FORCE-Interior: Poisson-flow generative prior cuts CT truncation artifacts
New generative framework reconstructs interior CT scans from severely truncated projections with unprecedented clarity.
Interior tomography reconstructs a region of interest (ROI) from truncated X-ray projections, but truncation makes the problem severely ill-posed, causing non-unique solutions and artifacts. Existing learning-based methods struggle to generalize across different truncation patterns and noise levels. Current generative models are not designed for interior settings and may produce anatomically plausible but measurement-inconsistent outputs.
To solve this, Kang Chen and colleagues introduced FORCE-Interior, a Poisson-flow generative framework that anchors reconstruction to measured projections via a full-FOV-constrained initialization and per-step data consistency. Experiments demonstrate superior reconstruction quality at high truncation levels, with competitive results at lower truncation. This approach promises more reliable CT imaging in applications requiring localized high-resolution scanning, such as dentistry or oncology.
- FORCE-Interior uses a Poisson flow generative prior combined with measurement-constrained initialization to handle severe projection truncation.
- Achieves superior structural and perceptual quality at two heavily truncated ROI sizes while maintaining projection-domain consistency.
- Outperforms existing learning-based and generative methods in generalizing across different truncation patterns and noise levels.
Why It Matters
Enables high-quality interior CT reconstruction from limited data, improving diagnostic imaging in dentistry, oncology, and interventional radiology.