Diffusion-Based 3d Motion Estimation from Sparse 2d Observations

• DOI: 10.2139/ssrn.4673120
• Publication Date: 2023-12-22T07:18:28Z
• Authors: Niklas Gunnarsson; Thomas B. Schön; Jens Sjölund; Peter Kimstrand

Intra-interventional imaging is a tool for monitoring and guiding ongoing treatment sessions. Ideally one would like the full 3D image at high temporal resolution, this however not possible due to acquisition time. In study, we consider scenario when observations are sparse consist only of 2D slices through volume. Given 2D-2D registrations between predefined volume observations, propose method estimate motion. This motion enables reconstruction anatomy. Our relies on conditioning-based denoising diffusion model generates estimates given observations. We reduce dimensionality process by embedding data in lower dimensional space using principal component analysis. The evaluated two experiments: first synthetically generated then medical lung images. show that stable across entire within 1mm bound defined error.