IRSDE-Despeckle: A Physics-Grounded Diffusion Model for Generalizable Ultrasound Despeckling

Shuoqi Chen; Yujia Wu; Geoffrey P. Luke

cs.CVFeb 26, 2026

IRSDE-Despeckle: A Physics-Grounded Diffusion Model for Generalizable Ultrasound Despeckling

Authors: Shuoqi Chen, Yujia Wu, Geoffrey P. Luke

Organizations: Thayer School of Engineering, Dartmouth College, Hanover, NH, USA

Abstract

Ultrasound imaging is widely used for real-time, noninvasive diagnosis, but speckle and related artifacts reduce image quality and can hinder interpretation. We present a diffusion-based ultrasound despeckling method built on the Image Restoration Stochastic Differential Equations framework. To enable supervised training, we curate large paired datasets by simulating ultrasound images from speckle-free magnetic resonance images using the Matlab UltraSound Toolbox. The proposed model reconstructs speckle-suppressed images while preserving anatomically meaningful edges and contrast. On a held-out simulated test set, our approach consistently outperforms classical filters and recent learning-based despeckling baselines. We quantify prediction uncertainty via cross-model variance and show that higher uncertainty correlates with higher reconstruction error, providing a practical indicator of difficult or failure-prone regions. Finally, we evaluate sensitivity to simulation probe settings and observe domain shift, motivating diversified training and adaptation for robust clinical deployment.

Figures & tables

Figure 1: Overview of our Image Restoration SDE (IRSDE) ultrasound despeckling pipeline. The forward image degradation SDE progressively perturbs a HQ image $x(0)$ into a terminal state $x(T)$ , which is approximately a noisy observation of the low-quality image. The reverse image restoration SDE then evolves $x(T)$ back towards $x(0)$ by simulating the corresponding reverse-time dynamics, where the drift includes the score term $\nabla_{x}\log p_{t}(x)$ , which is parameterized via a conditional time dependent U-Net noise predictor $\hat{\epsilon}_{\phi}(x_{t},\mu,t)$ .

Figure 2: Overview of our data generation scheme. a) MRI slices were segmented and augmented to maximize anatomically informative content. b) The processed 2D grayscale MRI samples were simulated into B-mode US images using the MUST toolbox. The pipeline output paired high quality and low quality images for training.

\tilde{m}_{t}=\tilde{\mu}+(\tilde{x}_{0}-\tilde{\mu})\,e^{-\bar{\theta}_{t}},\qquad\tilde{v}_{t}=\lambda^{2}\bigl(1-e^{-2\bar{\theta}_{t}}\bigr)

Algorithm 1 IRSDE-Despeckle Training with Normalization

Figure 3: Visual comparison of despeckling on a simulated ultrasound image from the held-out test dataset from UMD. Columns show the paired LQ–HQ images, our IRSDE despeckling result, and representative comparisons (DnCNN, OBNLM, MSRResNet, Speckle2Self, and ultrasound denoising GAN). The red boxes in the top row mark an example region of interest, and the corresponding zoomed-in croped images are shown in the bottom row. Our method suppresses speckle while preserving edges and contrast in the highlighted region, whereas several comparison SOTA methods either leave residual speckle or oversmooth fine structures and attenuate local contrast.

Figure 4: Quantitative and visual comparison on a real ultrasound image from BrEaST Lesions USG dataset using contrast-to-noise ratio (CNR). For this visualization, CNR is shown in decibels as $\mathrm{CNR}_{\mathrm{dB}}=20\log_{10}(\mathrm{CNR})$ . The first panel shows the speckled input with the foreground ROI (solid red) and background ROI (dashed red) used for CNR computation. Remaining panels show the despeckled outputs from baseline methods and our method (IRSDE Despeckle), with the corresponding CNR value overlaid. Our method achieves higher CNR and better preserves lesion-to-background contrast while suppressing speckle, whereas several baselines either oversmooth and reduce contrast or leave residual texture that lowers CNR.

Figure 5: Model prediction uncertainty and error visualization using a 5-fold cross-validation ensemble on selected samples from the MRNet knee MRI dataset. Examples with lower cross-fold variance are shown in the top rows, while examples with higher variance are shown in the bottom rows. Regions with higher variance indicate greater disagreement among fold models (higher epistemic uncertainty) and often coincide with localized hallucination artifacts and larger reconstruction errors. Images are displayed in 8-bit grayscale (0–255) for visualization, whereas MSE and variance are computed in the normalized image space.

Figure 6: Relationship between cross-model variance and reconstruction error (MSE) on the held-out UMD MRI test set. The x-axis represents the normalized per-image cross-fold predictive variance $\sigma^{2}$ estimated from a 5-fold ensemble, while the y-axis represents the normalized per-image mean squared error with respect to the HQ target. The dashed line shows the least-squares linear fit with Pearson correlation $r=0.60$ , indicating that higher ensemble variance, and thus higher prediction uncertainty, is associated with larger reconstruction error.

Figure 7: Qualitative comparison of transducer specific degradation and restoration on the Duke Liver MRI dataset. The leftmost panel shows the high quality reference image, the top row presents simulated ultrasound images generated in MUST under different transducer settings ( L11-5v , L12-3v, C5-2v, and P4-2v), and the bottom row shows the corresponding despeckled predictions produced by our method.

Transducer Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
C5-2v	$21.29\pm 3.22$	$0.47\pm 0.11$	$0.51\pm 0.08$
P4-2v	$20.24\pm 3.47$	$0.43\pm 0.11$	$0.48\pm 0.10$
L12-3v	$23.92\pm 3.23$	$0.58\pm 0.12$	$0.39\pm 0.12$
L11-5v	$24.11\pm 3.16$	$0.56\pm 0.10$	$0.39\pm 0.11$

Table 2: Despeckling performance of our despeckling model on simulated test images using the held out UMD dataset generated with different transducer settings. Metrics are reported as mean

\pm

standard deviation over all test images.

Sampling Steps $T$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time (s) $\downarrow$
$100$	20.95	0.60	0.30	1.04
$50$	20.49	0.58	0.30	0.52
$20$	17.90	0.42	0.43	0.21
$10$	14.17	0.22	0.66	0.11

Table 3: Quantitative despeckling performance and average inference time per

128\times 128

input image for models trained with different numbers of reverse diffusion steps

T

. Metrics are reported on the held-out UMD test set.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Qualitative comparison on ultrasound data simluated from full-view MRI 2D slices. Columns show the low-quality (LQ) speckled input, high-quality (HQ) reference patch, and outputs from baseline methods and IRSDE Despeckle. Due to space limitations, we show selected learning-based baselines. The examples illustrate tradeoffs between speckle suppression and structural preservation.

Figure 2: Qualitative comparison on ultrasound data simulated from ROI patches of MRI 2D slices. Columns show the low-quality (LQ) speckled input, high-quality (HQ) reference patch, and outputs from baseline methods and IRSDE Despeckle. Due to space limitations, we show selected learning-based baselines. The examples illustrate tradeoffs between speckle suppression and structural preservation.

Method	CNR $\uparrow$	ENL $\uparrow$	Speckle SNR (dB) $\uparrow$	BRISQUE $\downarrow$
DnCNN	0.80 $\pm$ 0.48	8.68 $\pm$ 6.19	8.25 $\pm$ 3.45	38.28 $\pm$ 11.79
FFDNet	0.74 $\pm$ 0.44	7.08 $\pm$ 4.56	7.44 $\pm$ 3.42	28.54 $\pm$ 8.73
IMDN	0.85 $\pm$ 0.49	10.08 $\pm$ 7.07	8.76 $\pm$ 3.81	58.16 $\pm$ 8.70
IRCNN	0.76 $\pm$ 0.47	9.14 $\pm$ 6.61	8.40 $\pm$ 3.60	27.61 $\pm$ 7.21
MSRResNet	0.85 $\pm$ 0.49	9.96 $\pm$ 7.01	8.71 $\pm$ 3.81	56.83 $\pm$ 7.42
SRMD	0.85 $\pm$ 0.49	9.99 $\pm$ 7.01	8.72 $\pm$ 3.82	60.20 $\pm$ 10.72

Appendix

Table 1: Referenceless ultrasound image quality metrics (mean

\pm

std) across all images. Best value in each column is in bold.

Panel of Figure 3 — Figure 3: Additional referenceless qualitative examples used for image quality assessment on the BrEaST tumors USG dataset. Each row corresponds to a separate test case, showing the input image and the outputs of the compared despeckling methods, along with the associated referenceless quality metrics when available. The red contour in the input image indicates the tumor annotation provided by the dataset. For clarity, we omit the background ROI overlay in this figure; an example showing both the foreground and background ROIs is provided in the main text (Fig. 4 ). Quantitative metrics include CNR, ENL, Speckle SNR, and BRISQUE, with green highlights marking the best value per metric across methods for that case. While some methods achieve higher scores on individual metrics, visual inspection reveals notable differences in structure preservation, edge fidelity, and tumor appearance, underscoring that metric improvements do not always align with perceptual or clinically relevant quality. These examples motivate interpreting referenceless metrics in conjunction with case-by-case visualization.

Explore similar work

Aug 30, 2026

Null-Space Diffusion Restoration with Adaptive Uncertainty-Guided Fusion for Ultrasound Speckle Reduction

Ultrasound B-mode imaging commonly suffers from speckle noise and artifacts, requiring a delicate balance between contrast, resolution, and preservation of anatomical structures. Although recently developed despeckling methods have achieved some progress, supervised learning approaches remain fundamentally limited by the ground truth paradox, which arises from the absence of noise-free, ground truth reference images in in vivo scenarios. Existing unsupervised diffusion-based methods typically enforce data consistency directly in the nonlinear log-compressed domain, which can disproportionately amplify background artifacts when mapped back to the envelope domain. To overcome these limitations, we propose an uncertainty-guided null-space diffusion (UGNS) framework, a novel label-free solution that enforces consistency correction on a stabilized positive-envelope proxy obtained via inverse log compression. The proposed UGNS introduces several technical novelties: (a) extraction of a structural prior in the stabilized envelope domain to produce a robust signal envelope that preserves anatomical structure, (b) development of an adaptive range-null reconstruction mechanism that uses an adaptive weight mask to preserve tissue regions via range-space projection, and (c) introduction of uncertainty-guided fusion in an adaptive way to mitigate sampling variability. Extensive and comparative experiments were conducted using the PICMUS benchmark and in vivo datasets. The results demonstrate that UGNS achieves competitive generalized contrast-to-noise ratio (gCNR) values across diverse datasets. In addition, it is successfully validated that UGNS effectively suppresses speckle noise while preserving fine spatial resolution. Code is available at https://github.com/yousirong/UGNS.git.

Juneyong Lee, Jaeyoung Choi

Department of Computer Science and Engineering, Hankuk University of Foreign Studies, Yongin-si 17035, Republic of Korea

Diffusion Models for Image Restoration Ultrasound Imaging Diffusion-Based Inverse Problems Diffusion Models Image Denoising

Ask PDF

Sep 26, 2026

Mask2Restore: Self-Supervised Ultrasound Despeckling via Inpainting

Medical ultrasound (US) is inherently degraded by speckle, a granular interference pattern that is often treated as a complex form of noise in image restoration. However, unlike random noise, US speckle originates from coherent scattering within tissue and is therefore highly spatially dependent and deterministic under fixed acquisition conditions, making US speckle suppression fundamentally different from natural image denoising. Because speckle-free US targets are unavailable in practice, self-supervised denoising is necessary. Blind-spot networks (BSN) are the dominant self-supervised paradigm for natural images, but their pixel-wise masking strategy assumes spatially independent noise, an assumption poorly matched to US speckle, which is spatially correlated over multiple pixels rather than pixel-wise independent. To address this mismatch, we propose Mask2Restore, a self-supervised US despeckling framework that reformulates despeckling as contextual inpainting with block-wise masking on single noisy images. Unlike pixel-wise BSN masking, block-wise masking addresses this multi-pixel speckle correlation by removing locally correlated speckle neighborhoods and shifting the reconstruction cues used by the network from adjacent speckle correlations to broader anatomical context. We further introduce cross-resolution context regularization (CRCR), which suppresses residual speckle bias by enforcing consistency across multi-resolution predictions. Experiments on simulated and in vivo carotid US, unseen fine-structure cases, and downstream cardiac segmentation demonstrate improved speckle-detail trade-offs, better preservation of fine anatomical structures, and practical value for subsequent image analysis.

Xuesong Li, Yingtai Xu, Zhongliang Jiang +2

Computer Aided Medical Procedures (CAMP), Technical University of Munich, Munich, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Department of Mechanical Engineering, The University of Hong Kong, Hong Kong SAR, China

Image Restoration Image Denoising Self-Supervised Image Denoising

Ask PDF

May 23, 2026

Physics-Guided Self-Supervised Statistical Residual Learning for Sonar Despeckling with Improved Generalization

This letter introduces a physics-informed self-supervised framework for sonar image despeckling that reformulates despeckling as residual consistency in the homomorphic log domain. By constraining the log-ratio residual to obey multiplicative speckle statistics, the proposed method eliminates the need for clean supervision while preventing degenerate identity solutions. A variance-targeted statistical loss combined with edge-aware structural regularization and median-guided curriculum stabilization enables effective speckle suppression with preserved structural fidelity. This formulation along with a lightweight neural network achieves state-of-the-art performance across multiple real sonar datasets and demonstrates excellent cross-dataset robustness, while remaining suitable for real-time deployment.

Swapna Pillai, Siddharth Singh Savner, Sujit Kumar Sahoo

School of Electrical Sciences, Indian Institute of Technology Goa, Goa 403401, India · Inria, Sophia Antipolis, France

Self-Supervised Image Denoising Self-Supervised Learning Underwater Imaging Residual Learning Physics-Guided Image Restoration Physics-Informed ML+1

Ask PDF