Reproduce “2D Versus 3D Diffusion for In Silico Training of Interventional X-ray AI Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sampath Rapuri, Jeremy Ko, Benjamin D. Killeen, Russell H. Taylor, Mathias Unberath
A controlled comparison of diffusion-based synthetic data for interventional X-ray AI finds that directly generated 2D radiographs can support landmark detection as well as, or better than, CT-derived DRRs for high-confidence predictions. DiffXray reached 4.40 ± 1.87 mm mean error at an NCC threshold of 0.60, while additional viewpoint augmentation reduced its error at lower confidence thresholds; the 3D DiffCT-DRR approach remained limited by CT synthesis and segmentation quality.
Key insights
- Direct 2D synthesis can be effective when the downstream task is inherently image-plane based and anatomical pose can be explicitly controlled through segmentation conditioning.
- 3D generation preserves volumetric consistency and may better support multi-view tasks, but its usefulness depends strongly on the fidelity, resolution, and segmentability of generated CT volumes.
- Expanding the range of simulated viewpoints improves DiffXray performance, especially when lower-confidence landmark predictions are included.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.