Prior-informed deep learning for the analysis and fusion of cardiovascular imaging modalities

Cardiovascular diseases remain the leading global cause of death, with coronary artery disease being the most prevalent form. Medical imaging provides crucial insights for diagnosis and treatment, particularly through modalities like coronary CT angiography (CCTA) and intravascular ultrasound (IVUS). However, analyzing these complex 3D images is challenging and time-consuming, requiring expert interpretation.
This thesis explores how incorporating anatomical prior knowledge can enhance AI-based analysis of cardiovascular images. Rather than treating medical images as generic pixel arrays, the work leverages known cardiac anatomy - specifically the tubular structure of coronary arteries and predictable heart chamber shapes - to guide deep learning algorithms.
The research presents several key contributions: (1) A generative adversarial network for correcting suboptimal contrast enhancement in CCTA images, improving both visual quality and automated analysis; (2) Methods for automatic stenosis and plaque quantification using ray-casting and geometric priors to create accurate surface meshes of coronary structures; (3) A generalized framework for cardiac surface meshing using geometric templates and graph neural networks; (4) An accelerated deformable image registration technique for aligning multimodality cardiac images; and (5) An automated pipeline for registering CCTA and IVUS data using classification-guided centerline alignment.
Results demonstrate that geometry-informed deep learning significantly improves accuracy and robustness compared to conventional voxel-based approaches, particularly in low-data regimes. The methods enable more reliable automated assessment of coronary artery disease while maintaining anatomical plausibility. This work advances the integration of domain knowledge with modern AI techniques for enhanced cardiovascular image analysis.