Evaluation and optimisation of synthetic coronary artery models in training of DL model predicting cardiovascular haemodynamic variables.
Authors
Affiliations (6)
Affiliations (6)
- Department of Mechanical Engineering, University College London, London, UK.
- Centre for Cardiovascular Medicine and Devices, Queen Mary University of London, London, UK.
- Centre for Cardiovascular Medicine and Devices, Queen Mary University of London, London, UK; Department of Cardiology, Barts Heart Centre, Barts Health NHS Trust, London, UK.
- School of Electronic Engineering and Computer Science, Queen Mary University of London, London, UK.
- Satsuma Lab, UCL Hawkes Institute, University College London, London, UK; UCL Respiratory, University College London, London, UK.
- Department of Mechanical Engineering, University College London, London, UK. Electronic address: [email protected].
Abstract
The generation of high-fidelity synthetic patient data has gained popularity as a solution to challenges in training Artificial Intelligence (AI) methods, including data scarcity, patient privacy, and the need for scalable training datasets. This paper presents a framework for generating, evaluating, and optimizing synthetic coronary artery surface geometries from X-ray coronary angiogram using a variety of methods. Patient Data Augmentation (PDA), Morphologically Equivalent Synthetic Models (MESM), Conditional Tabular Generative Adversarial Networks (CTGANs), and Copulas are used to generate the synthetic data. A rigorous test harness comprising seven key metrics is introduced to evaluate each method, spanning data fidelity, utility, and privacy. While PDA offers the highest fidelity, it is computationally expensive and poses privacy risks. In contrast, the fully synthetic MESM approach is fast, secure, and scalable, though initially less realistic. To address this, we implement genetic algorithms that optimize MESM cohorts across the multiple dimensions in the test harness. Alignment with real patient data is improved by 14% with single-objective optimization and by 11% with multi-objective optimization. This paper introduces a system for producing an optimized, privacy-preserving synthetic dataset that maintains competitive performance in downstream deep learning tasks.