CFD-based Synthetic Data Generation for Machine Learning based Pressure Drop Assessment in Aortic Stenosis

Teodor Ionut MATEI, Andreea Bianca POPESCU, Cosmin Ioan NITA, Costin Florian CIUSDEL, Lucian Mihai ITU

Abstract

Aortic stenosis occurs when the aortic valve does not fully open during the systole, which reduces and partially blocks the blood flow to the systemic circulation. The clinical diagnosis and treatment decision depend on the functional severity of the stenosis, which is assessed based on the trans-valvular peak pressure drop. The pressure drop is routinely estimated analytically, which may lead to sub-optimal results since certain hemodynamic aspects are not fully captured, like pressure recovery or blood flow turbulence. A promising solution lies in the use of machine learning (ML) for estimating the pressure drop based on patient-specific characteristics. Although a ML-based solution could provide the desired results in real time, the training of an accurate neural network would require a large database of invasively measured pressure drops, which is difficult and costly to set up. This study introduces a method for generating synthetic datasets based on a generic aortic valve model. This model is customized in order to create diverse yet physiologically normal valve shapes by modifying three anatomical parameters, namely the aorta diameter, blood velocity and valve area, by reducing it. High-fidelity computational fluid dynamics (CFD) simulations were conducted to compute reference pressure drop values. The efficacy of the generated dataset was assessed by employing it for training an ML model for pressure drop estimation.

Keywords

Synthetic Data, Computational Fluid Dynamics, Aortic Stenosis, Machine Learning.

View full article