DECADE: A Temporally-Consistent Unsupervised Diffusion Model for Enhanced Rb-82 Dynamic Cardiac PET Image Denoising
Authors: Yinchi Zhou, Liang Guo, Huidong Xie, Yuexi Du, Ashley Wang, Menghua Xia, Tian Yu, Ramesh Fazzone-Chettiar, +9 more
Organizations: Department of Biomedical Engineering, Yale University, New Haven, CT, USA · Department of Radiology and Biomedical Imaging, Yale School of Medicine, New Haven, CT, USA · Department of Medicine (Cardiology), Yale School of Medicine, New Haven, CT, USA · Siemens Medical Solutions USA, Inc., Knoxville, TN, USA · Department of Nuclear Medicine, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland
Rb-82 dynamic cardiac PET imaging is widely used for the clinical diagnosis of coronary artery disease (CAD), but its short half-life results in high noise levels that degrade dynamic frame quality and parametric imaging. The lack of paired clean-noisy training data, rapid tracer kinetics, and frame-dependent noise variations further limit the effectiveness of existing deep learning denoising methods. We propose DECADE (A Temporally-Consistent Unsupervised Diffusion model for Enhanced Rb-82 CArdiac PET DEnoising), an unsupervised diffusion framework that generalizes across early- to late-phase dynamic frames. DECADE incorporates temporal consistency during both training and iterative sampling, using noisy frames as guidance to preserve quantitative accuracy. The method was trained and evaluated on datasets acquired from Siemens Vision 450 and Siemens Biograph Vision Quadra scanners. On the Vision 450 dataset, DECADE consistently produced high-quality dynamic and parametric images with reduced noise while preserving myocardial blood flow (MBF) and myocardial flow reserve (MFR). On the Quadra dataset, using 15%-count images as input and full-count images as reference, DECADE outperformed UNet-based and other diffusion models in image quality and K1/MBF quantification. The proposed framework enables effective unsupervised denoising of Rb-82 dynamic cardiac PET without paired training data, supporting clearer visualization while maintaining quantitative integrity.
Figures & tables
Fig. 1: DECADE consists of a diffusion model training stage and a sampling stage. During Training Stage A, the model is pre-trained on 2-7 minute mean static PET images. Three Gaussian-smoothed consecutive noisy frames are then used as conditional input into ControlNet to refine the pre-trained model. In the sampling stage, the unconditional diffusion model is applied for time steps from 1000 to 950, followed by the use of ControlNet from time steps 950 to 0. Noisy frames are incorporated at each time step to guide the denoising process.
Fig. 2: Example dynamic frames and K1 images of a rest scan from the Vision 450 dataset. The radiology report indicates the patient has no myocardial defect. DECADE achieved superior denoising performance across all dynamic frames. The denoised frames and K1 map produced by DECADE exhibit the lowest noise levels and the most similar tracer distribution to the original dynamic frames.
Fig. 3: An example subject from the Vision 450 dataset with a confirmed myocardial defect. The clinical report indicates a partially reversible perfusion defect in the basal to apical inferior, basal inferolateral, and basal inferoseptal walls. The catheteriztion study confirms the appearance of dominant LCx with severe calcified stenosis. Dynamic frames, K1 images, k2 images, and polar maps before and after denoising are shown. K1 images are displayed in axial view, horizontal long-axis view (HLA), vertical long-axis view (VLA), and short-axis view (SA). DECADE achieves superior denoising performance across all dynamic frames over time, producing K1 maps with reduced noise. Notably, the stress polar map’s noise artifact near the apex region disappears after denoising, and the boundary of the myocardium becomes more apparent. Importantly, the myocardial defect, pointed out by a blue arrow in the VLA and SA views of parametric images and polar map, remains visible and even easier to detect after denoising due to improved parametric image quality. The defect in the k2 image is also indicated by higher uptake in the defect region of the VLA and SA view.
Fig. 4: A normal subject from the Vision 450 dataset. The radiology report indicates this patient has no myocardial defects. DECADE also produces visually appealing denoised results for this patient on dynamic frames. The denoised images exhibit a more uniform distribution and significantly lower noise levels. The contour of the myocardium is better depicted in all views of the K1 and k2 images.
Fig. 5: Comparison results on an example rest scan from the Quadra Bern dataset. 15% dynamic frames are used as noisy inputs, and 100% scans are used as ground truth to compare with denoised results. DECADE successfully generalizes to the downsampled Quadra Bern dataset without fine-tuning or re-training. The denoised images from DECADE preserve structural details of anatomical regions and ensure quantitative accuracy. In contrast, underestimation occurs with N2V, and overestimation occurs with the self-supervised UNet and DDIM for dynamic frames.
SSIM ↑
PSNR ↑
NMSE ↓
Noisy Frames
0.59 ± 0.12
26.72 ± 4.03
0.47 ± 0.16
N2V
0.66 ± 0.11
27.82 ± 3.09
0.40 ± 0.10
Self-supervised UNet
0.68 ± 0.07
26.84 ± 2.23
0.40 ± 0.10
DDIM
0.58 ± 0.07
24.02 ± 2.94
0.61 ± 0.11
DECADE
0.73 ± 0.10
30.64 ± 3.40
0.29 ± 0.08
TABLE II: Comparison across different denoising methods applied to 15% count-level images, with 100% count images serving as the reference, from 5 subjects scanned using Siemens Biograph Vision Quadra scanners (250 images in total)
Fig. 6: The ablation study on the effect of image guidance and temporal consistency using both the Vision 450 dataset and the Quadra dataset. The early frame of a rest scan is shown. Adding image guidance using DPS ensures quantitative accuracy similar to the original dynamic frames, as indicated by the ”DPS Only” column. Refining the pre-trained diffusion model using noisy frames through ControlNet further enhances qualitative accuracy and helps preserve structural details in anatomical regions with weak signals. The ”DPS w/Single Frame Control” uses the current Gaussian-smoothed frame as control, while DECADE uses three consecutive frames to add temporal consistency. Using multiple frames as conditional input further increases the model’s performance in more challenging regions with relatively weak signals, such as the cavity between the left ventricle and right ventricle, or the edges of anatomical regions, resulting in the best denoised images. Colored arrows indicate artifacts that appear where temporal consistency and image guidance are not added. The white arrow points to the connecting artifact in the septal wall, the cyan arrow points to the edge of the right atrium wall, and the purple arrow points to the boundary between the vena cava and the right atrium.
SSIM ↑
PSNR ↑
NMSE ↓
15% Dynamic Frames
0.587
26.724
0.469
DPS Only
0.603
28.093
0.388
DPS w/ single frame control
0.717
30.063
0.315
DECADE
0.733
30.641
0.291
TABLE IV: Ablation studies on 5 subjects scanned using Siemens Biograph Vision Quadra scanners (250 images in total), 15% frames are used as input images and full-count images are used as the reference.
Real-time cardiac cine MRI enables visualization of the beating heart during free breathing, but severe undersampling and motion make reconstruction highly challenging. A central challenge for reconstruction is incorporating powerful priors of cardiac anatomy while remaining computationally efficient. We propose Piecewise Dynamic Diffusion Regularization (PDDR), a reconstruction method that integrates a spatiotemporal diffusion model as a generative prior within a variational reconstruction framework for cine MRI. The model employs dedicated spatial layers to encode anatomical structure and temporal layers to capture cardiac motion learned from gated cine data. PDDR leverages the dynamic prior in a piecewise manner, enabling the efficient use of spatiotemporal diffusion models for processing of long real-time sequences. Experiments on retrospectively accelerated and prospective real-time cine MRI demonstrate that PDDR outperforms classical, unsupervised, and diffusion-based methods, delivering high-quality reconstructions with substantially reduced computation time compared to state-of-the-art baselines. These results highlight PDDR as a practical and scalable solution for free-breathing, real-time cardiac MRI. Code is available at https://github.com/MLI-lab/pddr.
Florian Fürnrohr, Reinhard Heckel
Technical University of Munich · Munich Center for Machine Learning (MCML)
Magnetic resonance imaging involves an inherent trade-off among spatial resolution, acquisition time, and noise. This trade-off contributes to long scan times and high cost. Deep learning has improved image denoising, but cardiac MRI remains difficult because high-resolution, rapid acquisitions generally lack corresponding low-noise ground truth. Self-supervised denoising offers a potential solution by learning from noisy image pairs or even single noisy acquisitions. However, we show that Noise2Void-style blind-spot denoising, which uses a mean squared error (MSE) loss and assumes zero-mean, independent and identically distributed (i.i.d.) noise, is poorly suited to MR magnitude images. When applied to short-axis T2-weighted and T1ρ-weighted cardiac MRI with synthetic Rician noise, it produces biased denoised images and biased parametric maps of T2 and T1ρ. To address this limitation, we formulate self-supervised denoising as maximum likelihood estimation under a known Rician noise model. This yields unbiased denoisers that are competitive with supervised baselines.
Diffusion-based unsupervised image registration has been explored for cardiac cine MR, but expensive multi-step inference limits practical use. We propose FlowReg, a flow-matching framework in displacement field space that achieves strong registration in as few as two steps and supports further refinement with more steps. FlowReg uses warmup-reflow training: a single-step network first acts as a teacher, then a student learns to refine from arbitrary intermediate states, removing the need for a pre-trained model as in existing methods. An Initial Guess strategy feeds back the model prediction as the next starting point, improving refinement from step two onward. On ACDC and MM2 across six tasks (including cross-dataset generalization), FlowReg outperforms the state of the art on five tasks (+0.6% mean Dice score on average), with the largest gain in the left ventricle (+1.09%), and reduces LVEF estimation error on all six tasks (-2.58 percentage points), using only 0.7% extra parameters and no segmentation labels. Code is available at https://github.com/mathpluscode/FlowReg.
Yunguan Fu, Wenjia Bai, Wen Yan +3
University College London · InstaDeep · Imperial College London