Geometry-Aware Hemodynamics via a Transformer Encoder and Anisotropic RBF Decoder

Reza Akbarian Bafghi, Sukirt Thakur, Maziar Raissi

Abstract:

Accurate and rapid estimation of hemodynamic metrics, such as pressure and wall
shear stress (WSS), is essential for diagnosing and managing Coronary Artery
Disease (CAD). Existing approaches, including invasive Fractional Flow Reserve
(FFR) measurements and computationally expensive Computational Fluid Dynam
ics (CFD) simulations, face challenges in invasiveness, cost, and speed. We present
a framework that accelerates non-invasive coronary hemodynamics prediction. The
model integrates 1D centerline and inlet flow rate into a transformer-based encoder,
followed by an anisotropic Radial Basis Function (RBF) decoder that aligns with
vessel morphology for continuous wall-based predictions. We also introduce a large
synthetic dataset of 4,000 single-vessel coronary artery geometries with correspond
ing steady-state flow simulations, enabling robust training and evaluation. Our
method improves accuracy and delivers orders-of-magnitude speedups over CFD on
this synthetic benchmark. While tested only on steady, single-vessel cases, it shows
promise for clinical acceleration; validation on clinical data and extension to multi
vessel and transient settings are important next steps. Dataset available at: https:
//huggingface.co/datasets/angioinsight/single-vessel-flow