Collaborative Research: Enhanced 4D-Flow MRI through Deep Data Assimilation for Hemodynamic Analysis of Cardiovascular Flows

血流与血管壁相互作用产生的作用力对诸如动脉瘤、动脉粥样硬化和血管痉挛等血管疾病的发生和发展具有重要影响。因此，详细和准确的血流分析可能是此类疾病预后和治疗的关键。目前有两种流行的方法用于研究3D血流。第一种是基于计算流体动力学(CFD)的模拟。第二种是通过使用相位对比磁共振成像(也称为4D-Flow MRI)等技术的直接非侵入性成像。CFD需要准确的血管几何形状、模型参数以及对边界流和初始条件的估计。这些都是耗时和非常困难的，如果不是不可能估计的话。此外，CFD的保真度受到模型假设的限制。另一方面，4D-Flow MRI直接测量体内的体积血流速度，但时空分辨率低，扫描图像受到噪声和图像伪影的污染。拟议的项目通过一种名为深度数据同化的新技术克服了CFD和4D-Flow MRI的局限性。这里使用深度神经网络对血液流动进行建模。训练过程使用4D-Flow MRI实现数据保真度，同时确保满足流体流动和磁共振的物理特性。然后，神经网络被用来生成准确的密集时空流场和流动相关参数，如壁面剪应力、涡度等。增强4D-Flow MRI的能力将使临床研究人员能够研究血流动力学对血管疾病发生和发展的影响。这将为疾病管理带来新的基于物理的流动图像分析工具，将显著降低成本并优化治疗计划。拟议项目的目标是能够从时间分辨的三维相位对比磁共振成像(4D-Flow MRI)对心血管流动进行准确和可靠的血流动力学分析。该方法利用物理信息的深度学习，将时变的流量(速度和压力)和场(磁矩)变量建模为深度神经网络。训练过程符合4D-Flow MRI数据，并将血流物理(Navier-Stokes方程)和MRI采集物理(Bloch方程)作为约束条件。在学习过程中创造性地设计损失函数，将达到超分辨率，衰减噪声，并消除各种图像伪影。自动微分将有助于速度相关的高阶血流动力学参数的截断误差无计算。将使用精心设计的体外实验来验证和优化该方法。提出的混合实验和深度学习方法将在心血管血流研究中创建一种新的范式，其中控制方程将直接应用于使用深度学习的低质量成像数据，以将可靠性和准确性提高到科学发现所需的水平。该项目将为研究生提供最新的工程深度学习技术培训，通过密尔沃基大学和北亚利桑那大学的众多项目让本科生参与研究，并通过密尔沃基大学的暑期项目接触边缘社区的高中生。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Forces resulting from blood flow interaction with walls of blood vessels have major impact on the initiation and progression of vascular diseases such as aneurysms, atherosclerosis, and vasospasms. Consequently, detailed and accurate blood flow analysis could be key to prognosis and treatment of such diseases. There are two popular modalities that are currently used to study 3D blood flow. The first is based on computational fluid dynamic (CFD) simulations. The second is through direct non-invasive imaging using techniques such as phase contrast magnetic resonance imaging (a.k.a 4D-Flow MRI). CFD requires accurate vascular geometry, model parameters, and estimates of boundary flow and initial conditions. These are time consuming and very difficult, if not impossible to estimate. Furthermore, the fidelity of CFD is limited by model assumptions. On the other hand, 4D-Flow MRI directly measures in-vivo volumetric blood flow velocities, but has low spatio-temporal resolution and the scans are contaminated by noise and image artifacts. The proposed project overcomes the limitations of both CFD and 4D-Flow MRI through a novel technique called deep data-assimilation. Here deep neural nets are used to model the blood flow. The training process imposes data fidelity with 4D-Flow MRI and simultaneously ensures that the physics of fluid flow and magnetic resonance are satisfied. The neural nets are then used to generate accurate dense spatio-temporal flow fields and flow dependent parameters such as wall shear stresses, vorticity etc. The ability to enhance 4D-Flow MRI will enable clinical researchers to investigate the impact of hemodynamics on the initiation and progression of vascular diseases. This will lead to novel physics-based flow image analysis tools for disease management that will significantly reduce cost and optimize treatment plans.The goal of the proposed project is to enable accurate and reliable hemodynamic analysis of cardio-vascular flows from time resolved three dimensional phase contrast magnetic resonance imaging (4D-Flow MRI). The proposed approach uses physics informed deep learning wherein time-varying flow (velocity and pressure) and field (magnetic moment) variables are modeled as deep neural nets. The training process fits 4D-Flow MRI data and also imposes blood flow physics (Navier-Stokes equation) and MRI acquisition physics (Bloch equations) as constraints. Creative design of loss functions in the learning process will achieve super-resolution, attenuate noise, and eliminate various image artifacts. Automatic differentiation will facilitate truncation error-free computation of velocity-dependent higher order hemodynamic parameters. Carefully designed in-vitro experiments will be used to validate and optimize the method. The proposed hybrid experimental and deep learning approach will create a new paradigm in cardiovascular flow research wherein the governing equations will be directly applied to low quality imaging data using deep learning to raise the reliability and accuracy to the level needed for scientific discovery. The project will provide opportunities to train graduate students in the latest deep-learning based techniques in engineering, engage undergraduate students in research through numerous programs at UW-Milwaukee and Northern Arizona University, and outreach to high school students belonging to marginalized communities through summer programs at UW-Milwaukee.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.