Coronary microvascular dysfunction (CMD), characterized by impaired regulation of blood flow in the coronary microcirculation, plays a key role in the pathogenesis of ischemic heart disease and is increasingly recognized as a contributor to adverse cardiovascular outcomes. Despite its clinical importance, CMD remains underdiagnosed due to the reliance on invasive procedures such as pressure wire-based measurements of the index of microcirculatory resistance (IMR) and coronary flow reserve (CFR), which are costly, time-consuming, and carry procedural risks. To date, no study has sought to quantify CMD indices using data-driven approaches while leveraging the rich information contained in coronary angiograms. To address these limitations, this study proposes a novel data-driven framework for inference of CMD indices based on coronary angiography. A physiologically validated multi-physics model was used to generate synthetic datasets for data-driven model training, consisting of CMD indices and computational angiograms with corresponding contrast intensity profiles (CIPs). Two neural network architectures were developed: a single-input-channel encoder-MLP model for IMR prediction and a dual-input-channel encoder-MLP model for CFR prediction, both incorporating epistemic uncertainty estimation to quantify prediction confidence. Results demonstrate that the data-driven models achieve high predictive accuracy when evaluated against physics-based synthetic datasets, and that the uncertainty estimates are positively correlated with prediction errors. Furthermore, the utility of CIPs as informative surrogates for coronary physiology is demonstrated, underscoring the potential of the proposed framework to enable accurate, real-time, image-based CMD assessment using routine angiography without the need for more invasive approaches.

翻译：冠状动脉微血管功能障碍（CMD）以冠状动脉微循环血流调节受损为特征，在缺血性心脏病的发病机制中起关键作用，且日益被认为是导致不良心血管结局的重要因素。尽管具有重要临床意义，但由于目前依赖基于压力导丝的微循环阻力指数（IMR）和冠状动脉血流储备（CFR）等有创测量手段（这些方法成本高、耗时长且具有操作风险），CMD的诊断仍显不足。迄今为止，尚无研究尝试利用数据驱动方法，同时结合冠状动脉血管造影所蕴含的丰富信息来量化CMD指标。为应对这些局限，本研究提出了一种基于冠状动脉血管造影的新型数据驱动框架，用于推断CMD指标。研究采用一个经过生理学验证的多物理场模型来生成用于数据驱动模型训练的合成数据集，该数据集包含CMD指标以及具有相应对比剂强度曲线（CIPs）的计算血管造影图像。开发了两种神经网络架构：用于IMR预测的单输入通道编码器-MLP模型和用于CFR预测的双输入通道编码器-MLP模型，两者均融入了认知不确定性估计以量化预测置信度。结果表明，当基于物理的合成数据集进行评估时，数据驱动模型实现了较高的预测准确性，且不确定性估计与预测误差呈正相关。此外，研究证明了CIPs作为冠状动脉生理学信息替代指标的实用性，凸显了所提框架的潜力：无需采用更具侵入性的方法，即可利用常规血管造影实现准确、实时、基于图像的CMD评估。