VascuSR: Iterative Topology-Aware Diffusion for CT Super-Resolution

Diffusion-based super-resolution can recover visually plausible CT detail, yet image-space reconstruction alone does not explicitly preserve the continuity and branching structure of fine vascular anatomy. We introduce **VascuSR**, an iterative anatomy-aware diffusion framework for $\times 8$ vascular CT super-resolution that couples image reconstruction with three-dimensional vascular refinement. Starting from an initial SR volume, VascuSR extracts hepatic- and portal-vein representations, refines them with a conditional 3D rectified flow, and feeds the refined anatomy back into the diffusion model to guide subsequent reconstruction. Across two public contrast-enhanced CT datasets, re-annotated MSD Task08 and HVM, VascuSR consistently improves vascular Dice and clDice over the matched ResShift backbone while maintaining comparable image-level reconstruction quality. On MSD, it further reduces the average number of connected vessel components by 46\%, indicating substantially improved vascular continuity. These results demonstrate the benefit of explicit volumetric anatomical feedback for structure-sensitive medical super-resolution.