PET Flow-Matching Prior

A PET scan acquired at 2% of the standard radiotracer dose records the same anatomy as a full-dose scan; what it loses is counts. Maximum-likelihood reconstruction amplifies that Poisson noise into streaks and speckle, and the classical fix — a hand-crafted smoothness or total-variation penalty — buys noise suppression by blurring the fine cortical structure a reader is looking for. Supervised networks do better, but they learn a mapping from paired low-dose and full-dose reconstructions, and that pairing must be re-collected for every scanner, protocol and tracer.

This model takes the prior out of the penalty function and puts it in a generator. A flow-matching network is trained once on high-quality PET images to regress a time-dependent velocity field, which defines a deterministic ODE transporting a Gaussian latent variable to the PET image distribution. Integrating that ODE gives a differentiable map from a latent vector to an anatomically plausible image. Reconstruction then becomes a Poisson likelihood problem constrained to the range of that map: the image must both explain the measured sinogram and be something the generator can produce. Researchers at Zhejiang University, working across the ZJU-UIUC Institute and three of its engineering colleges, released the work as a preprint in August 2026.

Because the generator is frozen, the only per-scan variable is the latent code. That is what separates this from a fit-per-scan regularizer: the same checkpoint reconstructs every study, and the paper demonstrates it on lesions and on a scanner the network never saw in training.

#Key Features

• Learned manifold constraint, not a penalty: The reconstruction must lie on the generator's image manifold, so structural plausibility is enforced by construction rather than traded against data fidelity through a tuning weight.
• Deterministic, differentiable sampling: Flow matching yields an ODE rather than a stochastic sampling chain, making the generator differentiable with respect to its latent so gradients come from autodiff.
• Ten to twenty function evaluations: A forward-Euler solver reaches stable quality in 10–20 steps, where diffusion posterior sampling needs on the order of 1,000.
• Split optimization via ADMM: A closed-form expectation-maximization surrogate handles data consistency and preserves voxel nonnegativity; an L-BFGS projection in latent space handles manifold proximity.
• Generalization beyond the training distribution: Lesions absent from training are recovered with the highest contrast recovery coefficient and lowest background standard deviation among the compared methods.

#Technical Details

The velocity field is a residual U-Net over a five-slice axial stack, with four encoder and four decoder stages joined by skip connections, residual blocks of two 3×3 convolution-BatchNorm-ReLU layers with 1×1 shortcut projections, base channel width 64 and encoder widths of 64, 128, 256 and 512. Training uses the conditional flow-matching loss on a linear Gaussian path, with Adam at a learning rate of 1×10⁻⁴, in PyTorch 2.8.0 on a single NVIDIA H100 NVL. Simulation work used 20 subject-realistic BrainWeb digital phantoms split 18/1/1, with tumors introduced only in the test case; clinical work used axial head slices from the Ultra-Low Dose PET Imaging Challenge 2025 cohorts — 372 patients on a Siemens Biograph Vision Quadra and 301 on a uEXPLORER — cropped to 128×128 and split 8:1:1 at the patient level. Projection data came from the parallelproj package, with randoms from delayed coincidences and scatter from openSSS.

Across 20 patients and five dose levels the method leads every baseline — maximum likelihood, total variation, the supervised unrolled FBSEM, deep image prior and diffusion posterior sampling. At the 2% dose it reaches 8.80% NRMSE, 33.81 dB PSNR and 0.955 SSIM against DPS's 11.51%, 31.47 dB and 0.918, in roughly 13 seconds per case over 10 ADMM iterations.

#Applications

The target is low-dose and short-acquisition PET, where cutting injected activity or scan time is desirable for pediatric imaging, longitudinal follow-up and total-body protocols but is normally paid for in image quality. Because no paired training data are required, the prior can be fitted to whatever high-quality images a site already holds and then applied to reduced-count acquisitions — useful in multi-center settings where scanners and reconstruction parameters differ.

#Impact

The work shows that flow matching, not only score-based diffusion, can serve as a learned image prior inside a physics-based inverse solver, and that its deterministic ODE formulation buys both differentiability and a two-orders-of-magnitude cut in generator calls — what makes generative-prior reconstruction practical at clinical turnaround times. The scope is narrow: 128×128 axial head slices, a single tracer, a multi-slice rather than fully volumetric generator, and a nonconvex problem with no convergence guarantee. The prior is learned from noisy high-quality reconstructions rather than true activity maps, and the authors state that broader validation across tracers, scanners, acquisition protocols and patient populations is required before clinical deployment. The model is unnamed in the preprint, and no code repository or trained checkpoints are linked.