From Medical Imaging to Hemodynamics: AI-Enabled Modeling of Cardiovascular Flow

Cardiovascular disease remains the leading cause of mortality in the United States, motivating the development of patient-specific tools for diagnosis, risk stratification, and treatment planning. Image-based computational fluid dynamics (CFD) has emerged as a powerful framework for characterizing cardiovascular hemodynamics by integrating medical imaging with physics-based simulation. This pipeline typically involves image acquisition, vascular segmentation, geometry generation, flow simulation, and post-processing, among which segmentation, geometry generation, and flow simulation are the most critical and technically challenging components. Conventional segmentation workflows are often manual or semi-automatic, making them time-consuming, operator-dependent, and difficult to scale. To address these limitations, this thesis develops AI-enabled segmentation frameworks that automatically generate simulation-ready vascular geometries, substantially improving efficiency while reducing user effort and variability. Beyond segmentation, the scarcity of high-quality patient-specific vascular geometries poses a major challenge for data-driven cardiovascular modeling, particularly for complex multibranch structures that are inadequately represented by traditional statistical shape models. This work introduces generative AI approaches for vascular geometry synthesis that combine hierarchical parameterizations with diffusion-based models, enabling the generation of anatomically consistent, CFD-ready geometries for data augmentation and population-level analysis. To overcome the high computational cost of traditional CFD solvers, this thesis further develops deep learning-based surrogate models that rapidly map vascular geometries to flow fields, as well as probabilistic modeling frameworks to capture complex flow phenomena and quantify uncertainty. In parallel, a GPU-accelerated differentiable finite-volume solver based on graph neural network message passing is introduced to enable efficient simulation on unstructured meshes. Together, this thesis presents a unified AI-enabled image-based CFD framework that advances the development of efficient, reliable, and clinically actionable cardiovascular digital twins.心血管疾病仍是美国人群死亡的首要诱因，这推动了针对患者个体的诊断、风险分层与治疗规划工具的研发。基于影像的计算流体力学（computational fluid dynamics, CFD）通过整合医学影像与物理仿真，已成为表征心血管血流动力学的有力框架。该流程通常涵盖影像采集、血管分割、几何建模、血流仿真与后处理等步骤，其中分割、几何建模与血流仿真正是最为关键且技术难度最高的环节。传统的分割流程多为手动或半自动操作，不仅耗时耗力，还依赖操作者水平，且难以实现规模化应用。为解决上述局限，本研究开发了基于人工智能的分割框架，可自动生成可直接用于仿真的血管几何模型，在大幅提升效率的同时，降低了人工投入与结果差异。除分割环节外，高质量患者特异性血管几何模型的匮乏，是制约数据驱动型心血管建模的一大瓶颈，尤其对于传统统计形状模型难以充分表征的复杂多分支血管结构而言。本研究提出了融合分层参数化与基于扩散的模型（diffusion-based models）的生成式AI（Generative AI）血管几何合成方法，可生成符合解剖学特征、可直接用于CFD的血管几何模型，用于数据增强与群体水平分析。为解决传统CFD求解器计算成本高昂的问题，本研究进一步开发了基于深度学习的代理模型，可快速将血管几何模型映射为流场；同时构建了概率建模框架，以捕捉复杂血流现象并量化不确定性。与此同时，本研究提出了一种基于图神经网络（Graph Neural Network, GNN）消息传递的GPU加速可微有限体积求解器，可在非结构化网格上实现高效仿真。综上，本研究构建了一套统一的基于影像的人工智能辅助CFD框架，推动了高效、可靠且可应用于临床的心血管数字孪生技术的发展。