Deep learning-based real-time prediction of coronary artery blood flow pressure from computed tomography angiography

The deep learning-based high-performance computational fluid dynamics (CFD) solution method is currently a hot, frontier topic in hemodynamic research. However, accurate predictions of the flow field with different coronary geometries and boundary conditions remain challenging. Given this, this study proposes a method based on deep learning and coronary computed tomography angiography (CTA) that achieves rapid and accurate solutions for blood flow pressure. We established a dataset based on patient-specific data from 370 patients and proposed a deep learning model with dual encoding of boundary condition and geometry. The model inputs boundary conditions obtained by patient-specific physiological parameters and coronary artery geometric information achieved by coronary CTA to iteratively predict the blood flow pressure along the centerline of the coronary artery in real-time. Statistical analysis was performed to evaluate the efficacy of the method by comparing it with CFD simulations. Testing on 112 cases, the root mean square error (RMSE) was 4.34% compared to the blood flow pressure obtained by CFD simulations. The computational efficiency of predictions using the trained deep learning model has improved by 180 times compared to CFD simulations (10 s VS 0.5 h). The proposed method in this study can provide accurate, real-time predictions of blood flow pressure for different coronary geometries and boundary conditions, which significantly improves computational efficiency and reduces costs while maintaining a high level of calculation accuracy.