acceptodds
Under review as a conference paper at ICLR 2027

PSDN: Physics-Guided Self-Supervised Dual-Stream Network for Ultrasound Image Denoising

Abstract

Ultrasound imaging is widely used in clinical diagnosis but inherently suffers from speckle noise, which degrades image quality and obscures vital anatomical details. Existing deep learning–based denoising methods are severely constrained by the lack of clean ground-truth images, while current self-supervised approaches often neglect acoustic physical mechanisms and rely on single-stream architectures, resulting in frequency aliasing and an inherent trade-off between noise suppression and edge preservation. To address these issues, we propose a Physics-Guided Self-Supervised Dual-Stream Module (PSDN) that effectively suppresses speckle noise while preserving anatomical structures and fine details. First, we propose a Physically Guided Adaptive Frequency Decomposition (PAFD) module to alleviate the aliasing between tissue details and speckles by extracting acoustic physical priors and dynamically decomposing the noisy image into purified frequency subbands. Second, a Structure-Detail Dual-Stream (S2DS) module is proposed to balance macro-structure recovery and micro-detail modeling by processing the decomposed subbands through parallel low- and high-frequency branches. Extensive experimental results across multiple datasets demonstrate that PSDN achieves state-of-the-art denoising performance without requiring clean reference images.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Ultrasound image denoising, Self-supervised learning, Physics-guided frequency decomposition.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FE7iOpenReview ↗
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
  1. 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.
  2. 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
  1. 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.
  2. 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
  1. 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)?
  2. 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.
  3. In real experimental ultrasound images, I would expect independent electronic noise as well. Is that addressed or a concern at all?
  4. 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?
  5. 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?
  6. I would like to see the subband naming and minor issues in the weaknesses sections fixed.
7 comments
AuthorsOpenReview ↗

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.

TO Q1

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 Q2

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.

Physical prior ablation experimentENL ↑\uparrowCNR ↑\uparrowHom ↑\uparrow
Full17.361.780.578
w/o A16.611.710.560
w/o S16.321.670.552
w/o T18.021.580.603
Loss term ablation experimentENL ↑\uparrowCNR ↑\uparrowHom ↑\uparrow
Full17.361.780.578
w/o Ltv15.941.620.541

TO Q3

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.

AuthorsOpenReview ↗

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 outputPSNR ↑\uparrowSSIM ↑\uparrowHom ↑\uparrowLPIPS ↓\downarrow
Speckle2Self final output18.590.7710.8030.231
PSDN final output16.340.6210.5360.484
PSDN pseudo-target13.760.3930.3420.705

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 settingENL ↑\uparrowCNR ↑\uparrowHom ↑\uparrow
Full PSDN17.361.780.578
S2DS removed and replaced with a single-level U-Net13.851.480.421
Single-stream reconstruction with a two-level U-Net14.081.520.458
Attention gating removed from the low-frequency branch15.621.640.519
Standard residual blocks in the high-frequency branch16.311.70.546

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.

Reviewer_x6JnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PSDN, a label-free training framework for ultrasound image despeckling without the need for paired clean reference images. The key features of this framework are as follows: Physics-Guided Adaptive Frequency Decomposition (PAFD) module derives image-statistical proxies for attenuation, scattering and tissue structure from the input image, using them to spatially modulate subbands produced by four fixed Log-Gabor filters. The Structure-Detail Dual-Stream (S2DS) network processes lower and higher frequency subbands via separate branches and fuses their features for reconstruction. The training objective combines pseudo-target reconstruction, scattering-weighted total variation and subband-response regularisation. The authors validated the method on three real-world ultrasound datasets covering the carotid artery, kidney and calf muscle, using metrics such as ENL, CNR, homogeneity and blind physician assessments. The reported ENL, CNR and physician scores showing superior performance to the compared baseline methods.

Critical strengths

The proposed method, PSDN, takes full account of the input conditions in real-world application scenarios and operates directly on B-mode ultrasound images without the requirement of paired clean targets or RF data. As regards the rationale behind the method’s design, it explicitly treats the highest-frequency subband as a mixture of anatomical details and speckle, using a soft retention mechanism based on structural proxies to address the risk of discarding useful high-frequency information. This sounds like a reasonable design choice for balancing despeckling and detail preservation under limited supervision.

The results across real ultrasound datasets demonstrate the effectiveness of the method. Table 1 illustrates the significant advantages in terms of performance metrics compared with a range of competing methods. During the evaluation phase, the preselection of ROIs shared across methods and independent, blind assessments by clinicians of anonymised, randomly ordered outputs strengthen the credibility of the comparisons.

Critical weaknesses

There exists limited evidence to support the validity of the structure-detail modelling mechanism. The single-stream ablation in Table 3 and the visualisation comparisons in Fig. 5 provide initial support for the effectiveness of the S2DS approach. However, the single-stream baseline lacks sufficient definition and instructions regarding its configuration. Consequently, it is difficult to attribute the final performance improvement entirely to the frequency-specific processing and structure-detail complementarity independently of other architectural differences such as the parameter capacity. Further experiments may be required to demonstrate the additional benefits of the approach itself by establishing more representative control groups.

The paper evaluates the proposed method on three real-world datasets using the ENL, CNR, and Hom metrics and aggregate physician scores. These results provide useful evidence to support the quality of the denoised image, but there is a lack of quantification to assess the reconstruction fidelity or anatomical-detail preservation against known references. The paper does not report reference-based ultrasound simulations, which have been used for evaluation in related work such as Speckle2Self (Li et al., 2025). A limited simulation study with well-defined reference targets, including comparisons with the baselines, would provide more direct evidence for the claimed balance between speckle suppression and anatomical preservation.

Questions for authors

How are the final images reconstructed in the PAFD-only ablation in Table 2 and Fig. 4? Is it strictly equivalent to the pseudo-target defined in Eq. (11), including the R4R_4 weighting and clipping, or does it involve additional reconstruction or post-processing?

Please describe the architecture, input subbands, and the number of parameters of the single-stream baseline in Table 3, and explain how its training setup corresponds to those of S2DS. What evidence supports attributing the gains specifically to the separation of low-frequency and high-frequency processing, independently of other architectural differences?

Section 4.2 suggests that stronger smoothing may explain Speckle2Self’s higher Hom. Given that Appendix F.4 evaluates Hom in homogeneous ROIs with strong boundaries excluded, could ROI-level evidence be provided to clarify whether PSDN’s lower Hom reflects retained tissue texture or residual speckle?

Would it be possible to provide a limited reference-based ultrasound simulation study comparing the final outputs of PSDN and the main baselines in terms of metrics such as PSNR and SSIM, while also evaluating the pseudo-targets against the same references? Such experiments may help quantify reconstruction errors and anatomical-detail preservation in both the pseudo-targets and final outputs.

2 comments
AuthorsOpenReview ↗

Response to Reviewer x6Jn

We thank the reviewer for recognizing the practical value of PSDN for real B-mode ultrasound images and its soft retention design for the highest-frequency subband.

TO Q1

The PAFD-only variant in Table 2 and Fig. 4 retains the full PAFD module. The four resulting subbands are concatenated along the channel dimension and fed into a single-level U-Net, which directly generates the final despeckled image. The (R_4) weighting in Eq. (10) and the clipping operation in Eq. (11) are used to construct the self-supervised pseudo-target, which is used only to compute the reconstruction loss during training. Figure 4 shows the U-Net predictions. The network outputs are evaluated directly without additional image post-processing.

The baseline in Table 2, with both PAFD and S2DS removed, uses a single-level U-Net with the same architecture. A shallow stem extracts input features from the original image. Thus, both variants use the same reconstruction backbone, receiving the PAFD subbands and shallow features from the original image, respectively. We have clarified the input and output procedures for both variants in the revised manuscript.

TO Q2

The single-stream baseline in Table 3 concatenates the four subbands generated by PAFD along the channel dimension and feeds them into a two-level U-Net with two downsampling and two upsampling stages. We adjusted the channel widths of the U-Net to bring its parameter count close to that of S2DS. Both models use the same subbands, pseudo-targets, loss functions, training settings, and evaluation data. The single-stream baseline processes all four subbands jointly, whereas S2DS feeds the first two subbands into the structure branch and the last two into the detail branch, then fuses the branch features for reconstruction.

To assess the effect of frequency-specific input assignment, we added a dual-branch control with mixed-frequency inputs. This control retains both S2DS branches, the fusion module, and the reconstruction head. It concatenates the four subbands, extracts features through a shared shallow stem, and feeds the same mixed-frequency features into both branches. We matched the parameter counts as closely as possible and used identical training and evaluation settings. Compared with the mixed-frequency control, the full S2DS achieved higher ENL, CNR, and Hom, indicating that assigning the low- and high-frequency subbands to the structure and detail branches, respectively, helps improve despeckling performance.

Branch input configurationENL ↑\uparrowCNR ↑\uparrowHom ↑\uparrow
Frequency-specific inputs (full S2DS)17.361.780.578
Shared mixed-frequency input14.911.620.525

TO Q3

As described in Appendix F.4, Hom is computed in homogeneous ROIs that exclude prominent boundaries and measures local gray-level uniformity. Both residual speckle and subtle tissue texture can reduce uniformity, so Hom should be interpreted alongside local image observations.

We additionally examined homogeneous ROIs and regions containing tissue texture at the same locations across the outputs of all methods. Magnified views of the homogeneous regions show that PSDN suppresses visibly granular speckle. Magnified views of the textured regions show that PSDN preserves spatially continuous tissue texture. The observations in textured regions provide additional evidence of detail preservation. The current results from homogeneous ROIs leave the respective contributions of residual speckle and tissue texture to the difference in Hom unresolved. We have added these local comparisons to the revised manuscript and presented the explanation for Speckle2Self’s higher Hom in Section 4.2 as a possible interpretation.

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 outputPSNR ↑\uparrowSSIM ↑\uparrowHom ↑\uparrowLPIPS ↓\downarrow
Speckle2Self final output18.590.7710.8030.231
PSDN final output16.340.6210.5360.484
PSDN pseudo-target13.760.3930.3420.705

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.

Thank you for the response, which clarified how the different components of the method were designed.

Reviewer_1wgpOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PSDN, a physics-guided self-supervised framework for ultrasound image despeckling without requiring clean reference images. The method introduces a Physics-Guided Adaptive Frequency Decomposition (PAFD) module that incorporates image-derived attenuation, scattering, and structural priors, followed by a Structure-Detail Dual-Stream (S2DS) network that separately processes low- and high-frequency components. Experiments on three ultrasound datasets demonstrate improvements over existing self-supervised denoising methods in equivalent number of looks (ENL), contrast-to-noise ratio (CNR), and physician evaluation scores. The main contribution is the integration of physics-inspired frequency decomposition with a dual-stream architecture to balance speckle suppression and anatomical detail preservation.

Critical strengths
  1. The paper addresses ultrasound despeckling without clean ground-truth images, which is an important challenge in medical imaging. The proposed self-supervised approach has practical value because obtaining noise-free reference images is generally infeasible in clinical ultrasound.

  2. The study shows consistent experimental improvements. PSDN achieves the highest ENL, CNR, and physician scores across all three evaluated datasets. The ablation studies also show that both PAFD and S2DS contribute to the reported improvements. These results provide reasonable evidence of the effectiveness of the proposed framework.

Critical weaknesses
  1. Limited methodological novelty and insufficient justification of the physics-guided design. The main components, including Log-Gabor frequency decomposition, handcrafted statistical priors, and dual-stream reconstruction, are largely based on established techniques. The proposed physics-guided priors are derived from image statistics rather than an explicit ultrasound image formation model. Although the authors demonstrate the effectiveness of the overall PAFD module, they do not isolate the contributions of individual priors or compare against a standard frequency decomposition without physics guidance. This makes it difficult to determine whether the claimed physics-guided formulation provides a meaningful improvement over simpler alternatives.

  2. Insufficient validation of anatomical detail preservation. The evaluation relies primarily on ENL, CNR, homogeneity, and subjective physician scores. These metrics do not directly measure whether fine anatomical structures are faithfully preserved, and improvements may partly result from stronger smoothing or contrast enhancement. Although the physician evaluation is useful, no inter-reader agreement or statistical significance analysis is reported. Additional quantitative evidence of structural fidelity, such as boundary preservation measurements or downstream segmentation performance, would strengthen the main claim of the paper.

Questions for authors
  1. How much improvement comes specifically from the proposed physics-guided priors? Could the authors provide an ablation comparing the full PAFD module against standard Log-Gabor decomposition and variants without individual attenuation, scattering, or structural priors?

  2. How do the authors verify that PSDN preserves genuine anatomical details rather than introducing or removing subtle structures? Could quantitative downstream task evaluations or anatomical boundary measurements be provided?

  3. In the supplementary material, the energy regularization term LeneL_{\mathrm{ene}} is calculated from the input image and frequency subbands. Since the Log-Gabor filters and physics-inspired priors appear to be fixed, does this term provide any gradient to the trainable reconstruction network? If not, what is its role in optimization?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1wgp

We thank the reviewer for their assessment of PSDN and for raising specific questions about the physics-guided priors, anatomical detail preservation, and the energy term.

TO Q1

The first experiment compares the full PSDN with a variant that uses standard four-band Log-Gabor decomposition. In the full model, PAFD uses attenuation, scattering, and tissue structure priors to spatially modulate the Log-Gabor subband responses. In the control variant, the four subbands produced by the same filter bank are fed directly into S2DS without prior-guided modulation. Both variants use the same reconstruction network and training settings, allowing us to isolate the contribution of prior-guided modulation.

Frequency decompositionENL ↑\uparrowCNR ↑\uparrowHom ↑\uparrow
Full PSDN with PAFD17.361.780.578
Log-Gabor (no priors)14.921.550.521

The second experiment uses the full PAFD as the baseline and evaluates three variants that remove the attenuation prior ((A)), the scattering prior ((S)), or the tissue structure prior ((T)). Each variant removes one prior while retaining the other two, allowing us to assess its contribution within the full framework.

Physics-guided prior ablationENL ↑\uparrowCNR ↑\uparrowHom ↑\uparrow
Full17.361.780.578
w/o A16.611.710.560
w/o S16.321.670.552
w/o T18.021.580.603

The full model outperformed the standard Log-Gabor variant across all three metrics, indicating that the physics-guided priors account for part of PSDN’s performance gains. The individual ablations further show that the attenuation and scattering priors improve despeckling performance. Removing the tissue structure prior increased ENL and Hom while reducing CNR, suggesting that this prior helps preserve tissue-to-background contrast while balancing the suppression of local intensity fluctuations.

TO Q2

We conducted an additional left ventricular cavity segmentation experiment on the CAMUS echocardiography dataset. We used U-Net to segment the original images and the images despeckled by PSDN, and compared the predictions with the reference annotations. We used the Dice coefficient to measure region overlap and the Hausdorff distance (HD) to measure boundary discrepancy.

U-Net inputDice ↑\uparrowHD ↓\downarrow
Original ultrasound images0.9034.12
PSDN-despeckled images0.9523.68

After PSDN processing, the predicted regions overlapped more closely with the reference annotations, and the predicted boundaries showed smaller deviations. Together, these results support the preservation of anatomical detail in the left ventricular cavity and along its boundary.

TO Q3

We thank the reviewer for pointing this out. Lene is computed from the input image and the frequency subbands generated by the fixed PAFD module. Therefore, its gradient with respect to the trainable parameters of the reconstruction network is zero, and it does not participate in network parameter updates. This term quantifies the relative discrepancy between the energy of the input image and the total response energy of the subbands, helping us monitor whether frequency decomposition causes substantial changes in response magnitude. In the original manuscript, this monitoring quantity was included alongside the training losses in the total loss equation, which may lead readers to mistakenly believe that it also participates in optimization. In the revised manuscript, we will remove Lene from the total loss equation, provide its definition separately, and clarify that it is used to monitor the response magnitude during frequency decomposition.

AuthorsOpenReview ↗

Response to Reviewer 1wgp

Thank you for your valuable feedback. We have responded to your questions and would appreciate it if you could review our responses when convenient and let us know whether they address your concerns. We would be happy to provide any further clarification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.