Starting over 70 years ago,1 clinicians began quantifying myocardial blood flow to diagnose and guide treatment. During the past seven decades, a large number of tools has been developed, both invasive and non-invasive. The ongoing utility of myocardial blood flow assessment testifies the underlying clinical need. Why does this patient have chest pain or dyspnea? Should I perform coronary angiography or how should I interpret its anatomy? Will the patient benefit from revascularization? And, if yes, which vessels require a procedure?

The ideal tool for studying myocardial blood flow possesses the following features. First, it should be non-invasive to avoid the small, but non-zero, risk of complications inherent to any catheterization and therefore facilitate serial assessment and require small or no radiation exposure. Second, minimal contraindications allow its application to the broadest range of patients regardless of body habitus, heart rate and rhythm, implanted devices, coagulopathy, and renal function. Third, the entire left ventricle must be evaluated independent of the complexity of coronary anatomy, including total occlusions, bypass grafts, and tortuous or calcified vessels. Fourth, since absolute flow depends on the amount of supplied muscle, only normalizing to myocardial mass (= perfusion) permits comparisons among patients, although serial exams in the same patient generally do not require this adjustment. Fifth, the technique should be repeatable with a known and acceptable test/retest variation under stable conditions. Finally, the resolution of the tool must permit distinction among the “revascularizable” vessels (generally at least 2 to 2.5 mm in reference diameter) and quantify their distribution territory as a fraction of the total left ventricle.

Since its initial description in animals2 and humans3 roughly 40 years ago for myocardial imaging, cardiac positron emission tomography (PET) has matured into the reference standard since it uniquely fulfills all of these ideal features. However, several software packages exist in current practice and, as admitted by a recent national statement on PET blood flow quantification, “significant variation remains among some vendor programs.”4 The current manuscript by Nesterov et al in this issue5 therefore provides a timely opportunity to consider how to judge existing PET flow software.

SCORECARD FOR FLOW MODELS

Table 1 summarizes our criteria for a scientifically valid and clinically practical flow model. First, the model should have been experimentally validated in animals using a reference standard such as an external flow probe6 or microspheres. Second, the model must be tested along the entire spectrum of flow that exists in nature, from the high levels in normal volunteers7 (distinct from clinical patients with mild risk factors that reduce average flows) to the low levels during ischemia8 (stress-induced perfusion defect with severe angina and/or significant electrocardiographic changes). Note that flow levels for ischemia must be quantified on a regional basis - not globally - given its generally circumscribed pathophysiology. Obviously a flow model should work with all the major pharmacologic stress agents, including dipyridamole,9 adenosine,10 and regadenoson.11

Table 1 Scorecard for judging a PET flow package
Full size table

Essentially every test in medicine displays some imprecision due to a combination of the physical limits of the technology plus biology itself. The net effect for myocardial imaging implies test/retest variation, even if done twice on the same day or a week or two apart under clinically stable circumstances. Short-term variation, as on the same day, largely reflects stochastic effects of the radiotracer and imaging chain amplified by the flow model, whereas long-term variation over a week or two also includes biologic fluctuations. We have demonstrated a 10% coefficient of variation in the short term that grows to 21% in the long term for rubidium-82 and a retention model.9

The arterial input represents a vital yet often underappreciated component of flow quantification. While radiotracer concentration in the arterial circulation does not vary much spatially due to mixing during transit, both cardiac PET12 and magnetic resonance13 perfusion imaging display heterogeneity among potential sampling sites (left atrium, left ventricular cavity, ascending aorta, descending aorta). Reasons for this variation in PET include partial volume effects, spillover from adjacent structures, and motion of intrathoracic anatomy during cardiac and respiratory cycles as well as from gravity during supine positioning, imaging anisotropy, and co-registration of emission and transmission images that focuses on aligning the left ventricular myocardium but not on other regions. Generally these effects reduce the imaged arterial activity and vary greatly among patients, indicating the need for individualized anatomic selection instead of using a fixed, “one-size-fits-none” region of interest. Recovering less than the full arterial input leads to overestimated and hence biased flow.

Results from a flow model must associate with clinical outcomes (as modified by medical treatment versus revascularization) in a large cohort followed for years, displaying a risk continuum with both size and severity.14 Incorporating flow quantification into clinical care and imaging reports should become the daily routine, as it has been at our center since April 2007 - now over 14 years and exceeding 9,200 sequential PET scans. Its application in practice should be demonstrated to be cost-effective15 and validated through randomized clinical trials.16

We have systematically examined each of these items (animal validation, flow spectrum from normal to ischemia, several pharmacologic stress medications, test/retest variability, arterial input customization, clinical outcomes, economic analysis, and randomized trial) in a series of publications mainly over the past decade,6,7,8,9,10,11,12,14,15,16 indicating that judging a PET flow model requires an edifice that cannot be built overnight and does not rest solely on any single aspect. Rather, multifaceted and interwoven evidence supports customized decisions for each patient based on the sound integration of PET flow physiology with unique clinical circumstances.

SAME DATA, DIFFERENT RESULTS?

With this background we can place the new results in context.5 In almost 50 patients with hypertrophic cardiomyopathy, dipyridamole stress N-13 ammonia PET images were obtained on a single scanner at one center in Italy. Reconstructed images were anonymized and shared among three centers (Italy, Finland, and Switzerland), each using a different software package (PMOD, Carimas, and FlowQuant, respectively) but the same one-tissue compartment model. After blinded analysis, global, territory, and segmental stress perfusion were compared among the three tools. Segmental values could vary by as much as 36% as quantified by the maximum/minimum ratio, and pairwise comparison of segments showed several outside of the their “acceptable agreement” range encompassing ± 20% in median value and ≥ 0.75 intraclass correlation coefficient.

Because the reconstructed PET data and time-activity curve model remained constant, the observed variation in the results can only have arisen due to model implementation or to differences in the input to the model (time-activity curves for the arterial input and myocardium uptake) or to a combination of both factors. The authors did not analyze the same time-activity curve in all three software packages, which would have uncovered any variations in implementation. Therefore, their results remain somewhat ambiguous regarding the mechanism(s) for the observed heterogeneity.

Nevertheless, that such frequent and sizable variations in calculated flow could arise from the same underlying data emphasizes the lack of portability among software packages. Consequently, in our opinion each vendor must be judged according to the integrative framework of Table 1. As described earlier, we believe that differences in the arterial input location12 likely account for a sizable portion of the variation in the flows observed in the current study.5 Additionally, many algorithms average over a region of interest in the myocardium to determine its uptake, which neglects basic physical principles of the point spread function and partial volume effects.17 To our knowledge, the partial volume factor remains “hidden in the black box” for almost all software vendors, providing a fudge factor never displayed to the user and hence not comparable among results.

Unexpectedly, the current analysis found the largest variation in quantification of stress perfusion in the lateral territory and segments.5 In normal volunteers the lateral wall has been shown by PET7 and invasive assessment18 to have equivalent perfusion to the other regions. The discrepant results in this new manuscript did not arise due to a perfusion defect, since the lateral territory and segments had the highest average stress perfusion.5 Because each software package started with the same input image, this finding suggests spatial heterogeneity when generating time-activity curves for the left ventricle.

FLOW PACKAGE, NOT FLOW MODEL

In our view, the current manuscript by Nesterov et al5 offers a profound yet unexpected insight: flow software must be judged by more than its underlying physiologic model. Indeed, we should instead speak of a PET flow “package” that explicitly acknowledges many design choices that affect the numeric values: time-activity curve versus retention flow model, peak versus average myocardial uptake, fixed and known versus variable and hidden partial volume correction, and customized versus static arterial input location. Additionally, each flow package should be rated using the report card in Table 1. We do not think of people as interchangeable, although we are all based on the same “model” - the same holds true for PET flow software.