Efficient Flow Matching for Sparse-View CT Reconstruction

A research team including Jiayang Shi, Lincen Yang, and K. Joost Batenburg has published a new paper on arXiv introducing FMCT and its efficient variant, EFMCT. These frameworks apply Flow Matching (FM) models to the critical problem of sparse-view Computed Tomography (CT) reconstruction. The work addresses a major bottleneck in medical imaging: traditional AI methods using Diffusion Models (DMs) rely on stochastic processes that create noise and require many computational steps, slowing down reconstruction in time-critical clinical settings. In contrast, FM uses a deterministic Ordinary Differential Equation (ODE), producing smooth trajectories that are naturally compatible with the repeated data consistency operations needed for accurate CT images.

The core technical breakthrough is the observation that velocity fields predicted by the FM model are highly correlated across adjacent steps. The team's EFMCT framework exploits this by reusing previously predicted velocity fields over consecutive steps, dramatically reducing the number of required Neural network Function Evaluations (NFEs). The paper includes theoretical analysis proving the error from this reuse is bounded. Experiments show FMCT/EFMCT achieve reconstruction quality competitive with state-of-the-art diffusion-based methods but with substantially improved computational efficiency. This advancement could enable faster diagnostic imaging and more practical AI deployment in hospitals. The researchers have open-sourced the codebase, allowing for further development and validation in real-world medical imaging pipelines.