Geometrically Constrained Stenosis Editing in Coronary Angiography via Entropic Optimal Transport

Researchers have developed OT-Bridge Editor, an AI method that uses optimal transport theory to synthesize realistic coronary angiography images with artificial stenosis lesions. The technique achieves 27.8% improvement in stenosis detection performance on benchmark datasets, addressing the critical shortage of high-quality medical imaging training data.

Medical AI systems struggle with scarcity of annotated coronary angiography data, limiting the deployment of automated stenosis detection tools in clinical settings. The OT-Bridge Editor tackles this by generating synthetic angiograms with geometrically accurate stenosis features, effectively expanding training datasets. Unlike conventional diffusion-based image editing that relies on soft guidance and produces imprecise results, this approach frames localized stenosis editing as a constrained entropic optimal transport problem, leveraging geometric constraints to maintain anatomical fidelity throughout the generation process.

The underlying innovation addresses a longstanding challenge in medical AI: data augmentation without sacrificing realism or diagnostic accuracy. Traditional synthetic data generation often introduces artifacts or fails to capture the complex spatial relationships in medical imaging. By grounding the generation process in geometric information, the method ensures that synthetic lesions appear structurally plausible, making them valuable training examples.

The empirical results demonstrate substantial clinical relevance. Testing on the public ARCADE benchmark yielded 27.8% relative improvement in detection precision, while multi-center validation produced 23.0% gains, suggesting the approach generalizes across different imaging protocols and patient populations. This consistency indicates the synthetic data successfully bridges distribution gaps that typically plague model generalization.

For the medical AI ecosystem, this represents progress toward broader clinical deployment of automated diagnosis tools. Better synthetic data generation reduces reliance on expensive, time-consuming manual annotation and helps address regional disparities in imaging dataset availability. The methodology could extend beyond coronary imaging to other modalities facing similar data scarcity challenges, making it strategically important for advancing clinical AI infrastructure globally.

• →OT-Bridge Editor uses entropic optimal transport to generate synthetic coronary angiography images with realistic stenosis lesions for training data augmentation.
• →Synthetic data improved stenosis detection by 27.8% on public benchmarks and 23.0% on multi-center datasets compared to models trained without synthetic augmentation.
• →The geometric constraint approach preserves anatomical structure better than traditional diffusion-based editing methods, reducing artifacts in medical imaging.
• →Data scarcity limits clinical translation of automated stenosis detection, making synthetic data generation a practical solution for expanding training capacity.
• →The technique demonstrates cross-center generalization, suggesting applicability across different imaging protocols and healthcare systems.