High performance neural network for solving coronary artery flow velocity field based on fluid component concentration

Rapid methods that can replace traditional inefficient computational fluid dynamics (CFD) for solving flow field are missing. We reconstructed three-dimensional (3D) coronary vascular tree models based on coronary computed tomography angiography (CCTA) images from 205 patients. Two fluid materials, blood and contrast agent, were mixed to simulate the flow field with concentration information under diverse boundary conditions, obtaining 2255 CFD simulations as deep learning samples. A dual-path physics-data multi-derived neural network (PDMNN) was designed, inputting geometric 3D point cloud and concentration information, respectively, and outputting 3D flow velocity field. Flow velocity in the coronary artery was clinically measured in 26 patients to verify the proposed PDMNN. For the 100 cases in a test set, the mean square error of the flow field velocity between the CFD calculations and the PDMNN predictions is 0.0309. However, the time taken by the PDMNN is significantly reduced (10 s VS 0.5 h). Clinically measured mean blood flow velocity and PDMNN predictions did not yield statistically significant differences (0.00 ± 0.05 m/s, P > 0.05). The proposed PDMNN present excellent computation accuracy and efficiency, holding a significant technical value for the clinical and engineering application.