Physics-Informed Implicit Neural Networks for 3D Coronary Artery Reconstruction.
Authors
Affiliations (4)
Affiliations (4)
- Unit of Medical Technology and Intelligent Information Systems, Dept. of Materials Science and Engineering, University of Ioannina, 45110, Ioannina, Greece.
- Electronics and Electrical Communication Engineering Dept., Faculty of Engineering, Tanta University, Tanta, Egypt.
- Unit of Medical Technology and Intelligent Information Systems, Dept. of Materials Science and Engineering, University of Ioannina, 45110, Ioannina, Greece. [email protected].
- Biomedical Research Institute, Foundation for Research and Technology-Hellas, 45110, Ioannina, Greece. [email protected].
Abstract
Three-dimensional (3D) reconstruction of coronary artery geometry from computed tomography angiography (CCTA)-derived segmentations is challenging because of thin vessel walls, tortuous trajectories, and small distal structures. We propose a physics-informed implicit neural network (PIINN) for case-specific reconstruction of paired coronary surfaces from pre-segmented CCTA-derived lumen and outer-wall geometries. For each coronary segment, a single network is optimized independently to represent the luminal and outer vessel-wall surfaces using dual signed distance fields (SDFs). The objective combines supervised SDF regression and boundary anchoring with Eikonal regularization and geometry- and anatomy-informed soft constraints on surface ordering, enclosure consistency, and local wall separation. A three-stage curriculum progressively adjusts the spatial sampling, learning rate, and constraint weights. The framework was evaluated on 15 coronary artery segments. The PIINN (256 × 6) achieved a near-surface RMSE of 0.26 mm and a sign accuracy approaching 90% for the outer SDF. Mesh reconstruction achieved an inner-surface ASSD of 0.166 mm and an HD95 of 0.455 mm, compared with the ASSD values of 0.290 mm for Vanilla INR and 1.057 mm for DeepSDF, with statistically significant improvements across the evaluated cases (p ≤ 0.003). Component-wise ablation further demonstrated distinct contributions from Eikonal regularization, anatomical constraints, region-based sampling, and curriculum optimization. These results support PIINN as a constrained continuous representation for case-specific coronary reconstruction from CCTA-derived segmentations; direct reconstruction from raw CCTA and generalization to unseen patients were not evaluated.