Data Augmentation Using Physics-Informed Plaque Growth Simulation

This project addresses the challenge of limited and imbalanced datasets in coronary artery calcium (CAC) analysis, particularly for rare high-risk cases such as high Agatston scores. Existing augmentation methods rely on heuristic or purely data-driven approaches that lack physiological realism. We propose a physics-informed generative framework for synthetic CAC data creation. Building on a validated atlas-based registration pipeline, the project models plaque formation as a stochastic process governed by latent hemodynamic factors. A reduced-order physics-informed neural network (PINN) will estimate endothelial shear stress (ESS), which will guide a stochastic differential equation (SDE)-based model to simulate realistic plaque growth patterns. The system will generate anatomically and physiologically plausible calcium distributions, with controllable parameters such as Agatston score, vessel territory, and lesion phenotype. Deliverables include a modular Python-based generation pipeline, a synthetic dataset of 300–500 scans covering long-tail cases, and a comprehensive evaluation using statistical, structural, and expert-based validation metrics. This work shifts CAC data augmentation from heuristic placement toward physics-driven simulation, enabling more robust and generalisable machine learning models for medical imaging.