3D coronary artery tree reconstruction from two non-simultaneous angiographic projections using deep learning

Thesis

3D reconstruction of coronary artery trees from invasive X-ray coronary angiography (ICA) remains challenging due to the limited number of available projections and the temporal mismatch between clinically acquired views, which introduces motion-related inconsistencies. This thesis aims to develop automated deep learning-based approaches for reconstructing 3D coronary artery trees from only two clinical ICA projections acquired from a dual-axis rotational single-plane C-arm system, while accounting for realistic acquisition constraints. By reducing reliance on multiple projections and manual interpretation, such approaches have the potential to lower radiation and contrast exposure and improve consistency in the assessment of coronary anatomy.

This thesis first shows that 3D coronary artery trees can be reconstructed from only two projections without requiring explicit 3D supervision, by leveraging self-supervised neural representations, demonstrating the feasibility of sparse-view reconstruction under idealised conditions without inter-projection motion. We then consider the more realistic setting of non-simultaneous clinical ICA projections and train on projections simulated from coronary computed tomography angiography data with added rigid transformations, enabling the model to handle inter-projection motion implicitly and generalise to real clinical acquisitions. Building on this, we introduce an explicit iterative motion correction framework that compensates for misalignment between projections via registration in the 2D projection plane, leading to more stable and robust reconstruction across a range of motion levels representative of clinical acquisitions.

Finally, we investigate the impact of representation on reconstruction and show that replacing voxel-based formulations with point cloud representations provides a more efficient and flexible approach for modelling sparse, complex coronary artery structures, enabling higher-resolution reconstruction with reduced computational cost. Overall, this thesis demonstrates that accurate and anatomically consistent 3D coronary artery tree reconstruction from only two clinically acquired projections is achievable under realistic imaging conditions, providing a principled learning-based framework with potential to reduce radiation and contrast exposure and improve consistency in clinical angiographic assessment.

Authors: Wang, Y

• DOI: 10.5287/ora-4r67eq9a2
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: iterative motion compensation; clinical non-simultaneous angiographic projections; transformer; self-supervised optimisation; generative adversarial network; point cloud-based adversarial learning; neural implicit representation; image registration
• Subjects: Deep learning (Machine learning)