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

VascuSR. Iterative Anatomy-Aware Diffusion for CT Super-Resolution. By Hairong Ni and four others, from the University of California, San Diego, the University of Michigan, and one other institution. Published on the fourteenth of October. No code repository is linked to it.

Generative super-resolution can bring back sharp detail in computed tomography images even from heavily downsampled observations. The catch is that the missing high-frequency content is not uniquely determined by the low-resolution input. Many sharp images fit the same blurry one. So a reconstruction can look plausible and still contain vascular detail the anatomy does not support: vessels missing, vessels that should not be there, or vessels connected the wrong way. The trouble lies in the word plausible. These are errors that look right, so the sharpness of the result is no evidence that its vessels are correct.

VascuSR is an iterative, anatomy-aware diffusion framework that pairs image reconstruction with explicit 3D guidance about the vessels. Rather than trusting the image generator to get vessels right, it gives the generator a vascular structure to follow, and keeps refining that structure.

The first ingredient is an explicit 3D vascular representation, which captures structural information beyond any single slice. It also makes refinement simpler, since a lightweight 3D generative refiner can attend to vascular structure without having to model the complex appearance of CT. The representation does two jobs, then: it carries anatomy across slices, and it narrows what the refiner has to learn.

The second ingredient connects that structure to the slice-by-slice super-resolution. A lightweight anatomy controller injects the refined structure into the diffusion decoder. Then the loop closes: each new super-resolved volume feeds back to update the guidance for the next pass. Structure guides the images, the images refine the structure, and the two are brought into agreement over iterations, which is where the word iterative in the title comes from.

Experiments use re-annotated versions of the MSD Task08 and HVM datasets. VascuSR beats interpolation and learning-based super-resolution baselines on segmentation accuracy, measured by Dice and clDice, and on 3D vascular structure, measured by connectivity and mean centerline distance. Image quality stays comparable to the super-resolution backbone it builds on. Against learning-based baselines, it cuts class-averaged connected-component counts by more than twenty percent on both datasets, and mean centerline distance by more than ten percent on MSD.

The framing is that sharpness and anatomical correctness are separate goals, and a sharp image can fail the second. VascuSR keeps image quality roughly where its backbone left it, and spends its effort on making the vessels connect the way they should.