A flow-matching method for rebuilding low-count PET images produced lower reported noise than MLEM, MAPEM and DDS, while showing visually higher contrast than PET-FlowDPS, according to an arXiv preprint.
At the 5% count level, the more severe of the two simulated low-count conditions, the proposed method preserved image contrast and structural details and recovered putamen uptake comparable to the full-dose reference. PET-FlowDPS underestimated putamen uptake.
The result is a comparison of reconstructed images, not a clinical outcome study. The evaluation offers algorithmic image-quality evidence; it does not establish diagnostic accuracy, patient benefit or clinical utility.
A learned guide for reconstruction
Flow matching is used here as a learned guide for the reconstruction. The proposed framework starts with a pretrained flow-matching prior, then combines that prior with refinement against the measured PET data and stochastic propagation inside an approximate Bayesian framework. In ordinary terms, the method uses learned image structure while still updating the image with the counts recorded by the scanner.
PET-FlowDPS incorporated Poisson-likelihood guidance with an expectation-maximization preconditioner. The proposed model combined a flow-based prior, PET-data refinement and stochastic propagation, while DDS used the same training datasets and network architecture as the corresponding flow model.
The test used simulated count loss
The primary test set comprised 10 clinical [18 F]FDG PET datasets from the Monash DaCRA fPET–fMRI dataset. Pretraining used 116 PET datasets, with 110 assigned to training and 6 to validation.
To create the low-count test conditions, the analysis simulated data at 10% and 5% of the original counts. The 10% setting was the primary evaluation, the 5% setting tested more severe count loss, and full-dose data supplied the reference images.
The study judged regional performance using mean uptake in the putamen and the coefficient of variation, or CoV, in white matter. CoV served as the noise measure, capturing how much values varied in that region relative to their average.
The analysis also examined bias and spatial standard deviation across low-count reconstructions. Here, bias means the average departure from the full-dose reference, while spatial standard deviation describes how much the reconstructed values varied across the image.
Noise fell, but the detail comparison mattered
At 10% of the original counts, DDS, PET-FlowDPS and the proposed method recovered putamen uptake comparable to the full-dose reference. The proposed method kept bias comparable to MLEM and MAPEM while achieving lower spatial standard deviation; PET-FlowDPS showed relatively increased spatial bias.
On the image-quality measures, the proposed method and PET-FlowDPS were reported as less noisy than MLEM, MAPEM and DDS. The proposed method also showed greater visual contrast than PET-FlowDPS in the reported images.
At 5%, the proposed method retained contrast and structural details while matching full-dose putamen uptake, whereas PET-FlowDPS underestimated that uptake. This was the more demanding simulated count condition in the evaluation.
These comparisons were descriptive. No confidence intervals or formal significance tests were reported, so the preprint does not quantify uncertainty around the differences.
Settings and speed shaped the result
The numerical settings affected the reported trade-off. In the sensitivity analysis, PSNR, a similarity score against the full-dose reference, peaked at K=50 flow-sampling steps and decreased after that. Favorable trade-offs were observed at 50 and 75 steps, and N=10 EM iterations was selected for subsequent experiments.
The selected configuration used 50 flow-sampling steps, 10 penalized EM iterations at each time point and a penalty strength of β=0.1.
Speed fell between the comparison methods in the reported timing. PET-FlowDPS took 0.40 minutes, the proposed method 1.85 minutes and DDS 3.93 minutes. The authors did not report timing uncertainty or broader hardware-normalized comparisons.
Where the evidence stops
The evidence has a narrow base: 10 clinical brain PET datasets and only the 10% and 5% simulated count levels. Performance was not established for larger cohorts, different scanners or other radiotracers.
Nor were the full-dose references noise-free ground truth. Those images contained statistical noise and correction errors, which means the comparison was against reconstructed references with their own imperfections.
Hyperparameters may need adjustment under other conditions. The probabilistic formulation also relies on approximations, so the outputs should not be treated as exact posterior samples. In addition, the effects of flow-sampling steps and inner EM iterations cannot be cleanly separated because total EM iterations increase as the number of sampling steps rises.
Promising, but still preliminary
The authors interpret the results as a promising use of flow matching as a PET generative prior, with a favorable balance between quantitative accuracy and denoising. They conclude that the proposed flow-based method improved image quality relative to MLEM, MAPEM, DDS and PET-FlowDPS.
They also suggest that the method’s MAP-based refinement may help explain its higher contrast, lower spatial bias and lower noise compared with PET-FlowDPS. That is an interpretation of this computational evaluation, not a demonstrated clinical effect.
The document is an arXiv preprint identified as arXiv:2608.20112v1 and dated 20 Aug 2026. The work was supported by NIH grants R01EB034692 and R01AG078250.
Paper data and sources
Original title: Flow Matching-Based PET Image Reconstruction
Authors: Fumio Hashimoto, Ziqian Huang, Tatsuya Yokota, Kuang Gong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text