Data-driven Methods for Automated Assessment of Coronary Artery Disease

Abstract

There are several diagnostic tools to assess stenosis severity in coronary artery disease. The gold standard is anatomical assessment via cardiologists’ visual inspection of X-ray angiography (XRA), which can be subjective. Quantitative Coronary Angiography (QCA) is a less subjective, semi-automated method to quantify stenosis severity. Alternatively, functional metrics that estimate how a stenosis affects blood flow, such as Fractional Flow Reserve (FFR), have demonstrated better diagnostic outcomes than anatomical assessment; however, they are not widely used due to their cost and risk. Ideally, quantitative and functional information could be derived directly from XRA images without the additional operator dependence, time, and cost associated with performing FFR or QCA.

This project aims to develop automated approaches for anatomical and functional quantification of disease severity using XRA images. To this end, we have developed algorithms for 1) coronary vessel segmentation, 2) stenosis detection and characterization, 3) 3D reconstruction of coronary trees, and 4) image-based flow extraction. These algorithms can be used in conjunction with computational fluid dynamics modeling to assess the functional significance of disease. We first present AngioNet, a neural network for coronary segmentation from XRA images. AngioNet’s key innovation is an Angiographic Processing Network which learns the best pre-processing filters to improve segmentation performance. AngioNet demonstrates state-of-the-art segmentation accuracy (Dice score=0.864) and does not segment the catheter in challenging cases where other neural networks fail.

Building upon AngioNet, we developed a neural network and image processing pipeline to automate the tasks of QCA: localizing, segmenting, and measuring stenoses. This pipeline was able to measure stenosis diameter within 0.206±0.155mm or approximately 1 pixel of ground truth measurements from QCA. It is also the first automated pipeline to quantify rather than categorize disease severity.

To assess functional disease severity, we require the patient’s 3D coronary geometry to perform hemodynamic simulations and compute FFR. To this end, we developed a machine learning approach for 3D vessel reconstruction from a series of uncalibrated 2D XRA images. Our neural network has demonstrated sub-pixel error in radius reconstruction (0.16±0.07mm) and 1% error in FFR computed in a reconstructed coronary tree.

In addition to the 3D coronary geometry, information about patient-specific flow or pressure is required to perform a hemodynamics simulation for FFR computation. We developed an algorithm that tracks vessel area in sequential XRA frames to estimate relative flow in each branch. We validated the algorithm using a simulation of dye transport under steady flow conditions as the ground truth. On average, the error in relative flow per branch was 5.15% for a healthy coronary tree and 3.68% in a coronary tree with stenosis.

We finally combined the methods developed in this thesis to compare stenosis severity and FFR using the above algorithms against patient-specific clinical measurements in a proof-of-concept analysis. In this patient, AngioNet and the stenosis characterization pipeline accurately captured the vessel anatomy and stenosis diameters. There was a substantial 11% error in computed FFR (0.98) compared to clinically measured FFR (0.88), mainly due to errors in 3D reconstruction. Reconstruction accuracy could be improved by retraining the network on a wider range of anatomical variations and incorporating patient motion between input images to match clinical practice. Another improvement would be to automate and incorporate motion compensation in the flow extraction algorithm so that it can be applied to clinical angiograms.