Session H14: Biological Fluid Dynamics: Data-driven Hemodynamics

Fast surrogate of 3-D patient-specific computational fluid dynamics using statistical shape modeling and deep Learning

Optimization and uncertainty quantification (UQ) have been playing an increasingly important role in computational hemodynamics. However, existing methods based on principled modeling and classic numerical techniques have faced significant challenges, particularly when it comes to complex 3D patient-specific shapes in the real world. First, it is notoriously difficult to parameterize the input space of arbitrarily complex 3-D geometries. Second, the process often involves massive forward simulations, which are extremely computationally demanding or even infeasible. We propose a novel deep learning solution to address these challenges and enable scalable geometric UQ and optimization. Specifically, a statistical generative model for 3-D patient-specific shapes will be constructed based on a handful of available baseline patient geometries.  An unsupervised shape correspondence solution is used to enable geometric morphing and a compact geometric design space can then be constructed by the statistical generative shape model. In order to build a fast forward map between geometric input space to the solution space of functional information, we propose a supervised deep learning solution, which will facilitate shape optimization and UQ analysis in a massively scalable manner.   

Inferring the left atrial appendage (LAA) hemodynamics from 4D CT contrast dynamics: reduced order models (ROMs) and physics informed neural networks (PINNs).

Atrial fibrillation (AF) is a common arrhythmia affecting > 40M people worldwide. During AF, blood inside the LAA becomes stagnant and can form clots, some of which can travel to the brain to cause a stroke. Current tools (CHA2DS2-VASc score) to predict stroke risk in AF patients are not personalized and have modest accuracy. We aim to infer each patient's clotting risk from 4D CT acquisitions of LAA contrast dynamics. We consider ROMs for near-real-time image analysis and high-fidelity PINNs. We run multiple inexpensive ROMs to derive optimal imaging settings balancing predictive accuracy with patient radiation dose. Our ground truth comprises patient-specific CFD simulations, including contrast agent dynamics. We find that advection-diffusion ROMs can infer the average blood residence time in the LAA but fail to capture its fine-scale spatio-temporal features. On the other hand, PINNs, albeit more computationally demanding, can fully infer LAA hemodynamics using each patient's 4D contrast agent concentration fields from CFD as training data. Finally, we show proof of concept of the application of ROMs to infer LAA residence time using 4D CT data acquired in the clinical setting.   

A Machine Learning Methodology for estimation of vascular characteristics using a single carotid waveform

Cardiovascular diseases are the leading causes of morbidity and mortality. Methods for early detection of vascular impairment provide intuition into cardiovascular disease pathogenesis. For instance, any change in arterial compliance may lead up to the onset of a clinically recognizable disease, and by knowing that caregivers can identify patients at risk even before the onset of clinical signs and symptoms. It is notable that vascular network characteristics such as total arterial compliance (TAC) and aortic characteristic impedance (ACI) are among the key factors for cardiovascular disease detection. Challenges for noninvasive and single waveform evaluation of TAC and ACI have raised the demand for employing machine learning (ML) tools. This study presents an ML-based methodology for calculating TAC and ACI from a single carotid waveform. This method was trained and tested on a large human cohort with an age range from 19 to 90 (Framingham Heart Study). The dataset includes both healthy and patients with cardiovascular diseases (53% female). The final model was also tested on a set of clinical data blinded to all stages of development. This can be clinically significant since carotid waveforms can be captured noninvasively using tonometry-type devices or even an iPhone camera.   

Towards resolving hemodynamic velocities from time-resolved contrast-enhanced magnetic resonance angiography using physics-informed machine learning

Time-resolved contrast-enhanced magnetic resonance angiography (TR-CEMRA), a non-invasive imaging technique to track contrast techniques in-vivo, is typically used to visualize unusual vascular flow patterns such as retrograde filling near an occlusion. In this research, we investigate the possibility of estimating the time-resolved three-dimensional blood velocities from TR-CEMRA data. We use the newly developed physics-informed neural networks (PINN). Variables of interest such as velocities, pressure, and contrast concentration are modeled as deep neural nets. The hidden variables (velocity and pressure) are inferred from assimilating time-resolved TR-CEMRA data during the training process. The training process also imposes the physics, namely, Navier-Stokes equation and mass conservation for blood flow, and advection-diffusion equation for contrast. The method requires time stamping every 2D Cartesian scan in the 3D region of interest.  Gaussian quadrature is used to volume average neural net outputs to match the TR-CEMRA acquisition process for data fidelity. We demonstrate examples in 2D and 3D aneurysm models wherein the reference flow and contrast dynamics were obtained with computational fluid dynamics simulations.   

Data-driven near-wall blood flow and wall shear stress modeling with physics-informed neural networks

High fidelity quantification of near-wall blood flow and wall shear stress (WSS) in diseased arteries is challenging. Experimental limitations in the quantification of WSS, uncertainty in computational models, and the complex multi-directional behavior in WSS vector fields challenge accurate WSS modeling. Physics-informed neural networks (PINN) provide a flexible machine learning framework to integrate mathematical equations such as Navier-Stokes with sparse measurement data. In this work, we demonstrate how PINNs can quantify WSS by leveraging sparse blood flow data collected away from the wall, which are easier to measure experimentally. We challenge the framework by not providing inlet and outlet boundary condition data. PINN combines the governing blood flow equations with the provided sparse velocity data away from the wall to obtain WSS. We demonstrate examples in 2D and 3D idealized stenosis and aneurysm models. Finally, inspired by the classical analytical Womersley solution for pulsatile flow in tubes, we demonstrate the application of Fourier features for efficient extension of the PINN approach to pulsatile blood flow problems.    

Brain Hemodynamic Predictions from Noninvasive Transcranial Doppler Ultrasound and Angiography Data Using Physics-Informed Neural Networks

Determining brain hemodynamics plays a critical role in the diagnosis and treatment of various cerebrovascular diseases. Therefore, there is a need to provide rapid, reliable, physiologically correct, and spatiotemporally resolved hemodynamic data for intracranial arteries. In this work, we put forth a deep learning framework that augments sparse clinical measurements with computational fluid dynamic (CFD) simulations to generate physically consistent hemodynamic parameters and vessel cross-sectional areas in the entire brain vasculature. Transcranial Doppler (TCD) ultrasound is one of the most common techniques in the current clinical workflow that enables the noninvasive measurement of blood flow velocity within the cerebral arteries. Our deep learning framework employs in vivo real-time TCD ultrasound velocity measurements at several spatial positions in the brain and the baseline vessel cross-sectional areas acquired from 3D magnetic resonance imaging (MRI) angiograms to provide high-resolution maps of velocity, area, and pressure in the entire vasculature. We validated the predictions of deep learning models against in vivo velocity measurements from the same subject obtained via 4D flow MRI scans. The key finding here is that the combined effects of uncertainties in outlet boundary condition subscription and physics deficiencies render the conventional purely physics-based computational models unsuccessful in recovering accurate brain hemodynamics. Nonetheless, fusing these models with in vivo clinical measurement through a data-driven approach ameliorates predictions of brain hemodynamic variables substantially. Finally, we showcase the clinical significance of the proposed technique in diagnosing the cerebral vasospasm (CVS) induced by intracranial aneurysm (IA) rupture by approximating the vasospastic vessel’s local diameters.   

Improving the Diagnostic Accuracy of Cardiac Auscultation using Supervised Learning: a Computational Hemoacoustic Study

Valve replacement is the primary mode of treatment for calcific aortic valve (AV) diseases, but prosthetic valves can fail prematurely and without warning, potentially causing death. Early diagnosis and intervention can avert such adverse outcomes. Heart sounds associated with unsteady blood flow and leaflet motion contain valuable information about valve function and auscultation-based diagnosis can provide a safe, cheap, and at-home means of preliminary screening for valve failure. However, its reliance on physician's proficiency leads to poor accuracy (<30%) which may be improved via machine learning. We present an in-silico analysis showing how these techniques can be combined to develop an accurate, non-invasive early detection modality for AV failure. We simulated transvalvular flow in 29 cases of healthy and mildly stenotic AVs using an immersed boundary method-based solver and coupled fluid-structure interaction with a simple acoustic transmission model. We describe hemodynamics underlying healthy and pathological AV sounds and train a linear discriminant classifier to detect AV anomaly using characteristics of the recorded sound. Diagnostic accuracy can thus be improved to around 90%, and its subjectivity can be alleviated using automated electronic auscultation.   

Analyzing Patient-Specific Coronary Arteries with and without Stents using Proper Orthogonal Decomposition

Incompressible transient CFD is run on patient-specific coronary arteries extracted from angiograms and intravascular ultrasound. The snapshot proper orthogonal decomposition (POD) method is used to generate POD modes for the patients' arteries, before and after stent placement. An examination of the highest energy modes highlight the areas of concern and illustrate the success of the stent procedure. It is shown that most of the dynamics of the diseased coronary arteries can be captured in the fewer than five modes, and less for the arteries with proper stent placement. Wall shear stress (WSS) distributions are examined as a high frictional force exerted parallel to coronary arterial walls may lead to or make plaque development worse. The POD modes generated from the patient-specific study are to be used in combination with machine learning techniques to develop predictive and diagnostic tools for healthcare in future work.   

All-in-one, physics-informed dealiasing method to regularize cardiac 4D flow MRI data.

Cardiac 4D flow MRI provides insight into cardiovascular pathophysiology. Of note, it can identify slow flow regions associated with thrombosis and cardioembolic stroke. However, clinically recommended velocity encodings (VENC) to prevent aliasing in the transvalvular jets yield a poor signal-to-noise ratio (SNR) in slow flow regions. Multi-VENC acquisitions are unfeasible given the current long acquisition times of 4D flow MRI. Low-VENC acquisitions require dealiasing, but existing tools cannot be easily integrated with other regularization steps (e.g., de-noising or velocity divergence removal). We present a single-step algorithm that simultaneously corrects aliasing in low-VENC acquisitions, removes image noise, and imposes physical constraints such as div(v)=0. We formulate a least-squares problem including a dealiasing penalty derived by generalizing existing methods from the meteorology field and additional penalties from physics-informed priors. Algorithm performance is tested on synthetic flows with varying SNR and VENC, and L1 vs. L2 regularizations are compared. The algorithm is tested on cardiac 4D flow MRIs of N=5 human subjects.   

Super-resolution study of aneurism based on AI PIV - preliminary report

The paper presents the results of CFD blood flow simulations in a 3D model of the cerebral aneurism for different flow rates and compares them with values obtained with Particle Image Velocimetry (PIV) in a 3D experimental model under pulsatile flow conditions. CFD simulation was done using Ansys CFX software. For validation of the experiments, two different PIV methods were employed: traditional PIV based on cross correlation algorithms and Artificial Intelligence PIV (AI PIV) based on Deep Learning and Convolutional Neural Networks, which allowed for a much higher spatial resolution of the resulting velocity field. A 3D model of cerebral aneurism was developed based on specific patient CT data. The protocol of the study was approved by the Institutional Review Board at the Medical University of Silesia in Katowice and all procedures were carried out in accordance with the relevant guidelines and regulations. The morphometric parameters were taken from patient whose vessel fulfils the Murray's law. The preliminary obtained simulation results, showed that better correlation between CFD simulation and types of PIV analysis is for data of AI PIV than ones obtained with traditional PIV. There was observed increase in the value of the shear stress of the vessel walls with an increase in the flow rate. Analysis of the results were promising and showed that AI PIV could be used to get more accurate results.