- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3(+1)
- Soundness
- 2
Read review
Paper summary
This paper introduces PSDN, a self-supervised, physics-informed dual-stream denoising method that specifically targets ultrasound imaging, where noise is characterized as spatially correlated speckles. Its primary contribution starts with a disentanglement that isolates structures from noise by decomposing sensor data into 4 radial frequency subbands using a log-Gabor filter. Two novel components then act on this disentangled frequency domain data: the Physics-Guided Adaptive Frequency Decomposition (PAFD) component modulates the target subbands based on physics-inspired priors (acoustic attenuation, local scattering, and tissue structure), where low-frequency bands are treated as structural proxies and high-frequency bands are treated as a mixture of detail and noise. Subsequently, a Structure-Detail Dual-Stream (S2DS) module takes the split low- and high-frequency bands and reconstructs them separately to handle image details. This combination effectively removes structure-correlated speckle noise in Ultrasound imaging compared with other existing methods, under metrics that are specifically designed for cases where noiseless labels are unavailable.
Critical strengths
- The overall method of the paper is well-motivated: both the subband disentanglement and physics-inspired priors modulating the target subband tie to ultrasound imaging physics, along with theoretical and prior-work justifications.
- The evaluation metrics are carefully chosen for the case where no ground-truth data is available. The authors give a rationale for each metric and guidance on how to read it. Moreover, the blind physician scoring links denoising quality to a clinically relevant judgment, which is a unique and valuable addition.
Critical weaknesses
- The proposed method’s training budget and model capacity are not controlled for in the results section. The baseline configurations are never introduced (parameters, training iterations, data, tuning), and PSDN’s dual-stream design may simply have more capacity. Since the paper’s central claim is that the physics-informed design drives the improvement, the gains cannot be attributed to it without a capacity- and budget-matched comparison.
- The paper lacks clarity when describing the method. In Fig. 2A, the authors named each subband as horizontal, vertical, and diagonal details, but according to Section 3.1, those names should correspond to the high-frequency subbands only; furthermore, the proposed log-Gabor filters are isotropic by the authors' definition, so directional names are not justified for any subband. Minor issues include text in Fig. 2B being too small to read and Appendix B having citation errors.
Questions for authors
- What motivated the chosen log-Gabor filter parameters? The ablation covers only small variations around them, but the reported values seem to overlap heavily. Have you tested substantially different band configurations, and can you show, using simulated data where speckle and structure are known, how speckle energy is distributed across the four bands (e.g., that the low-frequency bands are largely free of it)?
- I would like to see how much each physics prior, as well as the loss term, contributes to the overall denoising performance; that should be an ablation item as well.
- In real experimental ultrasound images, I would expect independent electronic noise as well. Is that addressed or a concern at all?
- Though the usage of label-free metrics is justified, prior work has also compared against simulated data to obtain absolute metrics, including PSNR and SSIM. Have any experiments been done against simulated datasets to validate the theoretical ceiling of the denoising capabilities of the proposed method?
- It seems like the S2DS novelty is limited. What is the reasoning behind the different architecture for low- versus high-frequency information? Are there architecture ablations on the S2DS module itself?
- I would like to see the subband naming and minor issues in the weaknesses sections fixed.
7 comments
Response to Reviewer FE7i Q4 to Q6
TO Q4
We conducted an additional reference-based simulation experiment on the synthetic ultrasound scene dataset used by Speckle2Self. For each noisy input, we used the corresponding clean image as a common reference to evaluate the final output of Speckle2Self, the final output of PSDN, and the PSDN pseudo-target . The clean reference images were used only for evaluation, and PSDN followed its original self-supervised training procedure. The results are shown below.
| Method or output | PSNR | SSIM | Hom | LPIPS |
|---|---|---|---|---|
| Speckle2Self final output | 18.59 | 0.771 | 0.803 | 0.231 |
| PSDN final output | 16.34 | 0.621 | 0.536 | 0.484 |
| PSDN pseudo-target | 13.76 | 0.393 | 0.342 | 0.705 |
On this synthetic dataset, PSDN’s reconstruction performance remains below that of Speckle2Self. Comparing the final PSDN output and the pseudo-target against the same clean reference images shows that S2DS further improves reconstruction fidelity relative to the pseudo-target, bringing the final output closer to the clean reference images in terms of structural and perceptual features.
TO Q5
The rationale for using different architectures for low- and high-frequency information is that the low-frequency subbands mainly contain large-scale tissue contours and slowly varying regional intensities. Therefore, the structure branch uses an encoder–decoder with attention gating to recover overall contours and structural continuity. The high-frequency subbands contain tissue boundaries, fine textures, and speckle, so the detail branch maintains the original resolution and uses dilated residual blocks to extract multiscale local features. The features from the two branches are then fused to reconstruct the despeckled image.
We performed two types of ablation using the full PSDN as the baseline. In the module-level ablation, we retained PAFD, concatenated the four subbands it generated, and fed them into a single-level U-Net to replace the entire S2DS module. The architecture-level ablation included three variants. We concatenated the four subbands and fed them into a two-level U-Net with a similar parameter count for single-stream reconstruction, removed attention gating from the low-frequency branch, and replaced the dilated residual blocks in the high-frequency branch with standard residual blocks. The full S2DS performed better than the single-stream baseline, supporting separate modeling of low-frequency structure and high-frequency detail before fusion. The ablations within the branches further show that both attention gating in the low-frequency branch and dilated residual blocks in the high-frequency branch help improve despeckling performance.
| Ablation setting | ENL | CNR | Hom |
|---|---|---|---|
| Full PSDN | 17.36 | 1.78 | 0.578 |
| S2DS removed and replaced with a single-level U-Net | 13.85 | 1.48 | 0.421 |
| Single-stream reconstruction with a two-level U-Net | 14.08 | 1.52 | 0.458 |
| Attention gating removed from the low-frequency branch | 15.62 | 1.64 | 0.519 |
| Standard residual blocks in the high-frequency branch | 16.31 | 1.7 | 0.546 |
TO Q6
In keeping with the radial and isotropic nature of the Log-Gabor filters, we removed directional labels such as “horizontal,” “vertical,” and “diagonal” from Fig. 2A and replaced them with descriptions that reflect the dominant frequency scale and response characteristics of each subband. We also enlarged the text in Fig. 2B and checked and corrected the citation errors in Appendix B. We have also added details on the baseline configurations, training budget, and model capacity to the revised manuscript.
I thank the authors for their efforts and explanations in providing additional experimental details and analysis. The response resolves my questions #1-3. However, I am yet to be convinced on items #4-6 since I did not see an updated manuscript. Has there been a revision? I would like to review the changes and additional settings/results, as I am leaning towards modifying my rating if the rest of my concerns are resolved.
Response to Reviewer FE7i Q1 to Q3
We thank the reviewer for recognizing the physical motivation behind PSDN and the value of including blinded physician ratings in our evaluation.
We use four radial filters with center frequencies of ((0.05, 0.15, 0.30, 0.45)) and bandwidth parameters of ((0.08, 0.12, 0.15, 0.10)) to cover a range of spatial frequencies from low to high. The center frequency determines the image scale to which each filter primarily responds, and the bandwidth controls its frequency coverage and overlap with adjacent bands. We also compared band configurations that differed substantially from the default settings. The results showed that placing the bands too close together led to redundant responses, whereas spacing them too far apart led to insufficient frequency coverage. Narrow bandwidths left gaps in coverage, whereas wide bandwidths increased subband overlap. Based on these comparisons, we selected the parameter configuration reported in the manuscript.
Regarding the distribution of speckle energy across the four bands, we do not assume that the low-frequency subbands are entirely free of speckle. Ultrasound speckle is spatially correlated, and prior studies suggest that it can still be present in low-frequency subbands. In our method, the low- and high-frequency labels indicate dominant responses at different spatial scales, with structural and speckle responses potentially coexisting within each band. PAFD further uses attenuation, scattering, and tissue structure priors to modulate subband responses across spatial locations. We treat the highest-frequency subband as a mixed response that may contain both speckle and genuine tissue detail.
To evaluate the contributions of the individual physical priors and training losses to despeckling performance, we conducted two additional sets of ablation experiments on the carotid ultrasound dataset. The first set is the physical prior ablation. Using the full PSDN as the baseline, we removed only one prior at a time, namely the attenuation prior ((A)), the scattering prior ((S)), or the tissue structure prior ((T)). We retained the other two priors and used the same network architecture and training settings. The results are shown below. After removing either the attenuation prior or the scattering prior, all three metrics decreased, indicating that both priors help improve despeckling performance. After removing the tissue structure prior, ENL and Hom increased while CNR decreased, indicating that the tissue structure prior helps maintain tissue-to-background contrast while suppressing local fluctuations.
The second set is the loss term ablation. Following the reviewer’s comment, we checked the definition of the energy term Lene in the original manuscript. This term is computed from the input image and the fixed PAFD subbands. Its gradient with respect to the trainable parameters of the reconstruction network is zero, and it does not participate in parameter updates. It is used to monitor the relative discrepancy between the energy of the input image and the total response energy of the subbands. In the revised manuscript, we will remove it from the total loss equation and explain its monitoring role separately.
The pseudo-target reconstruction loss Lrec provides the network with the basic reconstruction signal. Therefore, we retained this term in both configurations and removed only the scattering-weighted total variation loss Ltv to evaluate its additional contribution to pseudo-target reconstruction. The results are shown below. After removing Ltv, ENL, CNR, and Hom decreased by 1.42, 0.16, and 0.037, respectively. This indicates that under the same pseudo-target reconstruction constraint, the scattering-weighted total variation loss helps reduce local gray-level fluctuations and improve tissue-to-background contrast.
Independent electronic noise is indeed a limitation of this study. Existing ultrasound despeckling methods mainly focus on speckle in images and its spatial correlation. PSDN likewise focuses on speckle suppression and does not explicitly model or specifically address independent electronic noise generated by the imaging system hardware. Our evaluation uses ultrasound images with clearly visible anatomical structures and no obvious hardware artifacts, and all methods are compared on the same images. Therefore, the reported results reflect PSDN’s despeckling performance on such images, and its ability to suppress electronic noise remains to be verified.