DDPM-MR-PETCon: PET/MR Denoising

• The paper demonstrates a novel integration of denoising diffusion models with MR anatomical priors and PET data-consistency constraints to enhance low-dose PET image quality.
• It employs a U-Net backbone with multi-modality fusion and transformer modules, achieving superior PSNR, SSIM, and low NMSE compared to traditional methods.
• The approach delivers robust performance in out-of-distribution clinical scenarios while effectively balancing anatomical fidelity and PET-specific structural preservation.

DDPM-MR-PETCon refers to a class of algorithms that combine denoising diffusion probabilistic models (DDPMs) with magnetic resonance (MR) anatomical priors and explicit PET data-consistency constraints for PET image denoising and restoration. These methods are specifically designed to address the challenges inherent in low-dose PET imaging—namely, high noise and low contrast—by leveraging MR structural guidance alongside robust probabilistic diffusion frameworks. The DDPM-MR-PETCon approach has been refined to accommodate multi-modality fusion, out-of-distribution (OOD) clinical data, and optimal trade-offs between anatomical fidelity and PET-unique structural preservation, situating it at the forefront of PET/MR image restoration research (Gong et al., 2022, Zhang et al., 12 Feb 2026).

1. Probabilistic Diffusion Models in PET Restoration

DDPMs are a class of iterative generative models that learn to recover clean images from noise by modeling the inverse of a known noising process. Given an image $x_0$, the forward (noising) process generates a sequence $x_1, ..., x_T$ via

$$q(x_t|x_{t-1}) = \mathcal N\left(x_t; \sqrt{1 - \beta_t}x_{t-1}, \beta_t I\right), \quad \beta_t \in (0,1),$$

and the closed-form marginal

$$q(x_t|x_0) = \mathcal N\left(x_t; \sqrt{\bar\alpha_t} x_0, (1 - \bar\alpha_t)I\right), \quad \bar\alpha_t = \prod_{s=1}^t (1-\beta_s).$$

The reverse process is parameterized by a neural network to predict the noise component at each step, and samples are drawn by iterative denoising (Gong et al., 2022, Zhang et al., 12 Feb 2026). The models are extensible to multi-channel, multi-modal inputs and are well-suited to exploiting high-resolution MR anatomy for guiding PET image enhancement.

2. MR Guidance and Multi-Modality Fusion

DDPM-MR-PETCon incorporates MR anatomical priors as conditional inputs to the denoising network. During the reverse diffusion, the MR image is typically concatenated with the noisy PET image, either at the input layer or as features at down- and up-sampling stages of the U-Net architecture. This design enables the model to prioritize anatomical consistency and resolve ambiguous structures in low-SNR PET images (Gong et al., 2022, Zhang et al., 12 Feb 2026).

Recent enhancements feature sophisticated fusion modules ("f_{mfuse}") that learn to aggregate global and local modality-specific features derived from both MR and low-dose PET through multi-stage transformer and invertible coupling blocks, ensuring spatially and texturally coherent representations. The fused feature map is dynamically injected into the diffusion U-Net at every denoising step, resulting in improved accuracy and stability of restoration, particularly in OOD domains (Zhang et al., 12 Feb 2026).

3. Data-Consistency and Bias Control: The PETCon Mechanism

A defining feature of DDPM-MR-PETCon is the explicit data-consistency constraint. Rather than relying exclusively on learned priors (which can introduce anatomical bias if the PET and MR anatomies diverge), PETCon enforces, at each denoising step, a correction that proportionally "pulls" the intermediate estimate toward the observed noisy PET measurement:

$$x_{t-1} = \frac{1}{\sqrt{\alpha_t}} \left[x_t - \frac{\beta_t}{\sqrt{1 - \bar\alpha_t}} \epsilon_\theta(x_t, t; x_{\mathrm{prior}})\right] - \frac{\sigma_t^2}{\sigma_d^2}(x_{\mathrm{noisy}} - x_t) + \sigma_t z$$

where $\epsilon_\theta$ is the denoising network, $x_{\mathrm{prior}}$ is the MR image, and $\sigma_d^2$ models PET noise variance (Gong et al., 2022). This term is critical for reducing quantification bias: it preserves true PET contrast while leveraging anatomical detail from MR, and dominates in regions where MR and PET may structurally disagree.

4. Training Paradigms and Architectural Variants

Training Strategies

• Noise-Adaptive Training: Training is performed with the DDPM conditioned only on the MR prior, not on the PET data directly. All PET consistency is imposed during inference, making the model adaptable to any noise level without retraining (Gong et al., 2022).
• Supervise-Assisted Multi-Stage Training: Models such as MFdiff (editor's term: "multi-fusion diffusion") first learn on simulated data, then fine-tune on limited OOD clinical sets to encode both generalized and dataset-specific priors (Zhang et al., 12 Feb 2026).

Network Architectures

• U-Net Backbone: Most methods use 2D or 3D U-Nets with residual, attention, and transformer blocks to model the conditional diffusion process (Gong et al., 2022, Zhang et al., 12 Feb 2026, Yoon et al., 2024).
• Patch-wise 3D Training: To manage GPU constraints, 3D volumetric patch-wise training is adopted in approaches like CSRD, maintaining spatial coherence while reducing resource demands (Yoon et al., 2024).

5. Quantitative and Qualitative Evaluation

Rigorous validation has demonstrated that DDPM-MR-PETCon and its derivatives outperform classical denoising (e.g., non-local means, U-Net) and other state-of-the-art deep learning models:

Method	PSNR (dB)	SSIM	NMSE
LPET (input)	21.13 ± 2.71	0.839 ± .041	0.176±.055
U-Net (PET+MR)	26.4	0.81	—
CSRD	37.28 ± 1.94	0.990 ± .006	0.009±.005
DDPM-MR-PETCon	37.93 ± 1.85	0.991 ± .008	0.008±.004

• DDPM-MR-PETCon achieves the lowest error metrics across simulated and in-vivo datasets, with particular superiority in local regions relevant to lesion detectability and morphometric assessment.
• The data-consistency constraint is essential for quantification accuracy; MR-only models exhibit large bias in regions where PET/MR anatomy diverges (Gong et al., 2022, Zhang et al., 12 Feb 2026).
• Uncertainty quantification is provided natively by the stochastic diffusion trajectory, an advantageous property for clinical applications requiring confidence estimation.

6. Applications, Limitations, and Future Directions

Applications

• Low-dose PET denoising and restoration: Robust enhancement of PET images acquired at drastically reduced dose or scan times.
• Out-of-Distribution Adaptation: Two-stage and fusion-based methods (e.g., MFdiff/DDPM-MR-PETCon) generalize effectively to OOD clinical scenarios with varying dose, timing, and injection protocols (Zhang et al., 12 Feb 2026).
• MR-less PET restoration: Synthetic MR images generated via PET-conditioned DPMs can guide anatomical PET restoration when no real MR is available, expanding utility (Gan et al., 2024).

Limitations

• Computational burden remains substantial (up to 50 min per 3D volume for 1000-step samplers) (Gong et al., 2022).
• Most implementations are restricted to 2D or small 3D patches due to GPU memory; full-volume end-to-end 3D diffusion remains a challenge (Yoon et al., 2024).
• Per-target registration and diffusion model fine-tuning may add operational complexity (Webber et al., 4 Jun 2025).

Future Research

• Development of more memory-efficient architectures and solvers (e.g., denoising diffusion ODEs, subspace samplers).
• General-purpose MR-PET registration and synthetic augmentation pipelines.
• Integration with segmentation priors and expansion to other tracers, body regions, and scanner geometries (Webber et al., 4 Jun 2025, Yoon et al., 2024).

7. Context Within PET/MR Image Processing

DDPM-MR-PETCon and related methods embody the latest advancements in probabilistic, multi-modal image restoration for PET/MR. They supersede conventional and early deep learning approaches by unifying explicit data-consistency, sophisticated multi-modality fusion, and noise-adaptive generative modeling. Other influential directions include patch-wise 3D residual diffusion (CSRD) (Yoon et al., 2024), pseudo-anatomy synthesis for anatomical priors (Gan et al., 2024), and direct manipulation of denoising trade-offs between MR anatomical fidelity and PET-unique features.

The methodological diversity within the DDPM-MR-PETCon framework enables broad adaptability, high restoration fidelity, and quantification reliability across a range of challenging PET imaging scenarios (Gong et al., 2022, Zhang et al., 12 Feb 2026, Webber et al., 4 Jun 2025, Yoon et al., 2024).