GenDiff: AI model slashes CT radiation by reconstructing ultra-low-dose scans
New diffusion model cuts radiation dose while preserving anatomical structures across unseen conditions.
GenDiff is a generalizable diffusion model designed for low-dose CT (LDCT) reconstruction. Traditional deep learning approaches are typically optimized for fixed dose levels or specific anatomical regions, limiting their clinical utility. The framework introduces a Dose-Anatomy Encoder that learns acquisition-aware embeddings from both radiation dose and anatomy. It then uses a dose- and anatomy-conditioned cold diffusion backbone for iterative refinement, coupled with a physics-consistency update that enforces fidelity to the CT forward model. A key innovation is the Structural Prior Refinement Module (SPRM), which preserves anatomical structures while suppressing dose-dependent artifacts like noise and streaks. The model was evaluated on multi-anatomy clinical datasets, unseen ultra-low-dose conditions, and out-of-distribution phantom and animal scans. Results show GenDiff consistently beats state-of-the-art CNN and diffusion-based methods in reconstruction quality while maintaining robustness across different dose levels, anatomical regions, and acquisition domains. This makes it a promising solution for practical low-dose CT imaging, potentially reducing patient radiation exposure without sacrificing diagnostic image quality.
- GenDiff jointly models continuous radiation dose and anatomy via a Dose-Anatomy Encoder and conditioned cold diffusion backbone.
- Achieves superior reconstruction quality on multi-anatomy clinical datasets, unseen ultra-low-dose scans, and out-of-distribution phantom/animal data.
- Outperforms state-of-the-art CNN and diffusion methods, demonstrating strong generalization across dose levels and anatomical regions.
Why It Matters
Enables safer CT scans with lower radiation while maintaining diagnostic image quality across diverse clinical scenarios.