One-Step Rectified Flow Reconstruction with Spectral Weighting for Limited-angle Cardiac CT
Limited-angle cardiac computed tomography (CT) suffers from severe directional artifacts and resolution degradation due to missing wedges in the frequency domain. Diffusion-based reconstruction methods have demonstrated strong artifact suppression capabilities but typically require iterative sampling at a high computational cost. To accelerate limited-angle CT reconstruction, we propose a rectified-flow framework trained with paired supervision for fast inference. Instead of generating images from Gaussian noise, the model learns a deterministic transport from artifact-contaminated Feldkamp-Davis-Kress (FDK) reconstructions to full-view reference images. A 2.5D slice-context modeling strategy is introduced to exploit inter-slice information while reconstructing only the central slice, improving structural consistency. Furthermore, we introduce a reconstruction-aware spectral loss that applies frequency-domain supervision to an approximate endpoint reconstruction inferred from the predicted flow velocity. By embedding spectral priors into the learning objective and using one-step rectified-flow inference (a single network evaluation), the proposed framework achieves fast reconstruction while preserving image fidelity. Experiments on XCAT simulations and clinical cardiac CT data show superior quantitative performance and substantially reduced inference time compared to the state-of-the-art baselines. A qualitative review by an experienced cardiac radiologist further suggests improved visualization of motion-sensitive coronary structures with reduced motion blur.
Authors
- Bahareh Morovati (ORCID: https://orcid.org/0009-0001-2923-8974)
- Shuyi Fan (ORCID: https://orcid.org/0009-0002-8780-6535)
- Changsheng Fang (ORCID: https://orcid.org/0009-0009-1778-7130)
- Hengyong Yu (ORCID: https://orcid.org/0000-0002-5852-0813)
- James Greg Terry (ORCID: https://orcid.org/0000-0001-6659-3441)
- Ge Wang (ORCID: https://orcid.org/0000-0002-2656-7705)
- L. P. Zhou (ORCID: https://orcid.org/0000-0001-6989-8080)
- Shuo Han (ORCID: https://orcid.org/0000-0002-6187-1307)
- Song Ni (ORCID: https://orcid.org/0009-0007-3914-9944)
- Haodong Li
- John Jeffrey Carr
Institutions
- University of Massachusetts Lowell (US)
- Rensselaer Polytechnic Institute (US)
- Vanderbilt University Medical Center (US)
Publication Details
- Journal
- IEEE Transactions on Medical Imaging
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1109/tmi.2026.3754715
- Primary Topic
- Cardiac Imaging and Diagnostics
- Type
- article
- Field-Weighted Citation Impact
- 0.00