VascuSR: iterative 3D vascular guidance restores topological fidelity in CT super-resolution

Coupling a 3D rectified flow vascular refiner with a slice-wise diffusion decoder resolves vessel hallucination and fragmentation, reducing disconnected vascular components by over 20%.

Primary source: arXiv:2610.20509: https://arxiv.org/abs/2610.20509

What’s new: Generative super-resolution models can restore high-frequency contrast in low-dose computed tomography (CT), but 2D slice-wise architectures frequently synthesize anatomically invalid structures: creating broken vessels, spurious bifurcations, or phantom connections across slices. In VascuSR: Iterative Anatomy-Aware Diffusion for CT Super-Resolution (arXiv:2610.20509, 14 Oct 2026), researchers from UCSD, University of Michigan, and MIT CSAIL introduce an iterative diffusion framework that explicitly decouples 3D vascular topology refinement from 2D appearance synthesis.

VascuSR operates through a dual-representation feedback loop. Starting from a preliminary super-resolved volume, an external vessel segmenter extracts a coarse 3D vascular representation. Because modeling raw 3D CT intensities is computationally prohibitive and prone to hallucinated texture, a lightweight 3D conditional rectified-flow refiner operates purely on the extracted vascular geometry, repairing discontinuities and filtering noise over ten Euler steps. A multi-scale anatomy controller then modulates intermediate feature maps of the 2D diffusion decoder, injecting refined 3D structural guidance back into the slice-wise reconstruction. As the super-resolved volume sharpens across feedback rounds, the 3D vascular prior updates dynamically. On the Medical Segmentation Decathlon (MSD Task08 hepatic vessel) and High-resolution Vascular Morphometry (HVM) datasets, VascuSR maintains standard pixel fidelity while dramatically improving topological metrics. It outperforms interpolation and learning-based super-resolution baselines in volumetric Dice and connectivity-aware clDice, reducing class-averaged disconnected component counts by over 20% on both cohorts and cutting mean vessel centerline distance by over 10% on MSD.

Why it matters: In contrast-enhanced clinical CT, surgical planning and stenosis grading depend on faithful vessel connectivity rather than sharp, hallucinated edges. Injecting explicit 3D topological priors into 2D diffusion decoders bridges the gap between computationally efficient slice-based generative models and the strict anatomical consistency required for clinical intervention.

Caveats: The framework relies on an initial vessel segmentation model to extract the coarse vascular topology, meaning severe acquisition artifacts that completely obscure large vessels could misdirect the 3D refiner. Performing slice-wise 2D diffusion alongside iterative 3D rectified-flow refinement increases overall inference latency compared to single-pass 2D super-resolution baselines. Finally, while validated on hepatic and vascular CT, performance on complex, highly tortuous microvasculature or coronary CTA remains exploratory.