New CT super-resolution method achieves 4x quality without paired training data

A team of researchers from Korea (Jeonghyun Noh, Hyun-Jic Oh, Won-Ki Jeong) has developed a novel zero-shot CT super-resolution method that reconstructs high-resolution 3D CT volumes from low-resolution inputs using diffusion models and 3D Gaussian splatting. The work, accepted early at MICCAI 2026, addresses a key clinical challenge: acquiring high-resolution CT scans is limited by radiation exposure. Traditional deep-learning super-resolution requires paired low-res/high-res datasets, which are rarely available. The new framework eliminates that need entirely.

The method operates in two stages. First, a diffusion model is trained on abundant X-ray projection data to upsample the low-resolution 2D projections, recovering fine details lost in the original scan. Second, the enhanced projections are used to reconstruct a 3D volume via 3D Gaussian splatting with a novel Negative Alpha Blending technique (NAB-GS), which models both positive and negative Gaussian densities to learn signed residuals. This allows the reconstruction to capture fine structural details that previous methods miss. On two public datasets, the framework achieved superior quantitative and qualitative results at 4x super-resolution, with expert evaluation confirming its clinical viability.