Machine Learning CT FFR: The Evolving Role of On-Site Techniques

Published Online:https://doi.org/10.1148/ryct.2020200228

See also the article by Martin et al in this issue.

Abdul Rahman Ihdayhid, MBBS, PhD, trained in interventional cardiology and                     cardiac CT at MonashHeart, Melbourne, Australia. He completed his PhD in cardiac                     CT and intracoronary physiology through Monash University. Dr Ihdayhid’s                     research is focused on investigating novel techniques to determine coronary                     blood flow utilizing cardiac CT.

Abdul Rahman Ihdayhid, MBBS, PhD, trained in interventional cardiology and cardiac CT at MonashHeart, Melbourne, Australia. He completed his PhD in cardiac CT and intracoronary physiology through Monash University. Dr Ihdayhid’s research is focused on investigating novel techniques to determine coronary blood flow utilizing cardiac CT.

Sagit Ben Zekry, MD, is an echocardiographist at Leviev Heart Center,                     Sheba Medical Center, Israel, and a graduate of an echocardiography fellowship                     at Methodist Hospital, Houston, Texas. Dr Ben Zekry’s research is focused                     on valvular disease and three-dimensional echocardiography.

Sagit Ben Zekry, MD, is an echocardiographist at Leviev Heart Center, Sheba Medical Center, Israel, and a graduate of an echocardiography fellowship at Methodist Hospital, Houston, Texas. Dr Ben Zekry’s research is focused on valvular disease and three-dimensional echocardiography.

The role of coronary CT angiography in the evaluation of low to intermediate risk acute chest pain in the emergency department (ED) is well established (1). As an anatomic test, coronary CT angiography provides excellent sensitivity at ruling out coronary artery disease (CAD) and safely helps triage patients to discharge from the ED. A further strength of coronary CT angiography is its ability to mitigate future cardiovascular risk by identifying subclinical atherosclerosis and therefore the potential to initiate preventative medical therapy (2). However, anatomic assessment alone has recognized limitations. Coronary CT angiography tends to overestimate angiographic severity of coronary lesions and can result in inappropriate referral to invasive angiography (1,3). Furthermore, anatomic assessment has limited specificity for identifying hemodynamically significant disease (4). Therefore, in the presence of intermediate stenosis, patients often require further testing. Taken together, given the higher prevalence of CAD in patients with acute chest pain and the need for rapid risk stratification in the ED, noninvasive techniques that assess both anatomic and hemodynamic significance warrant consideration.

Clinical trials have established the role of intracoronary physiology using invasive fractional flow reserve (FFR) to augment anatomic assessment of CAD. The past decade has seen the application of computational fluid dynamic (CFD) techniques to noninvasively derive FFR from coronary CT angiography (CT FFR). The sole commercially available technique (FFRCT) requires remote analysis and is based on three-dimensional modeling of coronary blood flow (5). There is well-established evidence supporting its diagnostic accuracy, prognostic value, and clinical utility in real-world practice, including in patients with acute chest pain (57). Reduced-order CT FFR techniques have been developed that require less intense computational analysis and hence can be derived on-site using a standard desktop computer (8,9). Although evidence supporting their clinical use is less than that for FFRCT, these techniques have the appeal of potentially shorter analysis time, convenience, and reduced costs. In particular, a recent study has demonstrated feasibility of using artificial intelligence (AI) algorithms to reduce the computational time of CT FFR to a matter of seconds while maintaining high diagnostic accuracy (9). The use of an on-site AI-based CT FFR technique in the ED is unknown and forms the basis of this study by Martin et al (15) in this issue of Radiology: Cardiothoracic Imaging.

This single-center retrospective study evaluated the potential for an on-site AI-based CT FFR technique to identify clinical outcomes compared with coronary CT angiography in patients presenting to an ED with acute chest pain. From 271 eligible patients, 159 patients (59%) were included, with the majority of exclusions due to inadequate image quality for CT FFR analysis. The study reported that at 1-year follow-up, a positive CT FFR result (≤ 0.80) was a superior predictor compared with obstructive disease at coronary CT angiography (≥ 50%) for the combined composite endpoint of revascularization not driven by index coronary CT angiography, myocardial infarction, unstable angina, and all-cause death (odds ratio, 3.4 vs 2.2; P < .01). The findings were driven by differences in revascularization. Moreover, the agreement with subsequent downstream noninvasive testing (SPECT, stress echocardiography, stress MRI) was higher with CT FFR than coronary CT angiography, although the observation was not statistically significant (P = .23).

The authors are to be commended for advancing the field of on-site CT FFR and applying this technique to new patient cohorts. Although this is an exploratory analysis, there are important issues with regard to the study methodology and results that warrant further discussion. First, the proportion of scans deemed acceptable for on-site CT FFR analysis was low (59%). Only three patients (1.9%) received sublingual nitroglycerin and 13 patients received β-blockers (8.2%). Heart rate control is an essential step to minimize motion artifact, and the absence of nitrate premedication can limit coronary luminal characterization. This can significantly affect image quality and therefore the accuracy of both anatomic and hemodynamic assessment at coronary CT angiography and CT FFR analysis (10). Recent prospective FFRCT studies in both the ambulatory and ED setting demonstrated that routine use of nitrates and rate-controlling medications reduced nonevaluable coronary CT angiography scans to 1%–3% (6,7).

Second, it is important to describe where on the anatomic model CT FFR was measured. As a clinical tool, the location at which CT FFR values should be read is unclear. In studies thus far, the diagnostic and prognostic value of CT FFR was assessed by taking the most distal value (6). It has been observed that FFR simulation at coronary CT angiography is associated with pressure drop-off in the distal vessel in the absence of any measurable coronary stenosis (6). Therefore, taking the most distal measurement can be associated with higher false-positive rate and potentially inappropriate referral for invasive procedures. An alternative strategy is to follow the invasive FFR guidelines and measure CT FFR at 20–30 mm distal to the coronary stenosis (11). This provides better assessment of the hemodynamic impact of a given stenosis and patients most likely to benefit from lesion-specific revascularization (11). This method, however, does not provide an assessment of the “pattern of disease” along the course of the vessel (ie, diffuse disease, tandem lesions), which is increasingly recognized in the invasive literature to have prognostic and therapeutic implications (12). Future techniques that consider other CT-derived metrics such as subtended myocardium may assist in the interpretation of CT FFR values in real-world clinical practice.

The important novelty of this study lies in the use of an on-site CT FFR technique to potentially streamline functional evaluation of CAD in an ED setting. Several single-center and retrospective studies have demonstrated promising preliminary diagnostic results for on-site CT FFR techniques (8,9). Many questions though remain to be answered before we can translate on-site CT FFR techniques from a research tool to one that directs clinical practice. Similar to the process undertaken by FFRCT, the diagnostic performance of these techniques needs to be rigorously assessed in an adequately powered, international, multicenter setting and compared across the spectrum of noninvasive diagnostic testing, in addition to prospective trials assessing clinical utility and prognostic value. Furthermore, CT FFR techniques as a whole need to mature into a technology that informs clinicians on decisions regarding both medical therapy and revascularization. Recently, the application of virtual stent placement techniques using FFRCT has been demonstrated to be feasible and highly accurate at predicting the hemodynamic benefit of percutaneous coronary intervention (13). This has the potential to reduce time and cost and shift the decision making about our most complex patients from the invasive to noninvasive domain.

From a technical perspective, reduced-order CFD modeling of an on-site CT FFR method is associated with analysis times of 10–30 minutes (8,14). The implementation of machine learning is a notable evolution of the technique, which is able to provide computation of CT FFR in essentially “real-time.” However, it should be emphasized that accurate segmentation of the coronary lumen remains an essential step in the derivation of CT FFR (14). The original validation study of this AI-based CT FFR technique reported high diagnostic performance in the setting of operators spending up to 60 minutes on luminal segmentation (9). A recent study illustrated that meticulous attention to luminal segmentation is associated with improved CT FFR results, although at the cost of a fourfold increase in analysis time (14). Therefore, while the reduced segmentation time (17 minutes) in this study may seem appealing, this cannot be commended without understanding the segmentation technique employed and its impact on diagnostic performance. Machine learning applied to data sets of manually segmented coronary geometry will be essential to improve the automation of coronary luminal segmentation and therefore significantly reduce the overall processing time of CT FFR.

Last, in the current study, there was significant discordance between CT FFR and noninvasive stress testing. A positive CT FFR result was associated with a negative SPECT and negative stress echocardiographic result in 43% and 35% of patients, respectively. Furthermore, of the 41 patients referred for invasive coronary angiography on the basis of their CT angiography result, CT FFR was positive in only 54% of cases. These results raise questions on the accuracy of the CT FFR analysis, and together with the small study size, limit drawing positive conclusions on the ability of this technique to guide downstream testing.

In conclusion, this study provides early insight into the potential role of on-site machine learning–based CT FFR in evaluating patients with acute chest pain. The results also emphasize the importance of good quality coronary CT angiography acquisition to facilitate accurate anatomic and hemodynamic assessment and guide appropriate decision making. As on-site CT FFR continues to evolve, automation of coronary luminal segmentation and prospective studies evaluating its effect on patient outcomes will be essential steps to translate this technique into clinical practice.

Disclosures of Conflicts of Interest: A.R.I. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: received consulting honorarium and payment for lectures from Canon Medical. Other relationships: disclosed no relevant relationships. S.B.Z. disclosed no relevant relationships.

References

  • 1. Hoffmann U, Truong QA, Schoenfeld DA, et al. Coronary CT angiography versus standard evaluation in acute chest pain. N Engl J Med 2012;367(4):299–308.
  • 2. Honigberg MC, Lander BS, Baliyan V, et al. Preventive Management of Nonobstructive CAD After Coronary CT Angiography in the Emergency Department. JACC Cardiovasc Imaging 2020;13(2 Pt 1):437–448.
  • 3. Collet C, Onuma Y, Andreini D, et al. Coronary computed tomography angiography for heart team decision-making in multivessel coronary artery disease. Eur Heart J 2018;39(41):3689–3698.
  • 4. Danad I, Raijmakers PG, Driessen RS, et al. Comparison of Coronary CT Angiography, SPECT, PET, and Hybrid Imaging for Diagnosis of Ischemic Heart Disease Determined by Fractional Flow Reserve. JAMA Cardiol 2017;2(10):1100–1107.
  • 5. Nørgaard BL, Leipsic J, Gaur S, et al. Diagnostic performance of noninvasive fractional flow reserve derived from coronary computed tomography angiography in suspected coronary artery disease: the NXT trial (Analysis of Coronary Blood Flow Using CT Angiography: Next Steps). J Am Coll Cardiol 2014;63(12):1145–1155.
  • 6. Fairbairn TA, Nieman K, Akasaka T, et al. Real-world clinical utility and impact on clinical decision-making of coronary computed tomography angiography-derived fractional flow reserve: lessons from the ADVANCE Registry. Eur Heart J 2018;39(41):3701–3711.
  • 7. Chinnaiyan KM, Safian RD, Gallagher ML, et al. Clinical Use of CT-Derived Fractional Flow Reserve in the Emergency Department. JACC Cardiovasc Imaging 2020;13(2 Pt 1):452–461.
  • 8. Ihdayhid AR, Sakaguchi T, Linde JJ, et al. Performance of computed tomography-derived fractional flow reserve using reduced-order modelling and static computed tomography stress myocardial perfusion imaging for detection of haemodynamically significant coronary stenosis. Eur Heart J Cardiovasc Imaging 2018;19(11):1234–1243.
  • 9. Coenen A, Kim YH, Kruk M, et al. Diagnostic Accuracy of a Machine-Learning Approach to Coronary Computed Tomographic Angiography-Based Fractional Flow Reserve: Result From the MACHINE Consortium. Circ Cardiovasc Imaging 2018;11(6):e007217.
  • 10. Pontone G, Weir-McCall JR, Baggiano A, et al. Determinants of Rejection Rate for Coronary CT Angiography Fractional Flow Reserve Analysis. Radiology 2019;292(3):597–605.
  • 11. Cami E, Tagami T, Raff G, et al. Assessment of lesion-specific ischemia using fractional flow reserve (FFR) profiles derived from coronary computed tomography angiography (FFRCT) and invasive pressure measurements (FFRINV): Importance of the site of measurement and implications for patient referral for invasive coronary angiography and percutaneous coronary intervention. J Cardiovasc Comput Tomogr 2018;12(6):480–492.
  • 12. Collet C, Sonck J, Vandeloo B, et al. Measurement of Hyperemic Pullback Pressure Gradients to Characterize Patterns of Coronary Atherosclerosis. J Am Coll Cardiol 2019;74(14):1772–1784.
  • 13. Modi BN, Sankaran S, Kim HJ, et al. Predicting the Physiological Effect of Revascularization in Serially Diseased Coronary Arteries. Circ Cardiovasc Interv 2019;12(2):e007577.
  • 14. Ihdayhid AR, Sakaguchi T, Kerrisk B, et al. Influence of operator expertise and coronary luminal segmentation technique on diagnostic performance, precision and reproducibility of reduced-order CT-derived fractional flow reserve technique. J Cardiovasc Comput Tomogr 2019. 10.1016/j.jcct.2019.11.014. Published online November 26, 2019.
  • 15. Martin SS, Mastrodicasa D, van Assen M, et al. Value of Machine Learning–based Coronary CT Fractional Flow Reserve Applied to Triple-Rule-Out CT Angiography in Acute Chest Pain. Radiol Cardiothorac Imaging 2020;2(3):e190137.

Article History

Received: Apr 17 2020
Revision requested: May 29 2020
Revision received: May 31 2020
Accepted: June 2 2020
Published online: June 25 2020