Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors
Authors: Hengjia Ran, Jie Luo, Yutao Zhu, Rui Hu, Fenglin Zhan, Jianyong Jiang, Huafeng Liu, Bo Zhao
Organizations: ZJU-UIUC Institute, Zhejiang University, Haining 314400, China · College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310027, China · College of Optical Science and Engineering, Zhejiang University, Hangzhou 310027, China · Vivo Mobile Communication Co., Ltd., Hangzhou 311100, China · Department of Nuclear Medicine, the First Affiliated Hospital of the University of Science and Technology of China, Hefei 230001, China · School of Physics and Astronomy and the Key Laboratory of Beam Technology of the Ministry of Education, Beijing Normal University, Beijing 100875, China · College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou 310027, China
Image reconstruction for positron emission tomography (PET) is an ill-posed Poisson inverse problem. In this work, we introduce an unsupervised, optimization-based reconstruction framework that employs a flow-matching generative model as a learned manifold prior. We train the generative model on high-quality PET images to learn a differentiable ordinary differential equation (ODE)-based transport from a Gaussian latent distribution to the PET image distribution. We incorporate this generator as an explicit manifold constraint into a regularized Poisson likelihood formulation. To solve the resulting problem, we develop an alternating direction method of multipliers (ADMM) algorithm that integrates an expectation-maximization (EM)-type surrogate update for Poisson data fidelity with gradient-based latent-space optimization. Numerical simulations and in vivo PET experiments demonstrate improved reconstruction accuracy compared with conventional reconstruction methods and recent deep learning-based approaches, particularly in low-count settings, with competitive computational efficiency.
Figures & tables
Fig. 1: Comparison of reconstruction methods at the 50% dose level in the BrainWeb simulation study. Two representative axial slices are shown, with the ground-truth and reconstructed images in the first and third rows and the corresponding relative error maps in the second and fourth rows. NRMSE, PSNR (dB), and SSIM are reported below each reconstruction.
Metric
Dose
Method
ML
TV
FBSEM
DIP
DPS
Proposed
NRMSE ( ↓ )
2%
24.77%
19.53%
16.91%
19.68%
16.24%
15.36%
5%
20.55%
16.92%
13.74%
17.06%
13.64%
12.73%
10%
17.84%
14.84%
12.72%
14.35%
12.09%
10.84%
25%
14.41%
12.27%
10.01%
11.25%
9.97%
8.55%
50%
12.21%
10.31%
8.29%
9.51%
8.33%
7.09%
TABLE I: Quantitative comparison of reconstruction methods at five dose levels, averaged over the middle 20 axial slices of the BrainWeb test subject. The best result for each metric at each dose level is highlighted in bold and the second-best result is underlined .
Fig. 2: Trade-off between tumor CRC and background STD in white-matter regions, evaluated over 10 independent noise realizations. The curves correspond to 15–50 iterations for ML, 1000–3100 iterations for TV, 5000–5400 iterations for DIP, data-consistency guidance weights ranging from 6 to 20 for DPS, and 6–15 iterations for the proposed method. FBSEM is shown as a single operating point corresponding to the output of the trained unrolled network.
Fig. 3: Empirical convergence and parameter sensitivity of the proposed algorithm. (a) Final NRMSE for different values of the ADMM penalty parameter ρ and latent regularization parameter λ . (b) NRMSE versus ADMM iteration for different numbers of EM-type updates in the Poisson data consistency subproblem. (c) NRMSE versus ADMM iteration for different numbers of L-BFGS iterations in the latent projection subproblem.
Fig. 4: Sensitivity of the proposed algorithm to initialization. NRMSE is plotted against ADMM iteration for zero, Gaussian random, uniform random, and ML-based initialization. It is clear that all four strategies approach nearly identical final NRMSE values.
Fig. 5: Effect of the number of Euler steps in the flow-matching ODE solver on reconstruction NRMSE. The NRMSE decreases sharply up to approximately 10 steps, with only marginal improvements thereafter.
Fig. 6: Comparison of reconstruction methods at the 25% dose level in the clinical-image-based simulation study. Representative axial, coronal, and sagittal views are shown from top to bottom, with the ground truth and reconstructed images followed by their corresponding relative error maps. NRMSE, PSNR (dB), and SSIM are reported below each reconstruction.
Metric
Dose
Method
ML
TV
FBSEM
DIP
DPS
Proposed
NRMSE ( ↓ )
2%
21.69%
17.27%
11.38%
17.08%
11.51%
8.80%
5%
17.50%
13.23%
9.70%
12.43%
9.14%
7.26%
10%
15.00%
11.18%
7.46%
9.76%
7.73%
6.27%
25%
12.44%
10.29%
6.15%
7.06%
6.09%
5.16%
50%
10.19%
8.44%
5.30%
5.73%
5.25%
4.51%
TABLE II: Quantitative comparison of reconstruction methods at five dose levels, averaged over 20 subjects in the clinical-image-based simulation study. The best result for each metric at each dose level is highlighted in bold , and the second-best result is underlined .
Fig. 7: Cross-scanner generalization evaluation using clinical-image-based simulated data. DPS and the proposed method are compared under matched (Siemens training / Siemens testing) and mismatched (United Imaging training / Siemens testing) settings. Representative axial, coronal, and sagittal views are shown from top to bottom, with corresponding relative error maps below each reconstruction. NRMSE, PSNR (dB), and SSIM are reported below each reconstruction. The proposed method exhibits less performance degradation under scanner mismatch than DPS.
Fig. 8: Comparison of reconstruction methods using in vivo PET data at the 10% count-level from an independent patient dataset. Representative axial, coronal, and sagittal views are shown from top to bottom. The full-count reconstruction is provided as a reference.
Fig. 9: Trade-off between mean caudate activity and background CoV in white-matter regions using 10% count-level in vivo PET data from 15 subjects. The full-count reconstruction is marked as the reference. The curves correspond to 9–30 iterations for ML, 500–4000 iterations for TV, 1800–2500 iterations for DIP, data-consistency guidance weights ranging from 10 to 45 for DPS, and 10–30 iterations for the proposed method.
Image reconstruction recovers images, volumes, or scenes from incomplete, noisy, indirect, or compressed measurements using physical models and learned priors.
Image reconstruction for positron emission tomography (PET) is a challenging inverse problem due to its inherent ill-posedness and Poisson counting noise, particularly in low-dose acquisitions. In this work, we introduce a novel Bayesian PET reconstruction framework that leverages a flow-matching generative prior learned from high-quality PET images. By incorporating this prior, we formulate PET reconstruction as a Bayesian inference problem and solve it via latent-space posterior sampling using the unadjusted Langevin algorithm. We evaluate the proposed method through numerical experiments and benchmark it against classical reconstruction methods and state-of-the-art deep learning-based baselines. The proposed method achieves improved reconstruction performance while enabling posterior uncertainty quantification.
Hengjia Ran, Huafeng Liu, Bo Zhao
ZJU-UIUC Institute, College of Information Science and Electronic Engineering, Zhejiang University, China · College of Biomedical Engineering and Instrument Science, Zhejiang University, China
Generative models have shown strong potential for positron emission tomography (PET) image reconstruction. Although diffusion model-based reconstruction methods have demonstrated promising performance, they often require many reverse sampling steps with data-consistency updates incorporated into the sampling process. Flow matching offers an attractive alternative because it can directly estimate clean images from intermediate states, allowing data-consistency refinement to be separated from flow propagation. In this work, we proposed flow matching-based PET image reconstruction methods. We first established PET-FlowDPS by incorporating Poisson likelihood guidance with an expectation-maximization (EM)-based preconditioner into the FlowDPS framework. We then proposed a model-based PET reconstruction method that used a pretrained flow matching model as a prior, in which the flow-based prior, PET data refinement, and stochastic propagation were interpreted within an approximate Bayesian framework. Experimental results using [18F]FDG brain PET datasets showed that the proposed method achieved better bias-variance trade-offs across different dose levels compared with other reference methods. These results demonstrated the potential of flow matching as a generative prior for quantitative PET image reconstruction.
Fumio Hashimoto, Ziqian Huang, Tatsuya Yokota +1
J. Crayton Pruitt Family Department of Biomedical Engineering, University of Florida, FL, USA · Department of Computer Science, Nagoya Institute of Technology, Aichi, Japan · RIKEN Center for Advanced Intelligence Project, Tokyo, Japan
Low-count Positron Emission Tomography (PET) reconstruction is severely hindered by the dissipative nature of prevailing generative models, where the inherent phase-space contraction leads to the numerical extinction (``wash-out'') of weak but diagnostically critical lesion signals. To overcome this geometric limitation, we propose FlowPET, a physics-informed framework that reformulates reconstruction as volume-preserving transport in a symplectic phase space. By parameterizing the posterior dynamics via a Separable Hamiltonian System, our approach guarantees a divergence-free vector field by construction, theoretically immunizing weak signals against probability mass collapse. To steer this conservative flow, we introduce conjugate boundary conditions based on the Range-Null space decomposition of the PET operator; this strictly enforces data consistency in the range space while confining stochastic uncertainty injection to the unobserved null space. We train the model via symplectic flow matching and perform inference using a symplectic leapfrog integrator. Extensive experiments on BrainWeb, clinical pediatric, and UDPET datasets demonstrate that FlowPET not only surpasses state-of-the-art deterministic and stochastic baselines in SSIM and PSNR but, more crucially, exhibits superior recovery of low-contrast lesions. The results confirm that imposing Hamiltonian structural constraints offers a robust geometric safeguard for medical inverse problems in high-noise regimes.
Zheng Zhang, Hao Tang, Yingying Hu +2
Centre for Smart Health, The Hong Kong Polytechnic University, Hong Kong, China · Department of Nuclear Medicine, Sun Yat-sen University Cancer Center, Guangzhou, China · Research Center for Medical AI, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China