Automatic geometric modeling of aortic valve and calcification using deep learning for biomechanical simulation

Accurate geometric reconstruction of the aortic valve system is essential for patient-specific biomechanical simulation and transcatheter aortic valve replacement planning. However, thin valve leaflets, irregular calcifications, and complex leaflet–calcification interfaces remain difficult to model automatically from cardiac computed tomography angiography. This study proposes a unified deep learning framework for automatic reconstruction of the aorta, valve leaflets, left ventricle, and calcifications. A shared 3D U-Net encoder extracts multi-scale anatomical features, while two complementary decoders perform template-based deformation for anatomically stable structures and signed distance function-based implicit reconstruction for calcifications. Joint geometric, mesh, and contact consistency constraints are introduced to preserve structural continuity and improve simulation readiness. Experiments showed that the proposed method achieved better surface accuracy, leaflet thickness consistency, calcification reconstruction, and volumetric mesh quality than representative baseline methods. The reconstructed models exhibited anatomically plausible leaflet–calcification interfaces, stable topology, and no obvious mesh inversion. Finite-element simulations further demonstrated the technical feasibility of using the reconstructed geometries for downstream biomechanical analysis. These results indicate that the proposed framework can generate accurate, anatomically coherent, and simulation-ready aortic valve models, providing a practical basis for automated biomechanical analysis and preoperative assessment in structural heart disease.