Machine Learning Methods for Assessment of Coronary Artery Disease and Microvascular Disease

Introduction: Coronary Artery Disease (CAD) is an important cause of mortality worldwide, accounting for millions of deaths annually. CAD can be caused by both epicardial CAD and Coronary Microvascular Disease (CMD). Epicardial CAD is caused by a buildup of atherosclerotic plaque in the large coronary arteries, often leading to luminal obstructions termed stenosis. In contrast, CMD is characterized by impaired blood flow and blood regulation in the microcirculation, posing a significant challenge in the field of cardiology. Both conditions disrupt the delicate balance between blood supply and demand, leading to adverse clinical outcomes, i.e. angina, myocardial infarction, and death. The most common method for diagnosing these conditions involves techniques in the cardiac catheterization laboratory that involve angiography and pressure measurement. Fractional Flow Reserve (FFR) and the Index of Microcirculatory Resistance (IMR) serve as gold standard quantitative metrics from a physiological perspective for CAD and CMD, respectively. Despite their ability to provide useful information for clinicians, the invasive nature of these assessments using an intravascular guide wire has led to their underutilization in current clinical practice. Consequently, non-invasive approaches that use computational techniques to estimate FFR and IMR accurately are highly desired. In this study, machine learning-based methods are developed to estimate virtual FFR (vFFR) and equivalent information for the IMR and CMD using dynamic data from angiography, the most commonly used modality for coronary artery disease assessment.

Materials and

Methods: First, an attention-based multi-fidelity machine learning framework is developed for vFFR assessment, comprising two stages: training and testing (Fig. 1). During the training phase, the synthetic vessel generator is used to define two sets of vessel arrays. The first set is used to train the feature extraction autoencoder, which extracts geometric features from the 3D vessel geometry, with additional attention to key regions, such as stenoses. The second set is used to generate corresponding hemodynamic data through both high-fidelity (HF) 3D CFD and low-fidelity (LF) 1D models. Subsequently, the feature fusion U-Net is trained using geometric features, and both the HF and LF hemodynamic data for fidelity mapping. At the patient-specific testing stage, anatomical data and boundary conditions data are fed to AttMulFid to conduct computational vFFR prediction in real time.
Additionally, to provide an interpretation of the parameters that control different scenarios of the contrast injection and washout in the dynamic angiogram, we developed a novel multi-physics CFD model of iodine contrast injection, as shown in Fig. 2. A key aspect of the model is lumped parameter models (LPMs) for the left heart, right coronary arteries, and aortic outflow, which serve as boundary conditions for the CFD simulation. The computation of the transport of the contrast agent within the coronary vessels provides a surrogate of the angiographic data, and the contrast intensity curve (CIC) can be extracted through projection, thresholding, and pixel count, as shown in Fig. 3. CIC therefore provides an assessment of CMD analogous to the IMR.

Results, Conclusions, and Discussions: Fig. 4 (a) illustrates geometry reconstruction results. The reconstructed geometries exhibit a high level of agreement with the ground truth geometries with minimal discrepancies. This suggests that the low-dimensional geometric features are capable of representing the entire geometric information presented in the original data. Fig. 4 (b) presents a gradient-based attention map to provide a visual interpretation of the learning process. The attention is notably larger in the stenosis regions. This observation demonstrates that the extracted features effectively identify stenoses and allocate a greater degree of attention to such regions. Fig. 4 (c) shows comparisons of vFFR predictions for LF, HF (extracted from pressure contour in Fig. 4 (d)), AttMulFid, and invasive FFR, along with epistemic and aleatoric uncertainties. vFFR prediction from AttMulFid presents a good alignment with the FFR predictions of the CFD model and invasive FFR measurement.
To demonstrate the impact of altered microvascular resistances on the contrast injection curves, we conducted a study whereby the parameters governing coronary microvascular resistance were gradually altered to simulate increased degrees of CMD. Fig. 5 presents CICs corresponding to three simulation scenarios: 1) a healthy baseline with a total RCA flow of 0.15 L/min; 2) a moderate disease with a total RCA flow of 0.08 L/min, obtained by uniformly increasing Ra, Rp, Rd by 100% relative to the baseline case; 3) a severe disease with a total RCA flow of 0.05 L/min, obtained by further uniformly increasing Ra, Rp, Rd by 50% relative to the moderate disease case. The simulated CICs demonstrate the substantial variability in the rising and falling slopes of the CICs. Slower rising and falling patterns in the CICs correspond to higher resistance values in the coronary microcirculation and a higher degree of CMD.
In summary, this study developed machine learning-based methods to computationally quantify the vFFR and investigate the relationship between the dynamics of the contrast and microcirculatory resistance as an assessment of CMD. Results show that AttMulFid provides an efficient, accurate vFFR assessment. Besides, a correlation between CIC and microcirculatory resistance is seen, as may be expected, and could have implications for future work.

Acknowledgements (Optional): This work is supported by NSF Grant # 2151555.