An arXiv preprint reports that Flow Matching methods outperformed diffusion-based baselines in a computational test of sparse-view CT reconstruction. FlowDPS, one of the Flow Matching approaches, had the best overall reported performance, reaching a peak PSNR of 36.64 dB, an SSIM of 0.9260 and an LPIPS score as low as 0.1576. These are the image-quality measures reported for the reconstructed scans.
A comparison on sparse-view scans
The comparison asked whether a high-resolution Rectified Flow Matching prior could produce better reconstruction and image-generation results than diffusion-based priors under matched experimental conditions. The evaluation included FlowDPS, PnP-Flow, FLOWER and ICTM, alongside the diffusion methods DPS and DDRM, at 60, 90 and 120 projection angles.
The experiments used 3,306 acquisitions of human chests at 256 by 256 pixels from 10 patients. The paper separately reports an approximately 350-slice hold-out set for selecting reconstruction settings and a 350-image reference set for the generation test.
How the model was built
Flow Matching is described as learning a deterministic, continuous-time velocity field that transports samples from a base distribution toward the data distribution. It then generates a result by following that learned motion through an ordinary differential equation. Rectified Flow Matching is presented as a simplified version of this formulation and supplies the prior used in the reconstruction pipeline.
In the reconstruction formulation, CT measurements are treated as the output of a linear forward or acquisition operator applied to an unknown image, plus Gaussian measurement noise. The algorithm seeks a maximum a posteriori, or MAP, estimate: a feasible image favored by the model and consistent with the measured data.
The prior was trained in two stages. The initial stage used 500 epochs, a batch size of 4 and approximately 50 hours of training. A second stage fine-tuned the model with reduced augmentation for 100 epochs.
Several implementation choices were added to keep the reconstruction numerically stable. Image values were normalized by adding 1 and dividing by 2; gradients were scaled, values were clamped or otherwise controlled, and CGLS-based corrections were used to address range mismatch, large operator-induced gradients and mismatch between an operator and its adjoint. Guidance scales were selected separately for each method and angle setting by grid search on the approximately 350-slice hold-out set, using PSNR as the selection criterion.
The generation results
To assess unconditional generation, KID was calculated against 350 new Mayo images that had not been seen during training. The comparison included both training stages and a diffusion model at the same sampling steps. The Flow Matching sampler used explicit Euler integration with 5, 10, 20, 50 or 100 integration steps.
Flow Matching achieved a lower KID than the diffusion baseline at every tested step count, with the largest stated gap in the low-step regime. KID also fell by 10% to 20% between the first and second training stages across the tested step counts.
FlowDPS led the reconstruction comparison
On sparse-view reconstruction, the paper reports better image quality as the number of projection angles increased. DPS was the weakest diffusion baseline, DDRM was stronger, and the Flow Matching methods outperformed the diffusion baselines, with especially large gains on LPIPS.
Visual comparisons described FlowDPS as the most plausible reconstruction, with fewer streaking artifacts and better preservation of fine structures. Even so, both FlowDPS and the diffusion methods missed some extremely small, symmetric bone structures at the bottom of the slice.
An early computational result
The document is an arXiv preprint, version 1 dated 28 August 2026, and states that the trained Flow Matching model and accompanying code are publicly released. The findings describe a computational comparison using the stated chest CT data and the three tested projection-angle settings.
Paper data and sources
Original title: Physics-Guided Flow Matching for CT Image Reconstruction
Authors: Davide Evangelista
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text