Author + information
- Received September 21, 2003
- Revision received July 26, 2004
- Accepted July 29, 2004
- Published online November 2, 2004.
- ↵*Reprint requests and correspondence:
Dr. Peter G. Danias, Cardiac MR Center, Hygeia Hospital, 4 Erythrou Stavrou Street and Kifissias Avenue, Maroussi 15123, Greece
Objectives This study was designed to define the current role of coronary magnetic resonance angiography (CMRA) for the diagnosis of coronary artery disease (CAD).
Background Coronary magnetic resonance angiography has been proposed as a promising noninvasive method for diagnosis of CAD, but individual studies evaluating its clinical value have been of limited sample size.
Methods We identified all studies (MEDLINE and EMBASE) that evaluated CAD by both CMRA and conventional angiography in ≥10 subjects during the period 1991 to January 2004. We recorded true and false positive and true and false negative CMRA assessments for detection of CAD using X-ray angiography as the reference standard. Analysis was done at segment, vessel, and subject level.
Results We analyzed 39 studies (41 separate comparisons). Across 25 studies (27 comparisons) with data on 4,620 segments (993 subjects), sensitivity and specificity for detection of CAD were 73% and 86%, respectively. Vessel-level analyses (16 studies, 2,041 vessels) showed sensitivity 75% and specificity 85%. Subject-level analyses (13 studies, 607 subjects) showed sensitivity 88% and specificity 56%. At the segment level, sensitivity was 69% to 79% for all but the left circumflex (61%) coronary artery; specificity was 82% to 91%. There was considerable between-study heterogeneity, but weighted summary receiver-operating characteristic curves agreed with these estimates. There were no major differences between subgroups based on technical or population characteristics, year of publication, reported blinding, or sample size.
Conclusions In evaluable segments of the native coronary arteries, CMRA has moderately high sensitivity for detecting significant proximal stenoses and may have value for exclusion of significant multivessel CAD in selected subjects considered for diagnostic catheterization.
Imaging of the coronary artery lumen is currently best evaluated with X-ray angiography with selective intracoronary contrast injection. Besides establishing the diagnosis of coronary artery disease (CAD), angiographic data regarding luminal stenoses guide the selection between medical and interventional/surgical therapies. The number of diagnostic cardiac catheterizations has been steadily increasing over the past few years (1). This comes at a considerable cost, and though the complication rate for each individual is proportionally small (2), in total, cardiac catheterizations account for considerable procedure-related morbidity. A noninvasive diagnostic alternative is desirable.
Among proposed alternatives, coronary magnetic resonance angiography (CMRA) offers high spatial resolution, eliminates the need for iodinated contrast, and does not involve exposure to ionizing radiation. The many technical challenges have been addressed with various approaches, and CMRA has been the focus of clinical research over the past decade. Using two-dimensional (2D) CMRA, initial encouraging results (3) were followed by less optimistic reports (4), thus creating confusion on whether CMRA can have a clinical role in the evaluation of CAD. Subsequent reports with 2D CMRA (5–12) and three-dimensional (3D) CMRA (13–45) continued to show a seemingly broad range of sensitivity and specificity for detection of proximal CAD. It is unknown whether this variability has been due to chance or reflects differences in hardware platform, technique implementation, and subject selection criteria. Studies conducted to date have been of limited size, and thus the confidence intervals for the observed sensitivity and specificity are large. Given the sample size limitations, single studies have even less power for documenting whether CMRA might perform better for specific coronary vessels. To address these issues and to provide an evidence-based evaluation of the clinical utility of this new imaging modality, we performed a comprehensive meta-analysis of all currently available studies comparing CMRA with conventional X-ray coronary angiography.
Eligibility criteria and search strategy
We considered all studies published in English, French, or German evaluating the presence of significant CAD in native coronary arteries by both CMRA and X-ray angiography in the same subjects. Studies were eligible regardless of whether they referred to subjects with suspected or proven CAD and regardless of technique used for CMRA. We excluded studies with fewer than 10 subjects examined with both methods, as these would be unlikely to contribute meaningful diagnostic information. Studies evaluating vessel patency alone, phantom-only evaluations, and animal studies were excluded. Studies limited exclusively to subjects with normal coronaries and studies assessing coronary grafts were also excluded.
We identified eligible studies by searching MEDLINE and EMBASE (last search January 2004). The search combined exploded keywords pertaining to CAD and CMRA (strategy available upon request). In addition, we perused bibliographies of retrieved articles and reviews. The retrieved studies were carefully examined to exclude duplication or overlap of subjects. Duplicate/overlapping data were counted only once in the meta-analysis, unless the same subjects had been evaluated at different time points and/or with different CMRA techniques. When a single study included more than one well-defined subject group or CMRA techniques, each group/approach was considered as a separate comparison. We communicated with all corresponding investigators of the identified reports to clarify unclear data and to obtain key data that were not reported in the articles. Meeting abstracts were excluded, as they would not provide adequately detailed data and their results might not be final.
The following information was extracted from each pertinent study: first author, year of publication, and journal; study population characteristics including sample size (number of subjects evaluated with both tests and number of subjects evaluated with X-ray angiography only); percentage of subjects with documented CAD; gender; mean (or median) age (and standard deviation); relative timing of the two imaging procedures and blinding of evaluators of one test to the results of the other and to the clinical condition of the tested subject; technical characteristics of the CMRA (including electrocardiogram gating and fat signal suppression [yes/no], respiratory motion suppression method [breath-holding vs. navigator (retrospective vs. prospective)], mode of acquisition [2D vs. 3D], echo and repetition times, spatial and temporal resolution); and hardware platform.
We recorded the true positive, false negative, false positive, and true negative CMRA assessments for detection of CAD using X-ray angiography as the reference standard. The 50% diameter stenosis cutoff was selected as the criterion for significant CAD, unless only a different threshold had been used. Data were recorded separately, whenever available, at the level of segments, vessels, and subjects. Separate data were also recorded at the segment level for individual major epicardial vessels (left main, left anterior descending, left circumflex, and right coronary arteries). All diagnostic accuracy estimates pertain to evaluable segments/vessels/subjects. Detailed data are available upon request.
Two investigators (A.R., P.G.D.) performed the data extraction independently. Discrepancies were solved by consensus and with discussion with the third investigator (J.P.A.I.). Databases were enhanced with additional data provided directly by primary investigators of several studies.
The main analysis was performed at the level of coronary artery segments, as most studies focused on this level of information. Secondary analyses combined the available vessel-level data and subject-level data. Separate segment-level analyses were performed also for each coronary vessel and for prespecified subgroups, defined so as to investigate the potential effect of key characteristics of the CMRA technique (acquisition technique [2D vs. 3D] and respiratory motion compensation technique), temporal evolution (year of publication [1991 to 2000 vs. 2001 to 2004]), study population characteristics (inclusion or not of subjects without CAD), potential for bias (reading of CMRA stated or not to be blind to the results of X-ray angiography and to clinical data; small vs. larger studies), and segments considered (including or not distal segments).
Sensitivity and specificity estimates for each analysis were independently combined across studies using a random effects model that weighs each proportion by the inverse of the sum of its variance and the between-study variance (46). Between-study heterogeneity was assessed for the sensitivity and specificity estimates using the chi-square test with n − 1 degrees of freedom (n is the number of studies) and Fisher exact test when numbers were small. Because sensitivity and specificity are interdependent, independent weighting may sometimes give spurious results and underestimate both parameters. Therefore, we also estimated weighted and unweighted summary receiver-operating characteristic (SROC) curves (47) that have been established as the most appropriate approach to bypass this problem of interdependence. The SROC curves are estimated from D = a + bS, where D is the difference of the logits (log odds) of the true positives (sensitivity) and false positives (1 − specificity) and S is the sum of these logits. When b is 0, the SROC curve is symmetric around the diagonal. Conversely, when b is significantly different from 0, the SROC curve is not symmetrical, and this suggests that the overall diagnostic performance varies in different parts of the curve, with unequal tradeoff between sensitivity and specificity across studies. This indicates significant heterogeneity across studies in the thresholds used, the study populations, or other parameters. Summary receiver-operating characteristic curves should not be extrapolated outside the range of observed values. Unweighted SROC curves consider all studies equally in the calculations, whereas weighted SROC curves weigh each study by the variance of D.
Finally, for the derived estimates of sensitivity and specificity for CAD at the subject level, we evaluated the negative and positive predictive value among subjects with 5%, 20%, 50%, and 80% pretest probability of CAD. The positive predictive value is calculated as pretest odds × (sensitivity/[1 − specificity])/(1 + [pretest odds × (sensitivity/[1 − specificity])]). The negative predictive value is calculated as pretest odds × ([1 − sensitivity]/specificity)/(1 + [pretest odds × ([1 − sensitivity]/specificity]).
Analyses were performed in SPSS 11.0 (SPSS Inc., Chicago, Illinois), Meta-Analyst (Joseph Lau, Boston, Massachusetts), Meta-Test (Joseph Lau) and StatXact 3.0 (Cytel Inc., Boston, Massachusetts). All p values are two-tailed.
We identified 41 potentially eligible and apparently non-overlapping studies (3–22,24–34,36–45) with 44 different comparisons of CMRA against X-ray angiography. Supplemental data were provided by primary investigators for 13 studies (14 comparisons). For two studies (5,7), we could not obtain sufficient information for constructing any complete 2-by-2 table for any analysis. The remaining 39 studies described 41 comparisons of the two imaging techniques in a total of 1,671 subjects with X-ray angiography, of whom 1,522 also had CMRA data (91.1%). Twelve studies (12 comparisons) with 370 subjects included only subjects with CAD and these were not considered in subject-level analyses because no specificity could be estimated in these studies. Of the remaining 29 comparisons (1,152 subjects), 8 did not specify how many subjects had CAD, whereas in 20 of the other 21 comparisons, patients with CAD made up half or more of the study population. Study population characteristics are shown in Table 1and technical characteristics of the CMRA are shown in Table 2.All studies used gradient echo (bright blood) sequences and only four studies (24,26,34,45) used intravenous contrast enhancement. Data acquisition was consistently gated to the electrocardiographicsignal and timed to diastole, and nulling of the epicardial fat signal was employed.
Data on diagnostic accuracy were available for 25 studies (27 comparisons) (4,620 segments of 993 subjects) at the segment level, 16 studies (2,041 vessels of 624 subjects) at the vessel level, and 13 studies (n = 607) at the subject level. The vast majority of studies defined significant CAD by using the 50% stenosis cutoff. However, two studies (34,44) used 70% as the cutoff for CAD; one other study (17) included patients with >30% stenoses, and all patients had >40% vessel diameter stenosis by conventional X-ray coronary angiography.
There was significant between-study heterogeneity in the sensitivity and specificity estimates in all analyses (Table 3),with the exception of the sensitivity estimates at the subject level that consistently showed high sensitivity (range 73% to 100%), whereas specificity varied widely across studies. Weighted random effects independent estimates fell very close to the weighted SROC curve (Fig. 1),suggesting that they are appropriate approximations of the overall diagnostic performance. Unweighted SROC curves were somewhat more favorable than weighted SROC curves, suggesting that smaller studies tended to show slightly more favorable diagnostic performance than larger studies. The SROC curve for the segment-level analysis showed statistically significant asymmetry: across studies, specificity diminished without any gains in sensitivity beyond 85% to 90%. The overall weighted estimate suggested a sensitivity of 73% with specificity of 86%. However, there seemed to be clusters of studies: one with low sensitivity (<70%) and high specificity (>85%), another with high sensitivity (>80%) and also high specificity (>85%), and a third one with variable sensitivity (60% to 92%) and low specificity (50% to 75%).
At the segment level, the diagnostic accuracy was relatively similar for the left main, left anterior descending, and right coronary arteries, with 69% to 79% of the lesions being detected, against a specificity of 81% to 91%. The sensitivity was considerably lower for lesions in the left circumflex coronary artery with only slightly over half the lesions being detected, and without the specificity being any better (Table 3, Fig. 2).There was modest heterogeneity in sensitivity estimates across studies. Heterogeneity reached formal statistical significance for the left anterior descending and the right coronary arteries, but not for the left main and left circumflex coronary arteries, possibly due to smaller numbers. There was significant between-study heterogeneity across studies for all specificity estimates. In the case of the left main coronary artery this heterogeneity was totally explained by two studies (18,28) that paradoxically showed very low specificity (27%). These two studies, along with another report (4) that also showed <50% specificity for the left anterior descending coronary artery, accounted for the observed between-study heterogeneity for this vessel's segments as well. Absolute differences in specificity were less pronounced between studies for the right and left circumflex coronary arteries. Random effects weighted estimates fell close to the weighted SROC curve, whereas unweighted SROC curves provided somewhat more favorable results. There was no statistically significant asymmetry for any SROC curve in these analyses.
Subgroup analyses at the segment level showed no evidence that the diagnostic performance might differ statistically significantly across subgroups (p > 0.05 for all between-subgroup heterogeneity in both sensitivity and specificity random effects estimates) (Table 4).Nevertheless, there were trends suggesting somewhat paradoxically better diagnostic performance in earlier studies, 2D studies, and studies using breath holding for respiratory motion compensation.
We estimated that for subjects with 5%, 20%, 50%, and 80% pretest probability of CAD, a positive CMRA would slightly increase the probability of CAD to 10%, 33%, 66%, and 89%, respectively. Given the same pretest probabilities, a negative test would decrease the risk of CAD to 1.1%, 5%, 18%, and 46%, respectively.
Coronary magnetic resonance angiography has been a promising alternative for invasive X-ray contrast angiography and has been intensely investigated over the past several years. Our meta-analysis combined data on the diagnostic performance of CMRA that represent the experience of many centers around the world. We showed that in subjects with intermediate/high likelihood for CAD, CMRA can detect about three-quarters of significant stenoses in the visualized segments of the major epicardial coronary arteries with a concomitant specificity of 86%. Coronary magnetic resonance angiography has very good diagnostic performance in all vessels except the left circumflex coronary artery (possibly due to the close relationship with the accompanying vein and adjacent blood pools of the left atrium and ventricle, and lower signal related to the greater distance from the receiving coil). Across populations with generally high prevalence of coronary stenoses, CMRA showed overall high sensitivity for identifying those with CAD, but gave false positive readings in about half of the healthy subjects. Thus, CMRA may be particularly useful when it is negative for decreasing the probability of CAD below 5%, especially in individuals with modest suspicion for CAD, for example, a pretest probability for CAD below 20%.
One should recognize that the visual evaluation of conventional angiography may not be a perfect reference standard (48). Therefore, the diagnostic performance of CMRA may have been slightly underestimated. Conversely, other biases such as reporting bias may have led to overestimation of CMRA performance.
The technical aspects of CMRA are quickly evolving. For example, 3D imaging is now routinely performed by most centers. Nevertheless, we did not observe any statistically significant differences in the diagnostic performance of earlier studies using less advanced techniques, as compared with more recent studies and more advanced technologies. We should acknowledge that the meta-analysis was not powered to detect modest subgroup differences and such differences might have been missed. Alternatively, these parameters may not be very important for the overall diagnostic accuracy of CMRA. If anything, we noted non-statistically significant trends for better diagnostic performance in earlier studies with less advanced techniques. If this is not a chance finding, it may reflect intangible differences in the subject selection process, center expertise and performance, and changes in the criteria used to characterize CMRA as “evaluable” or not. The single multicenter study reported to date demonstrated 93% sensitivity and 42% specificity at the subject level (29). These values reflect a higher sensitivity and lower specificity than the main meta-analysis estimates, but the tradeoff is of fairly similar magnitude. The anticipated negative predictive values based on the results of the multicenter study would be slightly more optimistic than those of the meta-analysis at pretest CAD probabilities of 20% (4% vs. 5%) or 50% (14.3% vs. 18%), but the differences are not large. The meta-analysis demonstrates in a much larger sample size (approximately a log scale higher than that of the multicenter study) that CMRA holds clinical promise across a wider range of diverse clinical centers.
The published reports on the diagnostic performance of CMRA are quite heterogeneous regarding study design and analytic methodologies. For example, evaluation of CAD has been performed in different ways (subject level vs. vessel level vs. segment level). However, heterogeneity is not necessarily a limitation in meta-analysis (49), and it provides a key opportunity to show the consistent performance of the method. We should also acknowledge that not all reports provided complete data. Nevertheless, we contacted several investigators to obtain clarifications or additional data that enhanced our analyses. More rigorous reporting of future clinical research on coronary artery imaging technologies should be encouraged. Our study attempted to address issues pertaining to different analytic approaches by performing separate analyses for each level. These results are largely consistent and show the anticipated increase in sensitivity and decrease in specificity as one moves from segments to vessels to subjects.
The vast majority of the included studies did not evaluate distal segments of the coronary vessels. As CMRA cannot yet visualize well the distal coronaries and branch vessels, its diagnostic performance may be lower for lesions in these segments. Until CMRA improves technically to the point where its performance for distal stenoses can be reliably evaluated, the clinical use of CMRA should only be restricted to proximal vessels. This may be adequate for many patients, as the decision regarding surgical therapy of CAD frequently relies on the presence of proximal stenoses. Even for the proximal coronary arteries, however, current technology only allows for evaluation of roughly 80% of the imaged segments. This limitation, also shared by other noninvasive approaches (32,50–52), is an important consideration for the widespread use of these approaches as alternatives to conventional angiography.
Our findings should be used cautiously when attempting to support the clinical utility of the method. Large intra-center differences likely make the technique still very much center-dependent and, therefore, larger experience is still required. Though various magnetic resonance imaging approaches have been described for coronary artery imaging, current published experience only involves gradient echo (“bright blood”) techniques. With this methodology, normally flowing blood appears bright and stenoses/occlusions appear dark. Such techniques have known pitfalls. For example, coronary artery calcifications may cause susceptibility artifacts and present as signal voids, erroneously suggesting the presence of significant stenoses. Other approaches (black blood imaging, spiral imaging, enhanced sequences, high field imaging) (53–55) may offer additional value. Imaging of the coronary vessel wall and atherosclerotic plaque (56–58) may also enhance our ability to understand the pathophysiology of atherosclerosis and impact on its course, to improve patients' outcomes. The utility of such approaches remains to be proven. The merits of CMRA would also have to be evaluated against other competing noninvasive coronary artery imaging technologies, and in particular computed tomography-based approaches, such as electron-beam and multislice spiral computed tomography (32,50–52,59,60).
In conclusion, across many studies CMRA has shown moderately high sensitivity and may have satisfactory negative predictive value for excluding significant major epicardial coronary artery stenoses in subjects with suspected CAD. Coronary magnetic resonance angiography may have value for exclusion of significant multivessel CAD in subjects referred for diagnostic catheterization, but the available data do not suffice to introduce CMRA as a widely applied screening tool, particularly for individuals with low likelihood of CAD.
The authors wish to thank Drs. Andre Duerinckx, Armin Huber, Aki Ikonen, Warren Manning, Konstantin Nikolaou, Dudley Pennel, Sven Plein, Franco Sardanelli, Robert vanGeuns, Cristoph Weber, and Pamela Woodard for responding to the request for additional data/clarifications and/or for offering their comments regarding this meta-analysis.
- Abbreviations and acronyms
- coronary artery disease
- coronary magnetic resonance angiography
- summary receiver-operating characteristic
- Received September 21, 2003.
- Revision received July 26, 2004.
- Accepted July 29, 2004.
- American College of Cardiology Foundation
- ↵American Heart Association. 2004 Heart and Stroke Statistical Update. Dallas, TX: 2004.
- Pennell D.J.,
- Bogren H.G.,
- Keegan J.,
- Firmin D.N.,
- Underwood S.R.
- Post J.C.,
- van Rossum A.C.,
- Hofman M.B.,
- de Cock C.C.,
- Valk J.,
- Visser C.A.
- Nitatori T.,
- Yokoyama K.,
- Hachiya J.,
- Seki T.,
- Yoshino H.,
- Ishikawa K.
- Sandstede J.,
- Pabst T.,
- Kenn W.,
- Beer M.,
- Neubauer S.,
- Hahn D.
- van Geuns R.J.,
- de Bruin H.G.,
- Rensing B.J,
- et al.
- Regenfus M.,
- Ropers D.,
- Achenbach S,
- et al.
- Ikonen A.E.,
- Manninen H.I.,
- Vainio P,
- et al.
- Klein C.,
- Schalla S.,
- Schnackenburg B,
- et al.
- Fleiss J.L.
- Littenberg B.,
- Moses L.E.
- Leber A.W.,
- Knez A.,
- Becker C,
- et al.
- Kuettner A.,
- Kopp A.F.,
- Schroeder S,
- et al.
- Mollet N.R.,
- Cademartiri F.,
- Nieman K,
- et al.
- Botnar R.M.,
- Stuber M.,
- Kissinger K.V.,
- Kim W.Y.,
- Spuentrup E.,
- Manning W.J.
- Fayad Z.A.,
- Fuster V.,
- Fallon J.T,
- et al.
- Budoff M.J.,
- Achenbach S.,
- Duerinckx A.
- Ropers D.,
- Baum U.,
- Pohle K,
- et al.