基于深度学习的无创肺动脉血流储备分数快速计算方法研究

靶病变血管的选择及手术顺序是影响经皮腔内肺动脉成形术远期预后的关键。尽管侵入性血流储备分数（FFR）可降低再灌注肺水肿的发生，但因无法综合评估整个肺动脉树而不能规划多支病变血管的手术顺序。前期我们运用3D-0D耦合仿真方法验证了整个肺动脉树无创FFR计算的可行性。为克服仿真方法因计算时间长而难以广泛应用于临床的问题，本项目将建立一种基于深度学习的无创肺动脉FFR快速计算方法。首先为克服难以获取大量临床患者数据作为训练数据库的问题，通过算法人工合成大量的可覆盖临床常见病变生理状况的肺动脉3D解剖模型，再运用3D-0D耦合仿真方法批量生成FFR数据标签集。然后提出一种融合流体力学知识的深度学习网络架构，以将流体力学Navier-Stokes方程编码到深度学习算法中，获得高鲁棒性的学习模型。本项目提出的无创肺动脉FFR快速计算方法可实时评估整个肺动脉树的FFR分布，有望辅助临床优化治疗策略。

The target lesion vessel selection and procedure sequence are keys to the long-term outcomes of balloon pulmonary angioplasty. Fractional flow reserve-guided strategy increases the safety of the procedure as it reduces the extent of reperfusion edema. However, Invasive FFR cannot comprehensively assess all diseased vessels in the entire pulmonary artery tree. Thus it cannot plan the procedure sequence of multiple diseased vessels. Previously, we verified the feasibility of noninvasive pulmonary FFR calculation using 3D-0D coupled simulation method. However, the computational fluid dynamic simulation method is hard to be used widely in clinical practice due to the long computation time. The study will establish a fast prediction method of noninvasive pulmonary FFR based on deep learning to overcome the simulation method's shortcoming. Firstly, as it is impossible to obtain a large amount of clinical patient data as a training database, we will create a large database of synthetically generated pulmonary artery trees to reflect the anatomical variations representative of chronic thromboembolic pulmonary patients. And then the FFR data label set is generated in batch using 3D-0D coupled simulation method. Then, we will design a fluid mechanics informed deep neural network, in which the Navier-Stokes equations can be encoded in the outputs by using automatic differentiation. We will obtain a highly robust learning model. The fast computational method of noninvasive pulmonary FFR proposed in this study can evaluate the FFR value nearly in real-time. It is expected to assist clinical planning of procedure strategies for balloon pulmonary angioplasty and improve the patient’s long-term outcomes.