Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors
summary
The gist
Image reconstruction for positron emission tomography (PET) is an ill-posed Poisson inverse problem that often suffers from severe noise amplification and artifacts In this work we introduce an…
In short
This work addresses noise amplification in PET image reconstruction by introducing an unsupervised optimization framework using a flow-matching generative model as a learned prior. The method learns the continuous manifold of clinically plausible PET images from high-quality data, ensuring reconstructed images adhere to this physical constraint while maintaining consistency with the measured physical data.
Key concepts
- Poisson Inverse Problem
- PET reconstruction is modeled as an ill-posed Poisson inverse problem because photon counting leads to noise. The goal is to estimate the true tracer distribution from noisy sinogram data, which requires finding a solution that satisfies the noisy measurements while remaining physically realistic.
- Flow-Matching Generative Model
- This model defines image generation as a deterministic transport process—an ODE—that maps a simple Gaussian latent variable into the complex distribution of actual PET images. This learned manifold acts as a strong prior, guiding the reconstruction toward clinically plausible structures.
- Augmented Lagrangian (ADMM)
- The solution uses an Augmented Lagrangian approach combined with the Alternating Direction Method of Multipliers (ADMM). This optimization scheme solves two subproblems iteratively: one to find the best image estimate based on physical data, and another to regularize the latent variable against the learned image manifold.
Terminology used across episodes
This episode discusses
The paper
Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors · Read on arXiv
Zhejiang University
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.
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors".
Dev: Here's the gist of the paper. Image reconstruction for positron emission tomography (PET) is an ill-posed Poisson inverse problem that often suffers from severe noise amplification and artifacts In this work we introduce an unsupervised optimization-based reconstruction framework that employs a flow-matching generative model as a learned manifold prior The proposed method learns the continuous manifold of clinically plausible PET images from high-quality reconstructions to ensure the reconstructed image adheres to this prior while enforcing physical data consistency Problem Formulation The PET aims to estimate the spatial distribution of radiotracer uptake x from sinogram data y Due to the photoncounting nature of PET detectors the measurement process can be modeled by y ∼ Poisson(y¯) The log-likelihood can be written as log p(yx) = M ∑ i=1 yi log(y¯i)−y¯i − M ∑ i=1 log(yi!), which corresponds to a maximum-likelihood (ML) reconstruction The model constructs a deterministic transport from a simple base distribution to the target image distribution through a neural ODE After training generating an image amounts to solving the ODE in (3) from t = 0 to t = 1 The flow-matching model defines a deterministic mapping x = Gθ (x0) from a Gaussian latent variable x0 to a PET image Solution Algorithm The proposed formulation in (10) results in a constrained optimization problem The augmented Lagrangian associated with (10) can be written as LA(x,x0,ν) = M ∑ i=1 y¯i −yi log(y¯i) + µ 2 kx0k 2 + hν, x−Gθ(x0)i+ ρ 2 kx−Gθ(x0)k 2, which is the augmented Lagrangian The ADMM algorithm consists of solving two subproblems: 1 Solve x n+1 = argmin x≥0 LA(x n,x n0,ν n) This is a penalized maximum-likelihood reconstruction problem 2 Solve x n+10 = argmin x0 LA(x n+1,x0,ν n) This is a regularized latent projection problem Where the gains come from The proposed method achieves superior noise suppression, structural preservation, and quantitative accuracy compared to conventional reconstruction methods and state-of-the-art deep learning baselines The flow matching model defines image generation through an ODE-based transport from a Gaussian latent space to the PET image distribution which makes the generator differentiable with respect to the latent variable In contrast to diffusion-based posterior sampling methods which typically require long stochastic sampling chains, the proposed method reaches stable performance with a small number of Euler steps This computational efficiency is important for practical PET reconstruction where repeated forward and backward projections are already computationally demanding Experimental Validation The proposed method consistently outperforms conventional reconstruction methods as well as state-of-the-art supervised and unsupervised deep-learning baselines The proposed method achieves the best overall performance with improved structure preservation and superior quantitative metrics relative to both FBSEM and DPS In comparison, our method consistently achieves the lowest NRMSE (e.g., 5.13%, 5.83%, and 5.51% across the three views), the highest PSNR (above 34 dB), and the highest SSIM (above 0.
Paper summary: Rosa: So we’ve been looking at this paper, "Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors," and what they did is use a learned flow matching model to guide image reconstruction in PET scans, which usually gets really noisy <ref:2608.29158#pg1>
Dev: Yeah, it basically says they trained a model on good images to learn the shape of realistic PET data, and then they use that shape as a constraint during the math problem instead of just trusting the raw sensor data <ref:2608.29158#pg1>
Taro: So it’s taking those complex physics problems and adding this learned prior—this idea of what a real image looks like—to keep the solution realistic <ref:2608.29158#pg3>
Rosa: Exactly. They use an augmented Lagrangian method to solve it, which means they are balancing two things: making sure the result matches the raw measurement data, and making sure it follows the smooth path learned by their flow matching model <ref:2608.29158#pg3>
Dev: And that’s where I get interested in it—if you can enforce a continuous structure like that without needing super slow iterative sampling chains, that’s a big deal for real-time systems <ref:2608.29158#pg2>
Taro: Because diffusion models take way too many steps to generate one image, but this flow matching thing gets away with just a few Euler steps, which is much faster for clinical use <ref:2608.29158#pg2>
Rosa: It’s about making the reconstruction computationally efficient while still maintaining that structural accuracy you see in the high quality metrics they report <ref:2608.29158#pg1>
Dev: So when you look at their results, they show very low error numbers, like root mean square error of normality around five point one three percent across different views. That’s what we need for reliable clinical translation <ref:2608.29158#pg2>
Taro: But I wonder about the robustness—what happens when the input data is really messy? Does this learned manifold hold up when the world gets unpredictable? <ref:2608.29158#pg1>
Rosa: That's a huge question, Taro. The authors did show how it performs well against existing supervised and unsupervised deep learning methods, suggesting it’s performing better on structure preservation and quantitative accuracy in their tests <ref:2608.29158#pg1>
Dev: It sounds like they are proving that using generative models as priors can give us better results than just traditional iterative reconstruction when you have these tricky Poisson noise issues <ref:2608.29158#pg1>
Taro: It seems like the real implication here is moving away from purely data-driven methods to methods that incorporate some prior physical understanding, even if that prior is a learned generative model <ref:2608.29158#pg3>
Rosa: It’s about using AI to help enforce reality during the optimization process, which should lead to cleaner, more accurate images for doctors <ref:2608.29158#pg1>
Dev: So we’re looking at this framework as a way to bridge the gap between high-quality training data and practical, fast reconstruction loops for PET imaging <ref:2608.29158#pg2>
Conclusion: Rosa: So we've been talking about this paper on "Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors," and what they actually did is use a learned flow matching model to guide image reconstruction in PET scans, which are usually really noisy problems.
Dev: Yeah, basically they trained a model on good images to learn the shape of realistic PET data and then they use that shape as a constraint during the math problem instead of just trusting the raw sensor data.
Taro: So it’s taking something that's inherently noisy and difficult to solve and using this learned model—this flow matching thing—to guide the solution toward realistic structures. I wonder how long this works reliably outside of a controlled lab setting, Rosa?
Rosa: That’s the question. The team trained this flow matching model on high-quality PET images to learn a deterministic ordinary differential equation transport from a standard Gaussian latent distribution to those empirical PET image distributions, which gives them a differentiable generator of anatomically plausible images.
Dev: That means when you want to generate an image, it boils down to numerically integrating that ordinary differential equation, and because it's deterministic, you get this continuous probability path that transports samples from the simple base distribution to the data distribution.
Taro: So instead of using diffusion models which need long iterative sampling trajectories with a lot of neural network evaluations, this flow matching approach seems to reach stable performance with just a small number of Euler steps. That could save a ton on computational time for routine clinical reconstruction.
Rosa: Exactly. And they incorporate this generator as an explicit manifold constraint into a regularized Poisson likelihood formulation when setting up the problem.
Dev: That leads them to this constrained optimization problem where they minimize something involving the Poisson likelihood term and a penalty term related to how far the reconstructed image is from what their flow matching model predicts based on some latent variable xzero.
Taro: So it’s solving this whole thing using an alternating direction method of multipliers algorithm, which has two main parts: first you solve for the image estimate x by minimizing the augmented Lagrangian term with respect to that part, and then you solve for the latent variable xzero by minimizing that same augmented Lagrangian with respect to xzero.
Rosa: The augmentation itself is this term involving the measurement record, which is the Poisson likelihood of your data, plus a penalty term controlled by some regularization parameter mu on how close your image estimate x is to what their flow matching model produces from a latent variable.
Dev: And that penalty term also involves the squared difference between the image estimate and what their flow matching model outputs given that latent variable, which enforces manifold proximity in the latent space.
Taro: It sounds like they’re using an expectation maximization type surrogate update for the first part, which handles data consistency, and then a gradient based latent space projection for the second part to enforce that proximity to the learned manifold.
Rosa: The paper shows this method consistently achieves superior noise suppression, better structural preservation, and more quantitative accuracy when compared against conventional reconstruction methods and state of the art unsupervised deep learning baselines.
Dev: They achieved some solid numbers too, consistently achieving the lowest root mean square error of normality, or NRMSE, like five point one three percent, five point eight three percent, and five point five one percent across three different views they tested.
Taro: And they also got a very high peak signal-to-noise ratio of above thirty-four decibels and a very high structural similarity index of above zero point five one. That shows how well the image structure is kept while cleaning up the noise.
Rosa: It’s interesting because it seems to work really well across different metrics, meaning it’s not just good at one thing; it preserves structure while also being accurate quantitatively.
Dev: The core idea is that by using this flow matching model to define the image generation process through an ordinary differential equation transport from a Gaussian latent space, they make the generator differentiable with respect to that latent variable.
Taro: That differentiability is key because it lets them use gradient-based optimization in these imaging inverse problems, which is what makes the whole alternating direction method of multipliers algorithm feasible.
Rosa: It moves away from methods like diffusion-based posterior sampling that require long stochastic sampling chains, which are computationally expensive and demand a lot of neural network evaluations.
Dev: So they get stable performance with only a small number of Euler steps for image generation, which is really important when you're thinking about practical PET reconstruction where forward and backward projections are already demanding.
Taro: The authors Hengjia Ran, Jie Luo, Yutao Zhu, Rui Hu, Huafeng Liu and Bo Zhao are the ones showing this out. They’re proving that this approach outperforms other methods in terms of noise suppression and structure preservation when compared against FBSEM and DPS.
Rosa: It seems like the future work would be exploring how this framework handles failure modes when the world misbehaves in a more complex way than just simple noise amplification, testing its robustness beyond the clean data they used to train it.
Dev: That makes sense. We'll see if this framework holds up when we push it outside of the pristine lab data where they trained it, and how gracefully it recovers when things get messy in real-world scenarios.
More episodes
- Daily Summary for 2026-11-19
- 2111.07910-Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- 2203.02149-HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- 2205.10102-Degradation-Aware Unfolding Half-Shuffle Transformer for Spectral Compressive Imaging
- 2311.10959-Structure-Aware Sparse-View X-ray 3D Reconstruction
- 2403.04116-Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis
- 2410.02572-Combining Pre- and Post-Demosaicking Noise Removal for RAW Video
- 2604.17453-Learned Nonlocal Feature Matching and Filtering for RAW Image Denoising
- 2603.20387-End-to-End Multi-Task Learning for Adjustable Joint Noise Reduction and Hearing Loss Compensation
- 2507.20825-Chirp-Permuted AFDM: A Versatile and Multifunctional Waveform for 6G