of superimposed ischemia challenging. In this study quantitation achieved a high diagnostic accuracy even though a significant number (n = 33) of cases in the study had myocardial infarction as assessed by LGE imaging. Therefore, the quantitative diagnostic accuracy reported in this study supports the robustness of this technique in real-world clinical cases.
Our data showed comparable diagnostic accuracy in single-vessel and multivessel disease with visual or quantitative analysis implying that there was no advantage in quantitative analysis in patients with different extents of CAD. Furthermore, we found similar diagnostic accuracies for the ability of visual or quantitative analysis to detect perfusion defects in the 3 coronary arteries (left circumflex, left anterior descending, and right coronary artery). Although MPR seems to perform slightly better than visual or stress MBF analysis, especially at sensitivities above 80% (Figure 1), these differences were not statistically significant. The high diagnostic accuracy observed using stress MBF alone agrees well with previous studies that analyzed stress-only images using semiquantitative (6,26,27), visual (28,29), and quantitative analyses. These observations demonstrate that stress data alone can yield excellent diagnostic performance and a rest perfusion study may not be necessary in a standard protocol to detect or exclude CAD. MPR is a measure of the potential flow increase the myocardium has in reserve before maximal vasodilation occurs. Whereas stress perfusion is uncoupled from oxygen demand resting perfusion is not (6), so factors influencing resting myocardial oxygen demand cannot be controlled for in a clinical setting. This uncontrolled aspect of the rest perfusion measurement may account for the fact that dividing by the rest MBF measurement did not improve diagnostic accuracy in our quantitative data.
Our finding that the diagnostic performance of quantitative perfusion is comparable with, but not significantly better than, visual analysis is consistent with previous, smaller studies (13,15). However, Mordini et al. (11) did report a diagnostic advantage using quantitation. This may be caused in part by the fact that Mordini et al. (11) measured the ratio between the endocardial segment and the median epicardial value and required at least 2 segments to fall below the threshold before a patient was classed as ischemic, whereas our study used the minimum segmental MBF score. Our study did not replicate this transmural subdivision strategy because of concerns over increasing the noise in the signal versus time curves.
MBF values in patients without ischemia were comparable with those published in studies of healthy volunteers (30,31). At 1.23 ml/min/g the resting MBF is somewhat higher than most studies, because of nonlinearity effects in the AIF, but still well within the range of MBF values quoted in the literature. The total exclusion rate was 3% (4 of 128). This compares favorably with other quantitative studies. For instance, Patel et al. (13) excluded 23% of patients and Costa et al. (14) excluded 16%.
The optimal threshold for abnormal perfusion from the ROC analysis was set at an MPR of 1.11 and a stress MBF of 1.27 ml/min/g. The MPR threshold is somewhat lower than other studies (Huber et al.  at 1.54; Patel et al.  at 1.55) possibly because of the high rest MBF measurements in our study. The stress MBF threshold of 1.27 ml/min/g was somewhat lower than that of Mordini et al. (11) at 1.58 ml/min/g, possibly because their model required 2 AHA segments below the cutoff threshold, whereas our model only required 1.
Perfusion CMR assesses myocardial ischemia, whereas QCA is a measure of coronary artery stenosis, which is itself an imperfect reference standard. Thus, false negative results could occur if lesions not causing ischemia (as assessed by CMR) were judged clinically significant on the basis of angiographic stenosis severity. Invasive measurement of fractional flow reserve is now the reference standard for the measurement of hemodynamic significance of a coronary artery stenosis, but was not routinely performed at the time of recruitment to the CE-MARC study.
The combination of a 0.05 mmol/kg contrast dose and a pre-pulse delay of 126 ms yields a nonlinear signal response to contrast agent concentration in the AIF, resulting in an overestimate of MBF. The lack of a linear AIF measurement constitutes a limitation to this retrospective dataset. This could potentially diminish the range of MBF estimates and reduce the performance of quantitative perfusion, including the benefits of rest perfusion. However, our analyses achieved a high sensitivity and specificity in agreement with other studies using dual-bolus techniques implying that nonlinearity errors have not profoundly affected the results. This agrees with previous work directly comparing dual- and single-bolus strategies and finding no significant difference in diagnostic performance (32).
The use of a shared pre-pulse to acquire all 3 perfusion slices results in different T1 contrast between the 3 image slices. This has been addressed by using the basal AIF for all 3 slices and by applying a linear correction to the myocardial curves. This approach may be subject to errors if the myocardial signal to concentration relationship is sufficiently nonlinear, although there were no significant differences in MBF between the 3 slices in the study population (Table 3).
Manual correction for breathing motion introduces an extra source of error into the measurements because the signal curves can be contaminated by high signal blood pixels in the left ventricle or other surrounding tissues, deteriorating the results of MBF quantitation. Although care was taken to avoid these errors, an automated, nonrigid registration might have improved our quantitative results. The use of quantitative perfusion analysis in clinical practice requires a known healthy/diseased threshold value. Currently this value may vary between studies because of variations in MBFs caused by differing methodologies. Before MBF measurements can be used widely standardization of these methods and multicenter studies are necessary to show that a single cutoff across different sites and CMR vendors is suitable and can still achieve the high diagnostic accuracies reported in this study. This is even more relevant if the rest perfusion measurement is to be discarded because expressing perfusion as a ratio can normalize systematic shifts in MBF, caused by differences in methodology, that remain if a stress-only perfusion measurement is used. It is also noteworthy that quantitative perfusion can impose limits on the acquisition, such as lower contrast dose and reduced image T1 weighting that can force a reduction in image quality and/or heart coverage, which may adversely affect visual assessment.
Quantitative myocardial perfusion has a high diagnostic accuracy for detecting CAD, but is not superior to expert visual analysis, even in multivessel disease. Rest perfusion data acquisition does not increase the diagnostic accuracy of quantitative myocardial perfusion and could be eliminated from the imaging protocol.
COMPETENCY IN MEDICAL KNOWLEDGE: This work has shown that quantitative myocardial perfusion estimates obtained from CMR have a high diagnostic accuracy equivalent to, but not better than, that of expert visual analysis. In addition, the use of a rest perfusion measurement did not improve diagnostic performance above stress perfusion quantitation alone. The clinical implications are that these observations support removal of rest perfusion imaging from the acquisition protocol.
TRANSLATIONAL OUTLOOK: For quantitative perfusion estimates to be accepted as a standard clinical tool several obstacles need to be overcome. First, the time-consuming analysis needs to be streamlined so that quantitative estimates are easily available to a nonexpert user within a reasonable time frame. Second, diagnosis using quantitative measurement requires a known healthy/diseased cutoff value. These cutoff points vary between studies because of the variation in quantitative values caused by the wide range of methods used. These include differences in contrast dose administered, CMR acquisition sequence, methods for correcting nonlinearity between contrast agent concentration and signal intensity, motion correction, and modeling methods used to generate the final flow value. If these measurements are to be used widely, standardization of these methods is required to reduce these variations. Multicenter studies would then be necessary to show that a single cutoff across different sites and CMR vendors is suitable and can still achieve the high diagnostic accuracies reported in this study.
The authors thank David Buckley and Steven Sourbron for their insightful comments on the perfusion quantitation method.
This study is independent research supported in part by the National Institute for Health Research (NIHR). The views expressed in this publication are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health. Dr. Biglands has received funding from NIHR fellowships (NIHR/RTF/01/08/014, and ICA-CL-2016-02-017). Dr. Magee is partially supported by WELMEC, a Centre of Excellence in Medical Engineering funded by the Wellcome Trust and Engineering and Physical Sciences Research Council (grant WT 088908/Z/09/Z). Dr. Radjenovic is partially supported by WELMEC, a Centre of Excellence in Medical Engineering funded by the Wellcome Trust and Engineering and Physical Sciences Research Council (grant WT 088908/Z/09/Z). Dr. Plein is funded by a British Heart Foundation fellowship (FS/10/62/28409). All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
- Abbreviations and Acronyms
- American Heart Association
- arterial input function
- area under the curve
- coronary artery disease
- cardiovascular magnetic resonance
- late gadolinium enhancement
- myocardial blood flow
- myocardial perfusion reserve
- quantitative coronary angiography
- receiver-operating characteristic
- Received February 23, 2017.
- Revision received January 26, 2018.
- Accepted February 22, 2018.
- 2018 American College of Cardiology Foundation
- Schwitter J.,
- Wacker C.M.,
- Wilke N.,
- et al.
- Walker S.,
- Girardin F.,
- McKenna C.,
- et al.
- Pletscher M.,
- Walker S.,
- Moschetti K.,
- et al.
- Schwitter J.,
- Nanz D.,
- Kneifel S.,
- et al.
- Fairbairn T.A.,
- Motwani M.,
- Mather A.N.,
- et al.
- Larghat A.M.,
- Swoboda P.P.,
- Biglands J.D.,
- Kearney M.T.,
- Greenwood J.P.,
- Plein S.
- Mordini F.E.,
- Haddad T.,
- Hsu L.-Y.,
- et al.
- Patel A.R.,
- Antkowiak P.F.,
- Kramer C.M.
- Costa M.A.,
- Shoemaker S.,
- Futamatsu H.,
- et al.
- Chiribiri A.,
- Hautvast G.L.T.F.,
- Lockie T.,
- et al.
- Biglands J.,
- Magee D.,
- Sourbron S.,
- Plein S.,
- Greenwood J.,
- Radjenovic A.
- Cerqueira M.,
- Weissman N.,
- Dilsizian V.,
- et al.
- Jerosch-Herold M.,
- Wilke N.,
- Stillman A.E.
- Schwitter J.,
- Wacker C.M.,
- van Rossum A.C.,
- et al.
- Schwitter J.,
- Wacker C.M.,
- Wilke N.,
- et al.
- Wang L.,
- Jerosch-Herold M.,
- Jacobs D.R.,
- Shahar E.,
- Folsom A.R.