Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Cardiovascular Engineering

AI-associated cardiovascular engineering

Anatomical and Functional Assessment of Coronary Artery Disease Using Angiography

Thursday, October 24, 2024
1:15 PM - 1:45 PM EST
Location: Room 328

Introduction:  Coronary artery disease (CAD) and coronary microvascular dysfunction (CMD) refer to pathophysiology in large (epicardial) and small (arterioles and capillaries) coronary vessels, respectively. Large vessel disease manifests as local or diffuse plaque buildup that restrict flow, whereas small vessel disease is triggered by endothelial dysfunction, inflammation, and hormonal imbalances that affect perfusion, not primarily related to plaque buildup. The combination of CAD and CMD limits the ability of the coronary circulation to supply blood to the myocardium.
Diagnosis of CAD and CMD generally relies on different tests for epicardial and microcirculatory disease. For instance, fractional flow reserve (FFR) is used for epicardial disease, whereas CMD is assessed using positron emission tomography (PET). Recently, percutaneous approaches have made it possible to perform a comprehensive assessment of both epicardial and microcirculation disease states in a single procedure, although this approach is rarely utilized due to its invasive nature and complicated interpretation. Conversely, there are over 4 million cardiac catheterization procedures performed yearly in the United States and Europe alone, making it one of the most used diagnostic procedures, although with a long-recognized low diagnostic yield.
Given the high number of cardiac catheterization procedures and its low diagnostic yield, there is a pressing need to develop methods to extract diagnostic information for both CAD and CMD from coronary angiography data.

Materials and

Methods:  Our group has developed a series of data-driven computational models and machine learning tools for anatomical and functional characterization of CAD and CMD using angiography. These tools include: 1) convolutional neural networks for segmentation of angiograms [1]; 2) machine-learning methods for 3D reconstruction of coronary trees from several 2D angiography videos [2]; 3) an attention-based multi-fidelity reduced-order model (ROM) for efficient and accurate estimation of FFR [3], and 4) a data-driven model for assessing CMD.

Results, Conclusions, and Discussions: Fig. 1, panel (I) demonstrates the architecture of a ROM for FFR estimation, including the training and testing stages (blue and red boxes, respectively). In the training stage, a synthetic vessel generator is used to train an autoencoder for geometric features extraction. These synthetic vessels are used to generate corresponding hemodynamic data using high-fidelity (HF) –3D CFD– and low-fidelity (LF) –1D– models. Hemodynamic data and geometric features are used to construct a multi-fidelity feature fusion U-Net. In the patient-specific testing stage, vessel geometry and flow boundary conditions are supplied to the ROM to predict velocity, pressure and FFR.
Fig. 1, panel (II) demonstrates the performance of the ROM. Panel (a) compares ground truth and reconstructed 3D geometries. Gradient-based attention is presented in Panel (b). Panel (c) compares FFR solutions by the LF, HF, the ROM, as well as the invasive measurement (a single value of FFR= 0.77, purple dash-dotted line). The HF CFD pressure contour is presented in Panel (d). The ROM solution accurately captures the FFR profiles produced by the HF model. Furthermore, the ROM prediction compares well with the invasive measurements.
Analysis of dynamic contrast uptake and washout provides insight into the resistance of coronary microcirculation and thus CMD. Fig 2 provides a schematic representation of the analysis. Angiograms are recursively segmented using AngioNet (Figs. 2a and 2b) [1]. A time series of the net intensity for each segmented angiogram is generated by summing the number of filled pixels in each frame (Fig. 2c). These intensity time series is smoothed using an exponential moving window. Based on these normalized intensity time series, ‘automated TIMI frame counts’ can be calculated (Fig. 2d), by defining initial and final stages of filling. Using consistent minimum and maximum filling thresholds, the in-flow and falling slopes can be defined. This automated tool can quickly and objectively identify differences in angiographic contrast dynamics among patients. Fig. 3 demonstrates the large variability in rising and falling slopes in a cohort of 104 patients following PCI. Warm colors correspond to slower filling and emptying of the contrast agent, therefore suggesting a higher degree of microvascular disease.

Acknowledgements (Optional):  [1] Iyer et al. 10.1038/s41598-021-97355-8
[2] Iyer et al.10.1038/s41598-023-44633-2
[3] Yang, Figueroa, Garikipati. arXiv:2311.11397