Generative super-resolution can recover sharp detail in CT images, but the missing high-frequency content is not uniquely determined by the low-resolution input. As a result, a model may produce visually convincing reconstructions that contain vessels that are missing, spurious, or incorrectly connected. The authors of VascuSR argue that this makes standard SR unreliable for vascular imaging, where anatomical plausibility matters as much as pixel-level sharpness.
VascuSR addresses this by adding an explicit 3D vascular representation to a diffusion-based SR pipeline. A lightweight 3D generative refiner works on this vascular structure rather than on full CT appearance, and an anatomy controller injects the refined structure into the slice-wise diffusion decoder. The framework then iterates: each newly generated SR volume provides feedback that updates the vascular guidance for the next pass.
On re-annotated versions of the MSD Task08 and HVM datasets, VascuSR outperformed interpolation and learning-based SR baselines on Dice and clDice segmentation metrics, as well as on 3D vascular connectivity and mean centerline distance. The authors report that it reduced class-averaged connected-component counts by more than 20% on both datasets and mean centerline distance by over 10% on MSD, while keeping image quality comparable to its SR backbone.
The work is a step toward making generative SR usable in clinical contexts where the downstream task, not just the image, must be reliable. By steering reconstruction with explicit anatomy, VascuSR shows that domain knowledge can be folded into diffusion models without sacrificing image quality.