Novel View Synthesis of Coronary Arteries from Rotational X-Ray Angiography

Abstract

Coronary artery disease (CAD) is commonly diagnosed and treated using X-ray coronary angiography, which requires multiple angiograms and the administration of contrast dye. While effective, the overuse of contrast dye can lead to adverse effects in patients, particularly those with kidney issues. This thesis explores a novel approach aimed at minimizing the need for multiple angiograms and contrast dye by leveraging neural radiance field (NeRF) algorithms to generate novel views of coronary arteries. By training these algorithms on rotational angiography data, we seek to develop a generalizable model capable of producing detailed coronary artery images without requiring patient-specific retraining, potentially reducing contrast dye usage. The study systematically evaluates different NeRF architectures. MedNeRF, one of the initial models tested, demonstrated its ability to generate realistic coronary artery images. The right coronary artery (RCA) was easier to model due to its well-defined structure, while the left coronary artery (LCA), which had a larger dataset, was the primary focus of the experiments. Building on this, StyleNeRF, which overcame the limitations of MedNeRF by producing highly detailed and accurate coronary artery images with a Fréchet Inception Distance (FID) score of 75.32, the best among the models tested. StyleNeRF's hierarchical latent space enabled efficient generation of high-resolution novel views, and a web-based tool was developed to allow clinicians to interpolate target views at different angles. Lastly, dynamic NeRF architectures were explored to account for heart motion and the flow of contrast dye, addressing the dynamic nature of coronary artery imaging. However, the results from dynamic NeRF models, including attempts to merge RoDyNeRF with generative frameworks, were inconclusive, indicating the need for further development in this area. This research demonstrates the potential of NeRF-based novel view synthesis as a promising technique to reduce the reliance on contrast dye in coronary artery imaging. While the results are promising, challenges remain, particularly in handling dynamic scenes and real-time applications. Future work will focus on advancing generative models, such as latent diffusion and Gaussian splatting, to improve the model's real-time capabilities and its applicability in clinical settings.