Machine-Learning CT-FFR and Extensive Coronary Calcium

Coronary computed tomography angiography (CTA) is a reliable modality for noninvasive assessment of coronary stenosis (1). Recent advances in computational fluid dynamics (CFD) have enabled patient-specific modeling of coronary blood flow and pressure from standard CTA images, without the need for additional imaging or medication (2). Based on this technique, noninvasive fractional flow reserve (FFR) derived from CTA can identify lesion-specific ischemia by invasive coronary angiography (ICA)-derived FFR, as shown in several clinical trials (3–5). Nevertheless, CFD models are complex; involve multiple finely calibrated geometric, hemodynamic, and material parameters; and may require long computation times (6).

To address this limitation, Itu et al. (7) developed a machine-learning (ML)-based approach for calculating CT-FFR on-site at standard workstations. The ML algorithm was trained using a database of 12,000 synthetically generated 3-dimensional coronary anatomies for which the CFD-based CT-FFR values were computed. The CFD-based results formed the ground truth training data for the ML-based CT-FFR model. A deep neural network with 28 geometric input features extracted from the synthetically generated database and 4 hidden layers was used to train the ML-based model. The ML application was then tested on CTA images from 87 patients and demonstrated high diagnostic accuracy compared to invasive FFR. The average computation time was 2.4 ± 0.44 s for the entire coronary tree, thus providing near real-time assessment of FFR (7).

Coenen et al. (8) compared the diagnostic performance of this ML-based approach against CFD-based CT-FFR in MACHINE (Machine Learning Based CT Angiography Derived FFR: A Multi-Center Registry) (comprising 351 patients and 525 vessels), with invasive FFR as the reference standard. ML-based (area under curve [AUC]: 0.84) and CFD-based CT-FFR (0.84) outperformed visual CTA assessment (0.69; p < 0.001), with no significant difference between the two. The diagnostic accuracy per vessel improved from 58% (54% to 63%) for CTA to 78% (75% to 82%) by adding ML-based CT-FFR.

Despite technical advances in CTA resulting in improved temporal and spatial resolution (9), extensive coronary artery calcium (CAC) remains perhaps its greatest challenge in the assessment of coronary artery disease (CAD) (10). The presence of extensive coronary calcification can often lead to overestimation of stenosis severity (11). A meta-analysis of 1,634 patients from 19 studies that examined the influence of CAC on the diagnostic performance of CTA showed a significant decline in the specificity of CTA to 42% (28% to 56%) at CAC scores >400 (12). Extensive CAC also influences accurate segmentation of the vessel lumen, an important component of noninvasive FFR computation. Notwithstanding, a substudy of the NXT (Analysis of Coronary Blood Flow Using CT Angiography: Next Steps) trial by Norgaard et al. (5) found no differences in diagnostic accuracy, sensitivity, or specificity of CFD-based CT-FFR across study-specific quartiles of CAC scoring, including patients at even the highest quartile with CAC scores ranging from 416 to 3,599. Similarly, a post hoc analysis of 42 patients with calcium-related artifacts from the DISCOVER-FLOW (Diagnosis of Ischemia-Causing Stenoses Obtained Via Noninvasive Fractional Flow Reserve) trial showed the superior accuracy of CFD-based CT-FFR over CTA (86% vs. 67%), predominantly due to an increase in specificity (82% vs. 29%) (13).

In this issue of iJACC, Tesche et al. (14) report on the impact of CAC on the diagnostic performance of ML-based CT-FFR in a substudy of 314 patients (and 482 vessels) from the MACHINE registry who had available CAC scores from noncontrast CT. Using ML CT-FFR, the authors generated patient-specific 3-dimensional meshes that allowed for determination of the CT-FFR value at any point within the coronary tree. With invasive FFR as reference, 228 vessels in 200 patients had functionally significant coronary stenoses (FFR ≤ 0.80). On a per-vessel and per-patient level, the accuracy, sensitivity, and specificity of CT-FFR to detect lesion-specific ischemia were not significantly different across CAC score categories. Per vessel, CT-FFR showed incremental discriminatory power over CTA at both high CAC scores (CAC ≥ 400: AUC 0.71 vs. 0.55; p = 0.04) and low-to-intermediate CAC scores (CAC > 0 to < 400: AUC 0.86 vs. 0.63; p < 0.001). The performance of CT-FFR declined as CAC score increased, just reaching statistical significance (AUC 0.85 [CAC > 0 to < 400] vs. 0.71 [CAC ≥ 400]; p = 0.04).

These findings demonstrate the high and superior diagnostic performance of ML-based CT-FFR compared to CTA across a wide range of CAC scores in a large patient cohort, lending further support to the potential role of CT-FFR as a gatekeeper to ICA and coronary revascularization.

Although currently a research prototype, the ML-based application is highly promising for the delivery of on-site, real-time FFR assessment. Results derived using the ML model, based on geometric features extracted from CTA, are nearly identical to CFD results, with a >80-fold reduction in computation time compared to CFD. Theoretically, the ML-based FFR calculation may be less influenced by CAC compared to CFD modeling, as the ML algorithm incorporates both local and global geometric data, with only 8 of the 28 input features corresponding to luminal stenosis (7). Moreover, the performance of the ML model likely will improve as advancements are made in the CFD model of coronary arterial flow on which it is based, including refined segmentation algorithms with larger cohorts of real-world CTA data.

In recent years, ML in cardiovascular imaging has emerged as a powerful tool for risk prediction and decision making (15). An ML model that combined clinical and anatomic CTA data was shown to predict 5-year all-cause mortality better than existing clinical or CTA metrics alone (16). Furthermore, ML-based CT-FFR in the original MACHINE registry study reclassified 62 of 85 (73%) false-positive CTA results (8), which has important implications for risk stratification and subsequent referral to ICA of patients with suspected CAD.

As with all developing artificial intelligence technologies, ML-based CT-FFR has limitations. The study by Tesche et al. (14) revealed a reduction in diagnostic performance of the ML model in patients with CAC ≥ 400, which also was reflected by the poor correlation between per-vessel ML-based CT-FFR and invasive FFR in this subgroup (r = 0.35) (14). Certainly, clear thresholds for performing ML-based CT-FFR in the setting of extensive CAC need to be defined before its application in a broad patient population. The extraction of anatomic features from CTA for the ML model also requires considerable time (up to 60 minutes) (7) and remains the rate-limiting step for on-site CT-FFR at point of care.

The present study represents a post hoc analysis of the MACHINE registry data, which were mostly retrospective and included variations in patient selection and study design between centers. In this real-world cohort with a broad range of CAC scores, ML-based CT-FFR improved identification of lesion-specific ischemia even at high CAC scores, with a reduction in performance of CT-FFR with an increase in CAC. Despite the study limitations, Tesche et al. (14) should be commended for their timely work demonstrating the robust diagnostic performance of ML-based CT-FFR in the setting of extensive CAC, the current Achilles heel of coronary CTA.

Finally, in a study that randomly assigned patients with suspected CAD to standard or high-resolution CTA, improved CT spatial resolution significantly reduced overestimation of stenosis severity in the setting of calcified plaques and resulted in higher diagnostic accuracy of CTA in comparison to ICA (17). With ongoing advancements in CT technology, the accuracy of coronary tree segmentation and extraction of quantitative measures from CTA are only likely to improve.