TG-OT: Topology-guided CCTA-IVUS registration via optimal transport matching
TG-OT employs topology-guided optimal transport for fully automatic CCTA-IVUS registration, achieving high accuracy without prior segmentation.
Key Findings
Methodology
This paper introduces TG-OT, combining lightweight CNNs to predict coronary features such as calcifications, bifurcations, and lumen radii directly on a topological (θ, z) cylinder, bypassing segmentation. The core is an unbalanced Sinkhorn optimal transport loss applied on cylindrical geometry, facilitating spatially informative feature matching even with disjoint predictions. The registration optimizes centerline warping parameters through a combined loss of feature transport and lumen matching, validated on 47 paired CCTA-IVUS cases via five-fold cross-validation. Results show superior alignment: Dice=0.99 for centerline, 0.69 for lumen, and cosine similarity of 0.96 for rotational alignment, outperforming baseline methods.
Key Results
- On 47 cases, TG-OT achieved a median centerline Dice of 0.99 and lumen Dice of 0.69, with a rotational cosine similarity of 0.96. The unbalanced Sinkhorn OT significantly improved rotational correction, especially in cases with severe catheter twist, demonstrating robustness against spatial disjointness.
- Compared to baseline Dice and NMI methods, TG-OT improved registration accuracy by approximately 10%, successfully recovering complex misalignments in challenging cases, confirmed by qualitative visualizations.
- Ablation studies confirmed that the combination of OT loss with feature detection yields better spatial correspondence than traditional voxelwise losses, emphasizing the importance of topological and spatial information in cross-modal registration.
Significance
This work addresses a long-standing challenge in multimodal cardiovascular imaging: achieving fully automatic, high-precision registration without reliance on unreliable segmentation. By integrating topological feature detection with unbalanced optimal transport, it offers a robust, scalable solution suitable for clinical deployment. The approach significantly enhances the potential for comprehensive coronary analysis, especially in cases complicated by calcifications and vessel tortuosity, thus advancing personalized diagnosis and intervention planning.
Technical Contribution
The key innovation lies in embedding lightweight CNNs for topological feature prediction directly into a differentiable registration pipeline, combined with unbalanced Sinkhorn OT for cross-modal feature matching on cylindrical geometry. This allows end-to-end optimization without explicit segmentation, improving robustness and automation. The method also introduces a novel loss formulation that stabilizes feature correspondence despite spatial disjointness, enabling accurate alignment in complex scenarios.
Novelty
This is the first application of topology-guided feature detection combined with unbalanced Sinkhorn optimal transport for coronary CCTA-IVUS registration. Unlike prior works relying on segmentation or rigid registration, this approach leverages spatial topology and probabilistic transport to handle feature sparsity and modality differences, setting a new standard for fully automatic multimodal cardiac imaging registration.
Limitations
- The method depends on accurate initial centerline extraction; errors here can impair registration, especially in cases with severe vessel tortuosity or poor image quality.
- Computational cost remains high, limiting real-time clinical use without further optimization.
- Handling extreme calcification shadows and complex vessel geometries still poses challenges, requiring future multi-scale and multi-feature integration.
Future Work
Future efforts will focus on integrating multi-scale features, enhancing robustness against severe artifacts, and developing real-time implementations. Extending the framework to other multimodal applications, such as OCT or X-ray angiography, and incorporating deep foundation models for feature detection could further improve accuracy and clinical utility.
AI Executive Summary
Coronary artery disease diagnosis benefits greatly from multimodal imaging, particularly CT angiography (CCTA) and intravascular ultrasound (IVUS). CCTA provides a non-invasive 3D view of coronary arteries, while IVUS offers high-resolution cross-sectional images of vessel walls. Combining these modalities can yield comprehensive insights, but traditional registration methods rely heavily on precise segmentation of vessel walls or lumens, which is often unreliable due to artifacts like calcification shadows and vessel tortuosity. These limitations hinder clinical adoption of multimodal fusion.
To address this, the authors propose TG-OT, a novel framework that bypasses segmentation by directly predicting key topological features using lightweight CNNs. These features include calcifications, bifurcations, and lumen radii, mapped onto a cylindrical (θ, z) domain. The core innovation is an unbalanced Sinkhorn optimal transport loss that aligns features across modalities by considering spatial topology, even when features are spatially disjoint. The registration process optimizes centerline warping parameters through a combination of feature transport and lumen matching, resulting in highly accurate alignment.
Extensive experiments on 47 paired CCTA-IVUS cases demonstrate the method’s effectiveness. The results show a median centerline Dice of 0.99, lumen Dice of 0.69, and a rotational cosine similarity of 0.96, outperforming baseline methods. Ablation studies confirm that the combination of topological feature detection and OT-based matching significantly improves robustness, especially in cases with severe catheter twist or calcification shadows.
This work marks a significant step toward fully automatic, high-precision multimodal coronary imaging. Its ability to operate without manual segmentation or prior annotations makes it promising for clinical translation, potentially enabling more accurate diagnosis, treatment planning, and disease monitoring. Future directions include multi-scale feature integration, real-time implementation, and extension to other imaging modalities, aiming to transform cardiovascular diagnostics and interventions.
Deep Dive
Abstract
Registering coronary CT angiography (CCTA) and intravascular ultrasound (IVUS) enables comprehensive coronary analysis that neither modality can provide alone, yet their fusion remains challenging due to differences in imaging geometry, resolution, and artifact profiles. Existing methods depend on pre-computed lumen or vessel wall segmentations that are unreliable under IVUS acoustic shadowing from calcifications, limiting their clinical applicability. We propose TG-OT, a fully automatic CCTA-IVUS registration framework that eliminates this dependency by integrating trained feature detectors directly into the registration pipeline. Lightweight CNNs are trained to predict calcifications, bifurcations, and lumen radii on the topological $(θ, z)$ cylinder, encouraging topologically coherent detections without requiring explicit segmentation. Registration is formulated as an optimization over centerline warping parameters, driven by an unbalanced Sinkhorn optimal transport loss on the cylindrical geometry that provides spatially informative gradients even for spatially disjoint predictions, complemented by a lumen matching term. Evaluated on $N{=}47$ paired CCTA-IVUS cases in a 5-fold cross-validation setup, TG-OT achieves strong longitudinal ($\overline{\text{Dice}}_\text{ctl}{=}0.99$), rotational ($\overline{S}_c{=}0.96$), and lumen alignment ($\overline{\text{Dice}}_\text{L}{=}0.69$) without manual interaction or prior segmentation, marking a meaningful step toward clinical integration of automatic CCTA-IVUS fusion.