Plug-and-play with Denoising Diffusion Probabilistic Model (DDPM) Denoiser for Highly Accelerated Cardiac Perfusion Image Reconstruction

Background: Magnetic resonance imaging (MRI) provides high-quality images but suffers from long acquisition times. To accelerate imaging, undersampled data must be reconstructed through inverse methods such as parallel imaging [1], compressed sensing [2], and learning-based approaches [3,4]. Plug-and-play (PnP) methods address this by alternating between data consistency and denoising, enabling the use of diverse priors. Recently, Denoising Diffusion Probabilistic Models (DDPMs) [5] have outperformed generative adversarial networks [6] by leveraging denoising tasks for training. Applied to MRI reconstruction via Bayes’ rule, DDPMs achieve performance rivaling or surpassing existing learning-based methods. First-pass cardiac perfusion imaging remains difficult due to low SNR and motion artifacts, with clinical protocols typically limited to R=2–3 acceleration. At R≥4, conventional methods often yield non-diagnostic images. In this work, we integrate a pretrained DDPM prior into a PnP framework to reconstruct highly accelerated, prospectively downsampled perfusion datasets at R=4, 6, and 8, including multiband acquisitions with net acceleration R=8.

Methods: Data: We trained on 300 clinically cardiac perfusion datasets (R=2) from 94 patients scanned on a 3T Siemens Skyra. Each patient contained 3–5 slices per R-R interval at 2×2.4 mm² resolution. Per-frame HICU [7] reconstructions served as ground truth (GT). Testing used: 1) single-band datasets at R=4, 6, and 8 with variable-density sampling, and 2) multi-band SMILE [8] datasets at R=8 with 1.5×1.5 mm² resolution.

DDPM training: A U-Net model was trained with AdamW for T=1000 diffusion steps. Training patches were cropped to 96×96, with real and imaginary parts as two channels.

Fast inference: Reconstructions were initialized with time-averaged data, reducing reverse steps from T=1000 to T=50 for faster inference.

Algorithm: PnP alternates between denoising and data consistency. We use a primal-dual splitting scheme, with σ controlling denoiser strength and λ weighting data consistency. Unlike the standard DDPM generative sampling procedure, here we use DDPM as deterministically denoiser, no additional stochastic noise is injected at each reverse step. (See Figure 1)

Results: Figure 2 presents representative PnP images with DDPM reconstructions from one clinical perfusion dataset at R=4, 6, and 8, compared to PnP with DnCNN based denoiser [9]. Our PnP-DDPM method trained only on 2D images provides high-quality reconstructions comparing to 2D+t DnCnn model.

Figure 3 demonstrates the PnP-DDPM has good reconstruction quality of multi-band SMILE perfusion at R=8.

Conclusion: We show that PnP methods with a DDPM prior can effectively reconstruct highly accelerated cardiac perfusion images with good image quality. While our current approach achieves promising results in per-frame reconstruction, future work could explore extending the model to 2D+t training for improved temporal consistency.

Figure 1 Algorithm of PnP with conventional denoiser and PnP with DDPM denoiser
Figure 2 Inference on the data acquired from three patients at R4, R6, and R8
Figure 3 Multiband SMILE perfusion data