TY - JOUR

T1 - A feasibility study of deep learning for predicting hemodynamics of human thoracic aorta

AU - Liang, Liang

AU - Mao, Wenbin

AU - Sun, Wei

N1 - Funding Information:
Research for this project is funded in part by AHA is American Heart Association Award 18TPA34230083 . Appendix The CFD simulations were performed using STAR-CCM+ (CD-adapco, Melville, NY) CFD software. The aortic wall was assumed to be rigid with no-slip boundary condition. The physics model employed an incompressible Newtonian fluid with a reference density ρ = 1056 k g / m 3 and dynamic viscosity μ = 0.0035 P a ∙ s for blood properties ( Mao et al., 2016a, 2016b ). Five layers of prism layer meshes were used near the aortic wall and 1 mm size polyhedral meshes were generated at the core of the domain. Approximately 150,000 to 350,000 mesh cells per model were generated and this mesh density was found to be sufficient to provide mesh-independent results. Since the flow is in the transitional or low-Reynolds turbulent flow regime, turbulence modeling was performed using the K-Omega SST model with low-Reynolds-number damping modification available in STAR-CCM+. The Navier-Stokes equations for 3D flow were solved by a second-order segregated iterative method (SIMPLE algorithm). A constant flow rate of 25 L/min was applied at the ascending aorta inlet, corresponding to a physiological flow rate in peak systole ( Bonfanti et al., 2017 ). A turbulence 1/7 power law velocity profile ( Štigler, 2014 ) was assumed with a cutoff radius of the 80% of the ascending aorta (AAo) radius to represent a physiological velocity profile downstream the aortic valve. Simplified Windkessel models were applied at the outlets of the brachiocephalic artery (BA), left common carotid artery (LCCA), left subclavian artery (LSA), and descending aorta (DAo). Since the compliance in the Windkessel model only affects the transient flow, only the total resistance is considered for the peak systolic flow through the aorta. Therefore, the relationship between the outlet pressure P and the flow rate Q could be defined by Q = ( P - P 0 ) / R , where P 0 is the capillary pressure of 4.4 kPa ( Alastruey et al., 2016 ) and R is the total resistance from the downstream peripheral vessels. The values of R were chosen as 118022, 477871, 375470, and 35,156 P a ∙ s / k g for the BA, LCCA, LSA, and DAo outlets respectively, ensuring a physiological flow split ratio of 19.3%, 5.2%, 6.4%, and 69.1% at the corresponding outlets of the BA, LCCA, LSA, and DAo ( Alastruey et al., 2016; Bonfanti et al., 2017 ). We note that a constant flow rate does not fully specify the inlet flow boundary condition. The specified velocity profile actually depends on the orifice size of ascending aorta because a cutoff radius of the 80% of the AAo radius was used to adjust the velocity profile, as explained in the above paragraph. This means different models may have different cutoff radii or different velocity profiles. The velocity profiles are tuned so that the flow rate (equals the average inlet velocity multiplies the orifice area of velocity profile) is 25 L/min. The prescribed velocity boundary condition is a simplification compared to the real aortic forward flow. However, the specified velocity profiles partially considered the effect of the aortic valve opening in peak systole by restricting the flow profile in the central region of the ascending aorta orifice. The pressure fields and flow fields from the CFD simulations have large variations across different geometries, and therefore the CFD simulation data are sufficient for this feasibility study.
Publisher Copyright:
© 2019 Elsevier Ltd

PY - 2020/1/23

Y1 - 2020/1/23

N2 - Numerical analysis methods including finite element analysis (FEA), computational fluid dynamics (CFD), and fluid–structure interaction (FSI) analysis have been used to study the biomechanics of human tissues and organs, as well as tissue-medical device interactions, and treatment strategies. However, for patient-specific computational analysis, complex procedures are usually required to set-up the models, and long computing time is needed to perform the simulation, preventing fast feedback to clinicians in time-sensitive clinical applications. In this study, by using machine learning techniques, we developed deep neural networks (DNNs) to directly estimate the steady-state distributions of pressure and flow velocity inside the thoracic aorta. After training on hemodynamic data from CFD simulations, the DNNs take as input a shape of the aorta and directly output the hemodynamic distributions in one second. The trained DNNs are capable of predicting the velocity magnitude field with an average error of 1.9608% and the pressure field with an average error of 1.4269%. This study demonstrates the feasibility and great potential of using DNNs as a fast and accurate surrogate model for hemodynamic analysis of large blood vessels.

AB - Numerical analysis methods including finite element analysis (FEA), computational fluid dynamics (CFD), and fluid–structure interaction (FSI) analysis have been used to study the biomechanics of human tissues and organs, as well as tissue-medical device interactions, and treatment strategies. However, for patient-specific computational analysis, complex procedures are usually required to set-up the models, and long computing time is needed to perform the simulation, preventing fast feedback to clinicians in time-sensitive clinical applications. In this study, by using machine learning techniques, we developed deep neural networks (DNNs) to directly estimate the steady-state distributions of pressure and flow velocity inside the thoracic aorta. After training on hemodynamic data from CFD simulations, the DNNs take as input a shape of the aorta and directly output the hemodynamic distributions in one second. The trained DNNs are capable of predicting the velocity magnitude field with an average error of 1.9608% and the pressure field with an average error of 1.4269%. This study demonstrates the feasibility and great potential of using DNNs as a fast and accurate surrogate model for hemodynamic analysis of large blood vessels.

KW - Computational fluid dynamics

KW - Deep neural network

KW - Hemodynamic analysis

KW - Machine learning

UR - http://www.scopus.com/inward/record.url?scp=85076534435&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85076534435&partnerID=8YFLogxK

U2 - 10.1016/j.jbiomech.2019.109544

DO - 10.1016/j.jbiomech.2019.109544

M3 - Article

C2 - 31806261

AN - SCOPUS:85076534435

VL - 99

JO - Journal of Biomechanics

JF - Journal of Biomechanics

SN - 0021-9290

M1 - 109544

ER -