Zeng Yao, Sun Zheng, Wang Mengfei, Li Zhuo, Liu Ao, Pan Meixiu, Zhao Haifeng, Li Yuehua
Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai 200233, PR China
University of Shanghai for Science and Technology, No.516 Jungong Road, Shanghai 200093, PR China
Comput Methods Programs Biomed. 2025 Sep;269:108924. doi: 10.1016/j.cmpb.2025.108924. Epub 2025 Jun 23.
Population-scale hemodynamic research faces limitations due to the trade-off between computationally expensive patient-specific Computational Fluid Dynamics (CFD) and overly idealized cylindrical models. To overcome this, we propose a novel Tier-2 workflow that integrates point-cloud statistical shape modeling (Pcd-SSM) with HDBSCAN clustering. This approach aims to efficiently characterize the C1 geometries of the internal carotid artery (ICA) and analyze their corresponding flow patterns.
Time-of-flight Magnetic Resonance Angiography (TOF-MRA) data from 229 ICAs (171 normal, 58 with 30-50 % stenosis) were converted into 1024-point correspondences using Point2SSM. Principal Component Analysis (PCA), retaining 95 % variance, was applied before unsupervised clustering. A bootstrap-FDR test on 97 normal cases established the global replacement limit, Ncrit = 40. Cluster mean models were then meshed and evaluated using steady, non-Newtonian CFD simulations. The results were benchmarked against individual simulations and diameter-based models.
This framework significantly enhances the accuracy (error reduction of 77-95 %) and efficiency (computational cost reduced by approximately 65 times) of hemodynamic simulations in medium-to-large cohorts (Tier 2). Applied to stenosed arterial segments, the model successfully captured approximately 91 % of velocity increases and 51 % of pressure drops, accurately revealing high wall shear stress distributions.
Our "tier-and-cluster" generalization framework, driven by deep learning Pcd-SSM, provides a unified and transferable paradigm for analyzing complex vascular morphology and blood flow. It offers a robust tool for population-level blood flow studies, individualized risk stratification, exploration of pathological mechanisms, and evaluation of intervention timing for vascular stenosis.
由于计算成本高昂的患者特异性计算流体动力学（CFD）与过度理想化的圆柱模型之间存在权衡，大规模血流动力学研究面临局限性。为克服这一问题，我们提出了一种新颖的二级工作流程，该流程将点云统计形状建模（Pcd-SSM）与HDBSCAN聚类相结合。此方法旨在有效表征颈内动脉（ICA）的C1几何形状并分析其相应的血流模式。
使用Point2SSM将来自229条ICA（171条正常，58条有30 - 50%狭窄）的飞行时间磁共振血管造影（TOF-MRA）数据转换为1024点对应关系。在无监督聚类之前应用主成分分析（PCA），保留95%的方差。对97例正常病例进行的自抽样错误发现率检验确定了全局替换极限，Ncrit = 40。然后对聚类平均模型进行网格划分，并使用稳态、非牛顿CFD模拟进行评估。结果与个体模拟和基于直径的模型进行了对比。
该框架显著提高了中大型队列（二级）血流动力学模拟的准确性（误差降低77 - 95%）和效率（计算成本降低约65倍）。应用于狭窄动脉段时，该模型成功捕获了约91%的速度增加和51%的压力下降，准确揭示了高壁面剪应力分布。
我们由深度学习Pcd-SSM驱动的“分层与聚类”泛化框架为分析复杂血管形态和血流提供了一个统一且可转移的范例。它为人群水平的血流研究、个体化风险分层、病理机制探索以及血管狭窄干预时机评估提供了一个强大的工具。
