Vascular CT super-resolution stops chasing pixels: vessels get their own 3D label space, and broken portal veins reconnect

Vascular CT super-resolution can now be judged and improved by three-dimensional vessel structure, not only by how plausible the image looks. Thin vessels are poorly determined by a low-resolution input, so one low-resolution image is consistent with several plausible high-resolution outcomes, and a visually natural reconstruction may still miss vessels, add spurious ones, or connect them incorrectly. Earlier super-resolution routes either mapped low resolution directly to high resolution, enforced agreement between reconstruction and the low-resolution input in image or measurement space, or fine-tuned a generator with segmentation losses; those improve pixel or segmentation scores without directly constraining continuity and branching in 3D. On two public contrast-enhanced CT datasets, re-annotated MSD Task08 and HVM, VascuSR scores higher Dice and clDice than bicubic interpolation, RCAN, SwinIR, HAT, ResShift, and a ResShift variant fine-tuned with Dice and soft-clDice losses, while its PSNR and SSIM stay close to the ResShift backbone and its LPIPS is best on MSD.

The approach learns vascular structure separately and feeds it back as a condition on image reconstruction. Each round extracts vessel probabilities from the current super-resolved volume, refines them into a binary vascular condition, and lets the diffusion decoder generate the next volume under that condition. A conditional 3D rectified-flow model performs the vascular correction by learning a path from noisy imperfect vessel predictions toward manual annotations, and a lightweight multi-scale anatomy controller turns the refined 3D masks into per-level conditioning features that are added to the decoder features. Feedback rounds are set to 3 on MSD and 2 on HVM; ablations show Dice and clDice improve with feedback and largely stabilize after two to three rounds, while a zero-mask setting that keeps the controller parameters but removes anatomical conditioning does not consistently beat ResShift. The added controller and FlowSeg3D carry modest parameter overhead relative to ResShift, and with eight parallel trajectories a three-round configuration processes one volume in about 6 minutes.

The reported scope is hepatic vascular contrast-enhanced CT: the re-annotated MSD Task08 and HVM public datasets, evaluated on image quality, vascular overlap, and three-dimensional vascular structure, with slice-wise inference that takes about 6 minutes per volume in the three-round configuration, which the authors place as acceptable for offline volumetric analysis. The authors give exploratory results on coronary CTA from ImageCAS and state that larger-scale evaluation across diverse vascular systems remains necessary.

The authors note that the framework performs super-resolution slice-wise and relies on a fixed vessel segmenter and a separately trained 3D refiner, which may limit volumetric consistency and let errors accumulate across feedback iterations. Ablations show Dice and clDice stabilizing after about two to three rounds, with later rounds fluctuating rather than improving consistently. The HVM manual annotations themselves contain isolated fragments and incomplete hepatic-vein coverage at the hepatocaval junctions, which explains the elevated component counts and the hepatic-vein largest-component fraction of about 0.60 on that dataset.