Automatic Model Construction for Patient-Specific Aortic Flow Simulations Using Geometric Deep Learning
DescriptionImage-based computational fluid dynamics (CFD) provides comprehensive hemodynamic flow information and has hence been widely used in cardiovascular disease diagnosis. Reliable CFD simulation result requires accurate reconstruction of the geometry from medical images such as computerized tomography or magnetic resonance imaging. The traditional reconstruction approach involves manual operation which induces artifacts and is time-inefficient. Recent development in artificial intelligence inspires a series of works applying machine learning algorithms or deep neural networks (DNN) to automate this process, where 2-D or 3-D DNNs are used to map from medical image data to the true geometry. However, those methods either are not fully automated (e.g., require centerlines or manual pre/post-processing) or suffer from surface irregularity, especially for complicated geometries (e.g., the vascular tree vessel structure). To fill the gap, we proposed an automatic geometry reconstruction algorithm for patient-specific aortic flow simulation. The model uses a 3-D U-net with a shape stream module to predict a pixel volume from raw medical images, from which a template geometry is reconstructed and deformed to the true geometry using a graph deformation module. After training on the Vascular Model Repository (VMR) dataset, our model outperforms state-of-art models and has great generalizability across different patients.
TimeTuesday, June 2716:00 - 16:30 CEST
LocationSanada I
Event Type
Life Sciences