Estimation of FFR in coronary arteries with deep learning

TL;DR

The paper tackles the challenge of noninvasively estimating fractional flow reserve (FFR) in coronary arteries by replacing time-intensive CFD simulations with a deep learning surrogate. It proposes a hybrid approach that uses 3D point clouds to learn implicit geometric features while augmenting them with explicit hand-crafted features, achieving rapid vFFR estimates from synthetic data. Across extensive experiments, the method demonstrates strong agreement with CFD-derived vFFR, robust performance across varying inflow conditions and stenosis severities, and clinically viable decision-making with high accuracy at the standard threshold of $0.8$. The work highlights significant potential for on-site, real-time vFFR screening and outlines concrete paths toward validation with real patient data and physics-informed extensions.

Abstract

Coronary artery disease (CAD) is one of the most common causes of death in the European Union and the USA. The crucial biomarker in its diagnosis is called Fractional Flow Reserve (FFR) and its in-vivo measurement is obtained via an invasive diagnostic technique in the form of coronagraphy. In order to address the invasive drawbacks associated with a procedure, a new approach virtual FFR (vFFR) measurement has emerged in recent years. This technique involves using computed tomography angiography (CTA) to obtain virtual measurements of FFR. By utilizing Computational Fluid Dynamics (CFD), vFFR estimates can be derived from CTA data, providing a promising in-silico alternative to traditional methods. However, the widespread adoption of vFFR from CTA as a diagnostic technique is hindered by two main challenges: time and computational requirements. In this work, we explore the usage of deep learning techniques as surrogate CFD engine models in the task of vFFR estimation in coronary arteries to drastically limit the required time and computational costs without a major drop in quality. We propose a novel approach to vFFR estimation by representing the input vessel geometry as a point cloud and utilizing the hybrid neural network that learns geometry representation based on both explicitly and implicitly given features. We evaluate the method from the clinical point of view and showcase that it can serve as a compelling replacement for commonly utilized CFD-based approaches.

Figures (27)

• Figure 1: Coronarography procedure. The procedure is supervised via coronary angiography imaging and if required, the angioplasty treatment is performed by placing a stent in the stenotic region. The white arrow marks the severe stenosis seen on the coronary angiography imaging.
• Figure 2: CFD-based vFFR estimation methods from CCTA are often derived from the common pipeline. Based on the CCTA image (A) the fine-grained anatomical 3D model (B) and physiology model defining boundary conditions (C), are constructed. Both of them serve as prerequisites of CFD numerical solver (D). The results are often visualised for the physician in the form of colour-coded 3D mesh (E).
• Figure 3: Workflow of Siemens framework. The model is trained on synthetic data offline and then deployed online to perform vFFR regression on real patients.
• Figure 4: Deep learning approach by Siemens. The simple $4$-layer MLP is utilized to regress local vFFR from hand-crafted features in the form of local and global vessel characteristics.
• Figure 5: DEEPVESSEL-FFR framework. This approach utilizes two phases, a simple feature encoder, similar to Siemens framework (\ref{['sec:2.1.1']}), and a latter RNN for more robust vFFR regression that includes upstream and downstream vessel characteristics.