Content area
Coronary Artery Disease (CAD) is the leading cause of death worldwide. Its clinical diagnosis typically involves two imaging steps. First, a non-contrast Computed Tomography (CT) scan is performed to assess coronary artery calcifications and perivascular adipose tissue, providing an initial stratification of cardiovascular risk. This is followed by a contrast-enhanced CT Angiography (CTA), which allows for detailed visualization of blood flow and precise characterization of stenosis.
While CT and CTA are highly complementary and jointly interpreted in clinical settings, their integration is currently limited to side-by-side visual inspection. Accurate spatial alignment of both modalities would enable richer diagnostic insights, for example, localizing calcifications to specific coronary branches and correlating them with perfusion deficits observed in CTA. However, achieving this alignment is challenging due to technical differences between scans and patient motion during separate acquisitions.
Although previous studies have attempted multimodal alignment using classical iterative methods, the registration performance has been poorly evaluated and remains potentially unreliable.
This dissertation proposes a benchmark of deep learning-based 3D deformable multimodal image registration frameworks for aligning CT and CTA scans, using data from a private and a public dataset, totaling 522 CT/CTA pairs. The methods were comprehensively evaluated at both the global level and, on a limited dataset of 6 patients, at a fine anatomical scale to assess their potential for enhancing CT-based CAD assessment.
Global alignment was assessed using similarity and segmentation-based overlap metrics across major thoracic structures visualized in CT and CTA. All models showed strong performance, with statistically significant improvements from the input, albeit with limited absolute changes. The best performing models were LessNet, which achieved a Mutual Information (MI) score of 0.335 ± 0.094, and HyperMorph, which achieved mean values of Dice and Mean Surface Distance (MSD) scores of 0.866 ± 0.102 and 1.254 ± 1.305 mm, respectively, across cardiac structures.
In contrast, none of the methods achieved reliable alignment of coronary arteries. Across all methods and patients, the average Dice and centerline-Dice (clDice) scores for coronary segmentations in CT and the corresponding aligned CTA were 0.454 ± 0.055 and 0.435 ± 0.087, respectively. The average centerline distance between structures was 4.065 ± 1.705 mm, revealing high variability and insufficient precision for clinical use.
This thesis highlights the current limitations of multimodal registration techniques for coronary alignment while demonstrating their utility in global cardiac anatomy alignment. It also outlines future directions for improving coronary-specific registration to enable fully integrated and vessel-aware CAD assessment.