PET-DPC: Diffusion with Posterior Physical Correction

• The paper demonstrates that PET-DPC achieves highly accurate PET reconstructions by integrating posterior physical corrections, improving metrics such as PSNR from 36.86 to 47.70 in brain studies.
• PET-DPC uses a Geometric TOF Probabilistic Image (GTP-image) as a physics-informed conditioning prior, effectively bridging the gap between raw detector data and refined image outputs.
• Empirical evaluations show that PET-DPC reduces reconstruction time by up to 85% for whole-body scans while maintaining data consistency and outperforming both end-to-end deep learning models and traditional iterative solvers.

PET-DPC denotes a conditional diffusion model with posterior physical correction for PET image reconstruction that was proposed to address the persistent tension between fast deep learning-based reconstruction and the quantitative reliability expected of clinical PET. In this framework, an innovative normalization procedure produces a Geometric TOF Probabilistic Image (GTP-image) as the conditioning prior, while physical information from the measured PET data is injected during diffusion sampling to perform posterior scatter, attenuation, and random corrections. The method was trained and validated on 300 brain and 50 whole-body PET datasets, a physical phantom, and 20 simulated brain datasets, and was reported to produce reconstructions closely aligned with fully corrected OSEM images, to generalize to out-of-distribution data, and to reduce reconstruction time relative to iterative methods (Hou et al., 20 Aug 2025).

1. Conceptual position and problem setting

PET-DPC was introduced in the context of a broader concern: although deep learning-based PET reconstruction can be fast, its clinical adoption has been limited by quantitative accuracy, artifacts, limited interpretability, data-driven dependence, and overfitting risks (Hou et al., 20 Aug 2025). The method therefore targets not only image fidelity but also explicit data consistency with the underlying PET measurement process.

The paper frames PET-DPC as an alternative to two unsatisfactory extremes. On one side are end-to-end deep learning models, which are computationally attractive but can lose quantitative fidelity. On the other side are traditional iterative methods, which remain quantitatively trusted but are substantially slower, especially for whole-body studies. PET-DPC seeks to retain the generative flexibility of a conditional DDPM while constraining sampling by the PET likelihood so that reconstruction remains physically plausible (Hou et al., 20 Aug 2025).

A frequent misconception is to treat PET-DPC as merely another image-domain denoiser. The method is more specific than that characterization: its defining feature is not diffusion alone, but diffusion with posterior physical correction. The ablation study reported in the paper makes this distinction operationally important, because the performance gains are attributed to the correction mechanism rather than to the unconditional generative model by itself (Hou et al., 20 Aug 2025).

2. Conditional diffusion architecture and reconstruction workflow

PET-DPC is founded on the conditional Denoising Diffusion Probabilistic Model (DDPM). During training, Gaussian noise is gradually added to a clean PET image $x_0$, and the network learns to estimate the injected noise at each diffusion step while conditioning on auxiliary information. The training objective is the mean squared error between predicted and true noise:

$$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$

where $x_c$ denotes the conditioning input (Hou et al., 20 Aug 2025).

The conditioning signal is the GTP-image, which is intended to be a coarse but physics-informed prior. The workflow summarized in the paper is:

1. Input: GTP-image and normalized list-mode data.
2. Conditioned Denoising: the diffusion model denoises using the GTP-image as condition.
3. Posterior Physical Correction: at each denoising step, a physics-based correction using measured PET data is applied during sampling (Hou et al., 20 Aug 2025).

This organization is technically significant because it separates two roles that are often conflated in reconstruction models. The conditional diffusion network provides a powerful image prior, while the posterior correction stage enforces agreement with the measured detector data. A plausible implication is that PET-DPC is designed to reduce the susceptibility of generative models to visually plausible but quantitatively incorrect reconstructions.

The paper compared PET-DPC with FastPET and a Conditional-DDPM without posterior correction, as well as with the traditional iterative method QuanTOF. The reported outcome was that PET-DPC reconstructed images more closely aligned with fully corrected OSEM images and outperformed end-to-end deep learning models in quantitative metrics, while in some cases surpassing traditional iterative methods (Hou et al., 20 Aug 2025).

3. Posterior physical correction and PET likelihood constraint

The central innovation of PET-DPC is the insertion of posterior physical correction into the reverse diffusion process. The paper states that, unlike pure data-driven methods, PET-DPC enforces PET physical constraints during sampling by using the measured data to steer the reconstruction toward physically plausible solutions (Hou et al., 20 Aug 2025).

The PET data model is written in terms of a Poisson mean for each event or detector bin:

$\lambda_i(x) = \sum_j P_{ij} x_j + T_i + S_i$

where $P_{ij}$ is the system matrix, $x_j$ is voxel activity, and $T_i$ and $S_i$ denote mean random and scatter events, respectively (Hou et al., 20 Aug 2025).

With measurement data constraint, the sampling SDE includes a likelihood-gradient term:

$$dx = -\beta(t)x_t\,dt + \sqrt{\beta(t)}\,dw - \beta(t)\nabla_{x_t}\log p(y|x_t)\,dt$$

and the Poisson log-likelihood is

$$\log p(y|x) = \sum_i y_i \log \lambda_i(x) - \lambda_i(x) - \log y_i!$$

(Hou et al., 20 Aug 2025).

In this formulation, the gradient $$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$0 functions as the posterior correction. The paper describes this term as directly enforcing data consistency, especially for scatter and attenuation corrections, and the abstract additionally identifies random corrections as part of the posterior physical correction mechanism (Hou et al., 20 Aug 2025). Sampling is implemented with a predictor-corrector (PC) algorithm that alternates between standard diffusion updates and gradient-based correction steps.

The ablation study makes the role of this mechanism explicit. PET-DPC with increasing numbers of correction steps from 1 to 5 showed monotonic improvement in quantitative accuracy and visual fidelity. The paper reports the following values for the clinical brain study:

Correction steps	Metrics
1-step	PSNR 36.86, SSIM 0.966, NRMSE 0.234
5-step	PSNR 47.70, SSIM 0.994, NRMSE 0.066

These results were interpreted in the paper as evidence that posterior correction is crucial for quantitative reliability and high-fidelity clinical PET reconstruction (Hou et al., 20 Aug 2025).

4. Geometric TOF Probabilistic Image (GTP-image)

The GTP-image is the conditioning prior used by PET-DPC. It is described as a physics-inspired, rapid back-projection from PET list-mode data that incorporates time-of-flight (TOF) information and detector geometry (Hou et al., 20 Aug 2025). Its purpose is to provide an auxiliary image closer to the true activity distribution than pure noise.

For each list-mode line of response, the method estimates the annihilation position using TOF information, applies a Gaussian TOF weighting and a geometric weighting, scales by detector efficiency, and then normalizes by the sensitivity map. The paper gives the following expressions:

$$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$1

$$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$2

$$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$3

(Hou et al., 20 Aug 2025).

The GTP-image serves as the conditional input throughout training and sampling. According to the paper, this conditioning prior improves convergence and fidelity because it anchors generation to the actual measurement geometry and TOF structure rather than requiring the diffusion model to infer the entire reconstruction trajectory from noise alone (Hou et al., 20 Aug 2025).

A second misconception is that the GTP-image by itself constitutes the essential contribution. The paper argues otherwise: the GTP-image is important as a conditioning prior, but the ablation results identify posterior correction as the critical component for implementing scatter and attenuation corrections and for enhancing reconstruction accuracy. This suggests that PET-DPC should be understood as a coupled system of physics-informed conditioning and likelihood-based correction, not as a single-stage prior image estimator (Hou et al., 20 Aug 2025).

5. Empirical evaluation, quantitative behavior, and runtime

The reported evaluation covers clinical brain studies, clinical whole-body studies, a physical phantom, and simulated brain data (Hou et al., 20 Aug 2025). Across these settings, PET-DPC was compared with FastPET, Conditional-DDPM without posterior correction, and QuanTOF.

For the clinical brain study, PET-DPC achieved the highest reported quantitative metrics:

• PSNR: 47.70
• SSIM: 0.994
• NRMSE: 0.066

The paper also states that PET-DPC showed the strongest tumor-to-background CNR fidelity versus QuanTOF and preserved tumor edges and fine brain structures (Hou et al., 20 Aug 2025).

For the clinical whole-body study, PET-DPC again outperformed FastPET and Conditional-DDPM, with:

• PSNR: 49.90
• SSIM: 0.991
• NRMSE: 0.162

The qualitative account emphasizes better preservation of small lesions and fine organ uptake areas, whereas FastPET was said to smooth small structures and Conditional-DDPM remained inferior in quantitative accuracy (Hou et al., 20 Aug 2025).

For the physical phantom, PET-DPC reconstructed the background with high uniformity and recovered the expected tumor-to-background ratio of 4:1, matching or exceeding QuanTOF, while FastPET and Conditional-DDPM exhibited non-uniform backgrounds and intensity underestimation (Hou et al., 20 Aug 2025). For the simulation study, PET-DPC recovered a gray:white matter ratio of 1.868, close to QuanTOF at 1.898, although both underrepresented the phantom’s actual ratio of 3.846:1, which the paper identifies as a known quantitative challenge in PET (Hou et al., 20 Aug 2025).

The runtime findings are reported in two forms in the paper. The abstract states that, compared to iterative methods, PET-DPC reduced reconstruction time by 50% for brain scans and 85% for whole-body scans (Hou et al., 20 Aug 2025). The detailed evaluation gives explicit examples of QuanTOF 82 s vs PET-DPC 47 s for brain and QuanTOF 17,081 s vs PET-DPC 2,707 s for whole-body, described as 43% saved and 84% saved, respectively (Hou et al., 20 Aug 2025). The paper also notes that FastPET remained the fastest method, but with notable loss in quantitative accuracy.

6. Relation to adjacent PET reconstruction and synthesis methods

PET-DPC belongs to a rapidly expanding set of PET methods that combine deep generative modeling with explicit domain structure, but its operating point is distinct. PMDM-PET addresses PET image reconstruction from sinograms by first producing a minimum-MSE posterior mean estimate with DeepPET and then using a conditional diffusion model to transport that estimate toward the ground-truth PET distribution, thereby targeting the perception-distortion tradeoff (Sun et al., 11 Mar 2025). PET-DPC, by contrast, emphasizes posterior physical correction during sampling rather than an explicit perception-distortion decomposition.

Relative to STPDnet, PET-DPC is also methodologically distinct. STPDnet is a spatial-temporal convolutional primal dual network for dynamic PET image reconstruction, embedding the forward and adjoint projection operators inside an unrolled iterative learning process and using $$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$4 convolutions to encode spatial and temporal correlations (Hu et al., 2023). PET-DPC does not adopt a primal-dual unrolling scheme; its physical consistency is enforced through the likelihood-gradient correction term within diffusion sampling (Hou et al., 20 Aug 2025).

Other recent PET diffusion models operate on different tasks or data domains. PJDM performs projection-domain PET tracer conversion, transforming $$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$5F-FDG PET sinograms into $$L_{\text{DDPM}} = \mathbb{E}_{x_0, \epsilon, t} \left[ \left\| \epsilon_\theta(x_t, t, x_c) - \epsilon \right\|^2 \right]$$6F-DOPA PET sinograms through a two-stage prior-guided joint diffusion model (Chen et al., 20 Jun 2025). st-DTPM addresses delayed scan PET image prediction through a spatial-temporal guided diffusion transformer probabilistic model with spatial concatenation of early and noisy delayed scans and temporal embedding of diffusion step and delay time (Hong et al., 2024). These methods illustrate that diffusion has been applied to reconstruction, tracer conversion, and delayed scan prediction, but PET-DPC is specifically a physics-constrained reconstruction method with posterior correction tied to PET measurement likelihood (Hou et al., 20 Aug 2025).

Within the broader PET deep learning landscape, PET-DPC also differs from approaches that prioritize image enhancement without anatomical inputs. For example, the U-Net framework of "Deep Learning-Based Partial Volume Correction in Standard and Low-Dose PET-CT Imaging" maps standard- or low-dose PET directly to full-dose plus partial-volume-corrected PET images without requiring anatomical data at inference time (Azimi et al., 2022). That work addresses joint denoising and PVC, whereas PET-DPC addresses reconstruction itself and explicitly integrates scatter, attenuation, and random correction within the generative sampling process (Hou et al., 20 Aug 2025).

The practical implication is that PET-DPC occupies a specific niche: it is neither a purely end-to-end reconstructor nor a conventional iterative solver, but a reconstruction framework in which a conditional diffusion prior is constrained by PET physics at inference time. The paper’s reported strengths—quantitative agreement with fully corrected OSEM, OOD generalization, preservation of background uniformity, and accurate tumor-to-background ratios—are the properties by which it should be evaluated (Hou et al., 20 Aug 2025). Its stated limitation is that the physics-based correction step remains the current computational bottleneck, particularly for high-event-count, large-volume scans, and fewer correction steps trade speed for reduced reliability (Hou et al., 20 Aug 2025).