Generative AI-Assisted Novel View Synthesis of Coronary Arteries for Angiography

Rotational cardiac angiography, though a useful imaging technique, has certain limitations. It can be time-consuming to acquire the images and may require expertise and special equipment for image acquisition. To overcome that, three-dimensional (3D) image reconstruction, specifically the novel view synthesis method, can help. One such method is Neural Radiance Fields (NeRF), which has gained popularity as a robust method in 3D scene reconstruction and generating novel views. In this study, we apply MedNeRF, an advanced generative neural radiance field architecture, to a novel dataset of 800 rotational coronary angiograms—a scale not previously explored. MedNeRF's design allows it to learn from multiple scenes simultaneously, a significant advantage over traditional models that focus on single-scene optimization. This feature is particularly beneficial for our dataset, where traditional NeRF techniques fail. Our exploration reveals that while MedNeRF has shown promising results in CT scans, applying it to the higher resolution and dynamic nature of X-ray angiograms presents unique challenges. To bridge this gap, we introduce preprocessing methods tailored for X-ray angiography, such as MSRCP and MSRCR, which enhance MedNeRF's qualitative output, highlighting the nuanced details essential for accurate medical analysis. However, we identified key limitations: MedNeRF's original configuration struggles with the high resolution of our data
$(\mathbf{512}\times \mathbf{512})$
and lacks the ability to capture temporal changes, crucial for fully understanding the dynamic processes in angiography. Addressing these challenges, our study pioneers in setting benchmarks for the application of NeRF technologies in this domain, demonstrating the need for integrating dynamic components into generative NeRF models to better suit medical imaging tasks. Our findings not only provide a foundational framework for utilizing NeRF in analyzing complex angiographic data but also open pathways for future research. We suggest the exploration of models that merge generative, and dynamic NeRF architectures to create more adaptable and detailed imaging solutions. This work paves the way for significant advancements in medical imaging, with the potential to enhance diagnostic accuracy and patient care by enabling clearer and more detailed visualization of angiographic data.