Chris Schwemmer

A machine-learning approach for computation of fractional flow reserve from coronary computed tomography

Fractional flow reserve (FFR) is a functional index quantifying the severity of coronary artery lesions and is clinically obtained using an invasive, catheter-based measurement. Recently, physics-based models have shown great promise in being able to noninvasively estimate FFR from patient-specific anatomical information, e.g., obtained from computed tomography scans of the heart and the coronary arteries. However, these models have high computational demand, limiting their clinical adoption. In this paper, we present a machine-learning-based model for predicting FFR as an alternative to physics-based approaches. The model is trained on a large database of synthetically generated coronary anatomies, where the target values are computed using the physics-based model. The trained model predicts FFR at each point along the centerline of the coronary tree, and its performance was assessed by comparing the predictions against physics-based computations and against invasively measured FFR for 87 patients and 125 lesions in total. Correlation between machine-learning and physics-based predictions was excellent (0.9994, P < 0.001), and no systematic bias was found in Bland-Altman analysis: mean difference was -0.00081 ± 0.0039. Invasive FFR ≤ 0.80 was found in 38 lesions out of 125 and was predicted by the machine-learning algorithm with a sensitivity of 81.6%, a specificity of 83.9%, and an accuracy of 83.2%. The correlation was 0.729 (P < 0.001). Compared with the physics-based computation, average execution time was reduced by more than 80 times, leading to near real-time assessment of FFR. Average execution time went down from 196.3 ± 78.5 s for the CFD model to ∼2.4 ± 0.44 s for the machine-learning model on a workstation with 3.4-GHz Intel i7 8-core processor.

Diagnostic Accuracy of a Machine-Learning Approach to Coronary Computed Tomographic Angiography–Based Fractional Flow Reserve: Result From the MACHINE Consortium

Background: Coronary computed tomographic angiography (CTA) is a reliable modality to detect coronary artery disease. However, CTA generally overestimates stenosis severity compared with invasive angiography, and angiographic stenosis does not necessarily imply hemodynamic relevance when fractional flow reserve (FFR) is used as reference. CTA-based FFR (CT-FFR), using computational fluid dynamics (CFD), improves the correlation with invasive FFR results but is computationally demanding. More recently, a new machine-learning (ML) CT-FFR algorithm has been developed based on a deep learning model, which can be performed on a regular workstation. In this large multicenter cohort, the diagnostic performance ML-based CT-FFR was compared with CTA and CFD-based CT-FFR for detection of functionally obstructive coronary artery disease. Methods and Results: At 5 centers in Europe, Asia, and the United States, 351 patients, including 525 vessels with invasive FFR comparison, were included. ML-based and CFD-based CT-FFR were performed on the CTA data, and diagnostic performance was evaluated using invasive FFR as reference. Correlation between ML-based and CFD-based CT-FFR was excellent (R=0.997). ML-based (area under curve, 0.84) and CFD-based CT-FFR (0.84) outperformed visual CTA (0.69; P<0.0001). On a per-vessel basis, diagnostic accuracy improved from 58% (95% confidence interval, 54%–63%) by CTA to 78% (75%–82%) by ML-based CT-FFR. The per-patient accuracy improved from 71% (66%–76%) by CTA to 85% (81%–89%) by adding ML-based CT-FFR as 62 of 85 (73%) false-positive CTA results could be correctly reclassified by adding ML-based CT-FFR. Conclusions: On-site CT-FFR based on ML improves the performance of CTA by correctly reclassifying hemodynamically nonsignificant stenosis and performs equally well as CFD-based CT-FFR.

Coronary CT Angiography–derived Fractional Flow Reserve: Machine Learning Algorithm versus Computational Fluid Dynamics Modeling

Purpose To compare two technical approaches for determination of coronary computed tomography (CT) angiography-derived fractional flow reserve (FFR)-FFR derived from coronary CT angiography based on computational fluid dynamics (hereafter, FFRCFD) and FFR derived from coronary CT angiography based on machine learning algorithm (hereafter, FFRML)-against coronary CT angiography and quantitative coronary angiography (QCA). Materials and Methods A total of 85 patients (mean age, 62 years ± 11 [standard deviation]; 62% men) who had undergone coronary CT angiography followed by invasive FFR were included in this single-center retrospective study. FFR values were derived on-site from coronary CT angiography data sets by using both FFRCFD and FFRML. The performance of both techniques for detecting lesion-specific ischemia was compared against visual stenosis grading at coronary CT angiography, QCA, and invasive FFR as the reference standard. Results On a per-lesion and per-patient level, FFRML showed a sensitivity of 79% and 90% and a specificity of 94% and 95%, respectively, for detecting lesion-specific ischemia. Meanwhile, FFRCFD resulted in a sensitivity of 79% and 89% and a specificity of 93% and 93%, respectively, on a per-lesion and per-patient basis (P = .86 and P = .92). On a per-lesion level, the area under the receiver operating characteristics curve (AUC) of 0.89 for FFRML and 0.89 for FFRCFD showed significantly higher discriminatory power for detecting lesion-specific ischemia compared with that of coronary CT angiography (AUC, 0.61) and QCA (AUC, 0.69) (all P < .0001). Also, on a per-patient level, FFRML (AUC, 0.91) and FFRCFD (AUC, 0.91) performed significantly better than did coronary CT angiography (AUC, 0.65) and QCA (AUC, 0.68) (all P < .0001). Processing time for FFRML was significantly shorter compared with that of FFRCFD (40.5 minutes ± 6.3 vs 43.4 minutes ± 7.1; P = .042). Conclusion The FFRML algorithm performs equally in detecting lesion-specific ischemia when compared with the FFRCFD approach. Both methods outperform accuracy of coronary CT angiography and QCA in the detection of flow-limiting stenosis.

Comparison of invasively measured FFR with FFR derived from coronary CT angiography for detection of lesion-specific ischemia: Results from a PC-based prototype algorithm

BACKGROUND We evaluated the diagnostic accuracy of a novel prototype for on-site determination of CT-based FFR (cFFR) on a standard personal computer (PC) compared to invasively measured FFR in patients with suspected coronary artery disease. METHODS A total of 91 vessels in 71 patients (mean age 65 ± 9 years) in whom coronary CT angiography had been performed due to suspicion of coronary artery disease, and who subsequently underwent invasive coronary angiography with FFR measurement were analyzed. For both cFFR and FFR, a threshold of ≤0.80 was used to indicate a hemodynamically relevant stenosis. The mean time needed to calculate cFFR was 12.4 ± 3.4 min. A very close correlation between cFFR and FFR could be shown (r = 0.85; p < 0.0001) with Bland-Altman analysis showing moderate agreement between FFR and cFFR with mild systematic overestimation of FFR values in CT (mean difference 0.0049, 95% limits of agreement ±2SD -0.007 to 0.008). Compared to FFR, the sensitivity of cFFR to detect hemodynamically significant lesions was 91% (19/21, 95% CI: 70%-99%), specificity was 96% (67/70, 95% CI: 88%-99%), positive predictive value 86% (95% CI: 65%-97%) and negative predictive value was 97% (95% CI: 90%-100%) with an accuracy of 93%. CONCLUSION cFFR obtained using an on-site algorithm implemented on a standard PC shows high diagnostic accuracy to detect lesions causing ischemia as compared to FFR. Importantly, the time needed for analysis is short which may be useful for improving clinical workflow.

Growing a Random Forest with Fuzzy Spatial Features for Fully Automatic Artery-Specific Coronary Calcium Scoring

The amount of coronary artery calcium (CAC) is a strong and independent predictor of coronary heart disease (CHD). The standard routine for CAC quantification is to perform non-contrasted coronary computed-tomography (CCT) on a patient and present the resulting image to an expert, who then uses this to label CAC in a tedious and time-consuming process. To improve this situation, we present an automatic CAC labeling system with high clinical practicability. In contrast to many other automatic calcium scoring systems, it does not require additional cardiac computed tomography angiography (CCTA) data for artery-specific labeling. Instead, an atlas-based feature approach in combination with a random forest (RF) classifier is used to incorporate fuzzy spatial knowledge from offline data. Overall detection of CAC volume on a test set with 40 patients yields an \(F_1\) score of 0.95 and 1.00 accuracy for risk class assignment. The intraclass correlation coefficient is 0.98 for the left anterior descending artery (LAD), 0.88 for the left circumflex artery (LCX), and 0.98 for the right coronary artery (RCA). The implemented system offers state-of-the-art accuracy with a processing time (< 30 s) by magnitudes lower than comparable systems to be found in the literature.