acceptodds
Under review as a conference paper at ICLR 2027

Bias-Structure-Correlation Trade off Learning for Self-Supervised Diffusion MRI Denoising

Abstract

Spatially correlated noise violates the input–target independence assumed by self-supervised diffusion MRI denoising. We identify a three-way trade-off among residual noise correlation, perturbation-induced structural distortion, and model-induced supervision bias, and propose a tensor-guided framework to balance them. By decomposing prediction error into conditional noise bias and structural mismatch, we derive an anisotropy-dependent bound on angular signal distortion under the diffusion tensor model. This bound motivates FA-guided spatio-angular shuffling that limits perturbations in highly anisotropic regions. To address residual noise dependence under restricted perturbations, we introduce a complementary tensor-reconstructed view that excludes the target direction from fitting, preventing direct target-noise reuse. Bidirectional auxiliary guidance and cross-view consistency combine the views to mitigate noise-related and model-induced biases. Experiments on a digital phantom and two in vivo datasets demonstrate improved DW image quality and preservation of diffusion-derived structures. Under 10% Rician noise, our method achieves 34.26 dB PSNR, outperforming the strongest baseline by 0.79 dB, and reduces FA estimation error by approximately 30% relative to the best competing method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-Supervised Denoising, Diffusion MRI, Bias-Structure-Correlation, Trade-off, Tensor-Guided

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xbcyOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a new self-supervised approach for denoising diffusion MRI 4D data, which extends the blind spot/splitting framework (which relies on splitting the noisy data in two subsets with noise realizations that are independent of each other) to diffusion MRI, taking into account the fact that noise is typically correlated in the spatial axes (due to parallel/accelerated imaging) and not on the angle axis. The paper also proposes using an additional branch network that improves the reconstructions further. Experiments on diffusion MRI data (denoising a single 4D instance at a time) show better performance than similar algorithms such as Patch2Self.

Critical strengths
  • The paper shows better performance than other self-supervised denoising frameworks specific to diffusion MRI data
Critical weaknesses
  • I find the paper extremely hard to follow. Having both significant experience in self-supervised denoising methods and in MRI reconstruction algorithms (albeit I'm not an expert on the specific case of diffusion MRI), I'm really struggling to understand the contributions.

    • The abstract needs an almost full rewrite: it is overly specific and fails to communicate in simpler terms that do not require the reader to have read the paper already (eg "tensor-guided framework", "FA-guided spatio-angular shuffling", etc.), uses unexplained acronyms, and is very hard to follow.
    • Many diffusion-MRI-specific points are not introduced ("FA", subfigures of Fig. 3, does "dataset" entail a single patient or a set of patients?)
    • The mathematical notation is unnecessarily dense, with a large number of symbols and a complex handling of subindices and subscrits on every symbol. The math formulations require an almost full rewrite, trying to focus on the essential symbols and simplifying the indexing logic. I believe that the main idea of the paper, ie wisely choosing the splitting strategy for diffusion MRI data, should be explainable without that level of complexity.
    • Equation 11 is not very clear, what is DD there? What are the various indices?
    • There are multiple flawed sentences, e.g., "Simulation dataset under Self-supervised performance also depends critically on how training observations are constructed."
    • I did not understand the role of fauxf_{\text{aux}} from section 3.3 - how does it differ from the main blind spot branch?
  • The paper seems to imply a few times that correlation between two pixels of the clean signal means having similar values or a low-lipschitz constant, e.g., "This tension is particularly important in dMRI because signal variation depends on both spatial location and diffusion direction" - in any image, signal variations are spatially correlated (for example, in natural images two neighbouring pixels might be strongly correlated without necessarily having similar values)?

Questions for authors
  • How does the computational time of your method compare with other baselines? Given the large scale of the 4D data, I would expect this to be a major factor when choosing an algorithm
  • Why do you train a network on one patient (a 4D dataset) at a time? Most splitting/blind spot methods are trained across a large number of subjects/patients/scenes and then simply evaluated on new ones.
3 comments
AuthorsOpenReview ↗

Response to Reviewer xbcy

W1: Our method is specifically designed for diffusion MRI denoising and exploits the directional dependence of diffusion signals. We therefore consider it necessary to retain established concepts such as the diffusion tensor and diffusion anisotropy to accurately explain the method’s rationale and theoretical contribution. However, we agree that their roles should be clear without requiring prior knowledge of the proposed framework. Accordingly, we have revised the abstract to explain how diffusion tensor information guides data perturbations, replaced “spatio-angular shuffling” with a direct description of shuffling across spatial locations and diffusion directions, and simplified the explanation of how the two views interact. We have also expanded FA, DW at their first occurrence. These revisions improve accessibility while preserving the technical precision and central contributions of the work.

Regarding dataset, it is the dataset of a single subject. This is common for self-supervised denoising in diffusion MRI, the well-established method such as Patch2Self, they also used the dataset of one subject to denoise.

As for mathematical notation, we agree that the notation and indexing in the original manuscript could be simplified. We also wish to clarify that our contribution extends beyond selecting a data-splitting strategy. Our theoretical analysis explains the trade-off between noise decorrelation and signal distortion induced by perturbations, providing a principled basis for anisotropy-guided shuffling. The complementary tensor-reconstructed view addresses residual noise dependence while introducing a distinct source of model-related error. Distinguishing these mechanisms clarifies why each component is needed and how they work together. Accordingly, we have removed unnecessary notation, standardized subscripts and superscripts, simplified indexing, and improved transitions between subsections. Detailed diffusion MRI formulations have been moved to the appendix so that the main text focuses on the essential analysis and methodological rationale. The original Eq. (11) describes tensor-based signal reconstruction: the diffusion tensor is estimated from measurements excluding the target direction and then used to reconstruct the target signal. We have revised this formulation to explicitly define the variables and indices specifying the directions used for tensor estimation and relocated the detailed calculation to the appendix.

During LaTeX editing, “Simulation dataset under” was inadvertently inserted into the sentence “Self-supervised performance also depends critically on how training observations are constructed.” This error has been corrected.

fauxf_{\mathrm{aux}} predicts one view from another using Noise2Noise-inspired supervision, whereas the main blind-spot branch predicts each voxel from its surrounding context, excluding its own measurement. Cross-view prediction provides complementary information.

W2: We agree that statistical correlation does not imply signal similarity or a small Lipschitz constant. Our Lipschitz condition is instead a local regularity assumption used to bound the clean-signal perturbation caused by shuffling in terms of the local Lipschitz constant and shuffling distance. While spatial correlation is also common in natural images, our statement emphasizes that shuffling in dMRI involves both spatial and diffusion-directional variations, without equating statistical correlation with local smoothness.

Q1: We have added a computational efficiency comparison with five representative baselines, reporting inference time, model parameters, and FLOPs, as shown in Table 1, page 7. Our method requires 237.86 s for inference, which is approximately 4.56× faster than MPPCA (1084.41 s), 9.10× faster than Patch2Self (2164.77 s), and 5.4% faster than Replace2Self (251.39 s). Although SAN2N (7.69 s) and UP2-MAE (155.60 s) achieve shorter inference times, our method maintains a competitive computational cost, with 3.677M parameters and 355.32G FLOPs. These results provide a quantitative assessment of the computational efficiency of our method and clarify its trade-off between denoising performance and inference cost for large-scale 4D dMRI processing.

Q2: Our method adopts a subject-specific self-supervised training strategy, as commonly used in dMRI denoising methods such as Patch2Self. It exploits the spatial and angular redundancy within each 4D dMRI dataset to learn directly from noisy measurements, without requiring external training data or clean references. This strategy allows the model to adapt to subject-specific signal characteristics and acquisition-dependent noise properties, particularly when diffusion-encoding schemes, b-values, and imaging protocols vary across datasets.

AuthorsOpenReview ↗

Response to Reviewer xbcy

We hope that our clarifications and additional experiments have addressed your main concerns. We would greatly appreciate your further feedback and welcome any remaining questions during the discussion period. We also hope that the new evidence and discussions will be taken into account in your final assessment.

AuthorsOpenReview ↗

Response to Reviewer xbcy

Dear Reviewer,

Thank you for your constructive feedback. We have taken your concerns seriously and made substantial revisions, including extensive additional experiments and theoretical clarification. We would like to briefly summarize our major efforts:

  1. Validation under spatially correlated noise: We added experiments with explicitly correlated Rician and nc-chi noise across multiple correlation levels. Our method consistently achieves the highest PSNR, outperforming the strongest baseline by 1.30 dB under the highest correlation level. We also added quantitative SNR evaluations on real dMRI data.

  2. Theoretical clarification and validation: We revised the error analysis to align with the actual self-supervised training objective, clarified the conditions for bias reduction, and provided empirical evidence supporting these conditions.

  3. Additional experiments and improved clarity: We conducted FA threshold sensitivity analyses, direction-exclusion ablations, and computational efficiency comparisons. We also substantially simplified the mathematical notation and improved the methodological explanations.

These revisions represent considerable experimental and analytical effort, undertaken specifically to address the concerns raised in the reviews. We sincerely hope that this new evidence will receive the careful consideration it deserves.

We would also like to respectfully emphasize that active participation in the discussion phase is an important responsibility of ICLR reviewers. The ICLR reviewing guidelines explicitly recognize responding to author rebuttals as a core part of reviewers' responsibilities. We believe that carefully considering and engaging with substantive author responses is essential to a fair and constructive review process, as well as a reflection of mutual respect between reviewers and authors.

We fully respect your right to maintain your original assessment. However, we would greatly appreciate a substantive response indicating whether our additional experiments and clarifications have addressed your concerns and, if not, which specific issues remain unresolved.

Thank you for your time and consideration. We look forward to your feedback.

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

This paper introduces a novel self-supervised denoising architecture that aims to remove spatially correlated noise in diffusion-weighted magnetic resonance images (dMRI). The pipeline depends on fractional-anisotropy (FA) mask-guided spatio-angular shuffling to reduce noise correlation, which demonstrates better feature preservation while successfully decorrelating noise. Then, a novel two-branch architecture processes 2 different views: the top branch reconstructs a tensor-fitted view of the raw dMRI data, excluding the current direction as well as the source direction if shuffled; the bottom branch performs a pixel shuffle using a neighboring pixel from a neighboring direction. Then, each view is denoised through a blind spot network while the other view is fed into the auxiliary network to provide additional structural information, before a learned fusion computes the final denoised images. On synthetic datasets, the proposed method achieved superior performance without requiring clean image ground truths; on real datasets, the qualitative results demonstrated a promising denoising capability.

Critical strengths
  1. The method and problem formulation are sound and rigorous. The authors provide detailed noise characterization and motivate the FA mask construction under Lipschitz continuity conditions to balance noise decorrelation and structural distortion. They pair this with principled denoising error decomposition and visual analysis of noise correlation and image-structure distortion (Fig. 1a, b). Furthermore, the cross-view dual processing leverages the structural similarities between the two reconstruction methods, thereby justifying this dual-branch architecture design. Overall, the core idea and the methodology are well explained.

  2. The results are comprehensive. The paper presented a decent-sized family of relevant, strong, and recent self-supervised denoising methods and demonstrated superiority in both DWI quality and downstream DTI metrics on phantom images. On top of strong performance against the baselines, the authors also provided a set of ablation studies to empirically verify the usefulness of each component, including architecture ablations and mask choice ablations. Lastly, the authors provided different noise distributions and conditions, further strengthening the robustness of the proposed method under different degradation schemes.

Critical weaknesses
  1. Clarity and bookkeeping are lacking. Important values of theoretical and empirical bounds that could define the method's behavior were labeled but never provided, including the FA threshold and the shuffling radius. Further, several quantities, such as fractional anisotropy, “patch” mask, and “perturbation levels”, were used but never defined in the main text, which slightly hinders readability since it is unclear what their meanings are.
Questions for authors
  1. Could you report the values for τFA\tau_{FA} in the FA mask definition and the shuffling radius, along with sensitivity tests around these quantities to see if the model performance changes?

  2. Is the direction exclusion practice ablated at all? It would be nice to see all directions versus q\\{q\\} removed versus q,rp\\{q,r_p\\} removed.

  3. In real datasets, what are the residuals supposed to show? And, under the assumption that no clean labels are available, are you able to compute metrics such as contrast-to-noise ratio or other label-free metrics to quantitatively support the method's performance in real datasets?

  4. Is there any evidence that this is going to be clinically relevant beyond scan-time reduction? And is hallucination observed at all? In Fig.4(b), does the proposed method visibly produce more fiber tracks than the noisy data?

3 comments
AuthorsOpenReview ↗

Response to Reviewer tDDW

W1: We have clarified the following points in the revised manuscript and appendix:

  1. Fractional anisotropy (FA) and threshold. FA is a dimensionless measure of diffusion anisotropy, calculated from the eigenvalues of a tensor fitted to noisy DWI signals across all acquired directions. The FA threshold is the 80th percentile of the estimated voxelwise FA distribution.
  2. Shuffling radius. This parameter defines the spatial neighborhood within which shuffling is permitted. We use a radius of 1 pixel.
  3. Mask ablation settings. The random-mask setting randomly masked 50% of spatial positions, whereas the patch-mask setting masked a single contiguous 20×2020\times20 region at a randomly selected location in the image.
  4. Perturbation levels and bounds. Spatial and angular bounds specify the permitted shuffling neighborhood and diffusion-direction range, respectively. Smaller neighborhoods and narrower angular ranges produce weaker perturbations.

We have added the FA formula and detailed experimental configurations to the appendix and specified the shuffling radius and FA percentile in the main text. We have standardized variable definitions throughout the manuscript, simplified subscripts and superscripts, and clarified the relationships among the different subsections to improve readability.

Q1: Thank you for this helpful suggestion. We have clarified that τFA\tau_{\mathrm{FA}} is empirically set to the 80th percentile of the estimated FA distribution across all voxels, and spatial shuffling is restricted to a 3×33\times3 neighborhood, corresponding to a spatial radius of 1. Following your suggestion, we have added the sensitivity test for the choosing of τFA\tau_{\mathrm{FA}}. Fig. 6(b) shows that lower percentile thresholds overestimate the protected regions (M=0M=0), whereas excessively high thresholds fragment these regions and impair denoising performance. The 80th percentile achieves the highest PSNR under both Rician and NC-Chi noise (34.26 and 26.31 dB, respectively), as well as the highest SSIM under Rician noise (0.960). Although the highest NC-Chi SSIM (0.952) is obtained at the 85th percentile, the 80th percentile provides a favorable balance between structural protection and denoising performance. These results support our choice of the 80th percentile as the default FA threshold.

Q2: We compared three tensor-fitting strategies: using all directions, excluding the target direction qq, and additionally excluding the replacement-source direction rpr_p at shuffled locations. Full results across four noise correlation levels are reported in revised Appendix A6, with representative results in Table R1. Excluding qq consistently improves PSNR over using all directions. Further excluding rpr_p yields additional gains of 0.93 dB, achieving the highest PSNR. SSIM also increases. These findings support the proposed direction-exclusion strategy for mitigating potential noise leakage during tensor fitting.

Table R1. Comparison of fitting strategies under Rician noise with ℓ=2.10\ell=2.10.

Fitting strategyPSNR (dB) ↑SSIM ↑
All directions29.57±0.4029.57\pm0.400.908±0.0060.908\pm0.006
Exclude qq30.09±0.5530.09\pm0.550.935±0.0040.935\pm0.004
Exclude qq and rbr_b31.02±0.4431.02\pm0.440.937±0.0030.937\pm0.003

Q3: Residual maps show the differences between noisy and denoised images. Recognizable anatomy in residuals may indicate unintended signal removal, whereas noise-like residuals without discernible anatomy qualitatively support structural preservation. Residual inspection alone cannot establish denoising quality. We added a reference-free SNR assessment for the real DW images in Fig. 6, using foreground/background intensity statistics and identical ROIs across methods within each dataset and b-value. Results are reported as mean ± standard deviation across 90 diffusion directions for HCP and 64 for Sherbrooke. Our method achieves the highest mean estimated SNR in five of six settings. These results provide complementary quantitative evidence of improved estimated SNR.

Q4: Our method aims to improve diffusion MRI quality under a fixed acquisition protocol. The additional real-data analysis shows improved estimated SNR, but does not establish clinical benefit; task-specific clinical validation remains necessary. In Fig. 4(b), all methods used the same number of seeds, tracking parameters, and filtering criteria. The apparent higher streamline density may reflect altered orientation estimates and trajectory propagation after denoising, but this explanation remains unverified. We have not quantified streamline counts, and visual density alone cannot establish an increase or distinguish recovered pathways from false positives. Our evaluation also cannot rule out denoising-induced spurious structures or connections. We therefore interpret Fig. 4(b) as a qualitative illustration of tractography changes, without claiming improved anatomical accuracy.

I thank the authors for providing additional experimental results and analysis regarding my concerns. A few follow-up suggestions:

  • In response to my Q3, I am assuming the SNR assessment was conducted on Fig. 5. For the SNR calculation, I didn't see the exact settings or definition, and it would be good to clarify how it was calculated (i.e., what was chosen to represent "signal" versus "noise", which I assumed was a foreground/background split).
  • As other reviewers suggested, some terminology could be confusing to an audience that doesn't have a general dMRI background or knowledge, and it should be clarified rigorously before being used in the introduction or methods.
  • I would highly suggest evaluations on downstream protocols to showcase the denoising capabilities are not purely visual.

Regardless, I think this work introduces a novel perspective on dMRI denoising and insights into how additional information in higher-dimensional imaging could benefit self-supervised blind spot networks. Therefore, I will keep my score.

AuthorsOpenReview ↗

Response to Reviewer tDDW

Thank you very much for your thoughtful follow-up suggestions and positive assessment of our work.

Regarding SNR evaluation, although we provided the calculation formula in our previous response, we acknowledge that the specific procedure for identifying foreground and background regions was not sufficiently explained. We will clarify these details in the manuscript and further refine the terminology to improve accessibility for readers without extensive dMRI expertise.

Your suggestion also inspired us to consider foreground/background segmentation as an additional downstream evaluation of denoising performance. Although the limited rebuttal period prevents us from completing this experiment, we plan to investigate it in future work. Regardless of the outcome, we believe this evaluation could provide valuable insights and help strengthen the paper.

We sincerely appreciate your constructive feedback, which has been very helpful in improving our work, and thank you again for recognizing its contributions.

Reviewer_m8npOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses the challenge of spatially correlated noise in self-supervised diffusion MRI denoising. The authors propose a framework combining fractional anisotropy (FA)-guided spatio-angular shuffling, direction-excluded diffusion tensor reconstruction, and cross-view consistency learning to balance noise decorrelation and structural preservation. The method achieves improved denoising performance over several baselines, including a 0.79 dB PSNR improvement under 10% Rician noise, with better preservation of diffusion-derived structures. The main contribution is a structurally adaptive self-supervision strategy supported by theoretical analysis and experimental evaluation.

Critical strengths
  1. The paper identifies an important trade-off between noise decorrelation and structural preservation in diffusion MRI. Using FA to guide perturbations is a reasonable and domain-specific design choice, and the direction-excluded tensor reconstruction provides complementary supervision. The combination is technically interesting and addresses limitations of existing self-supervised denoising approaches.

  2. Promising quantitative results and ablation studies. The method achieves the best overall performance on the simulated dataset, including improvements in PSNR, FA estimation, and fiber orientation accuracy. The ablation experiments also demonstrate the contributions of FA-guided masking and cross-view consistency. Evaluating diffusion-derived biomarkers in addition to image quality is particularly valuable for this application.

Critical weaknesses
  1. The experiments do not adequately validate the central spatial noise correlation problem. The paper is motivated by spatially correlated noise, but Appendix A.5 indicates that the simulated noise innovations are independent across voxels. The spatial dependence discussed in the appendix instead arises from signal-dependent magnitude bias, which is different from spatially correlated stochastic noise introduced by MRI acquisition or reconstruction. Therefore, the simulation experiments do not directly establish the effectiveness of the proposed approach under the noise conditions motivating the method. Furthermore, the evaluations on real datasets are primarily qualitative, making it difficult to determine whether the method provides meaningful improvements under realistic spatial noise correlation.

  2. The theoretical justification for bias reduction is not fully convincing. The bias reduction bound in Eq. 17 relies on the assumption in Eq. 14 that the auxiliary view has lower conditional noise bias, rather than establishing this property theoretically or empirically. Similarly, the benefit of consistency regularization depends on the relative biases of the two branches. There is also a gap between the analysis in Eq. 2, which considers prediction of the original target signal, and the actual objective in Eq. 13, which supervises predictions using the shuffled observation. These issues weaken the theoretical support for the proposed framework, even though the empirical results are promising.

Questions for authors
  1. Can the authors provide experiments with explicitly spatially correlated noise, varying the correlation strength or spatial extent? This would help establish whether the proposed method addresses the central problem more effectively than existing approaches.

  2. How is the FA guidance map estimated from noisy diffusion images? If the target diffusion direction contributes to FA estimation, could this introduce information leakage into the self-supervised learning process? An ablation examining the sensitivity to FA estimation errors would be helpful.

  3. Can the authors clarify how the supervision on the shuffled observation in Eq. 13 relates to the original clean-signal prediction objective in Eq. 2? Additionally, is there empirical evidence supporting the conditional bias assumptions underlying Eqs. 14 and 21?

1 comment
AuthorsOpenReview ↗

Response to Reviewer m8np

W1& Q1: We clarify that, from a blind-spot denoising perspective, signal-dependent Rician and nc-chi noise can induce spatial noise dependence even with voxel-wise independent Gaussian noise, as neighboring observations may provide information about the target noise, resulting in conditional noise bias. This mechanism was explained in our original supplementary material. Motivated by your suggestion, we further introduced explicit spatial covariance into the Gaussian components to simulate spatially correlated Rician and nc-chi noise. The procedure is detailed in revised Appendix A.3, with results in Fig. 4 (p.7) and the revised appendix (Fig.A3), respectively. These experiments further support the robustness of our method to spatially correlated noise. The Rician experiments cover four correlation levels, including the uncorrelated setting. Our method achieves the highest PSNR across all four settings. At the strongest correlation level (ℓ=2.10\ell=2.10), it outperforms all baselines, with a 1.30 dB PSNR advantage over the second-best method. These controlled experiments directly address the missing evaluation under stochastic spatial noise correlation.

We have also added quantitative SNR measurements for the real-data evaluations, annotated directly on the DW images in Fig.5(a) and Fig.5(b). The detailed table is given in Appendix (Table A2).

W2 & Q3 (1) Conditional bias reduction. The inequality in the original Eq. (14) was an assumption about the relative conditional noise biases. Direction-excluded tensor fitting removes the direct use of the target and selected source directions in constructing the auxiliary view, thereby reducing direct noise sharing with the shuffled target. This provides a motivation for a smaller auxiliary conditional noise bias, but does not establish the bias ordering universally. We therefore formulate the fusion result explicitly as a conditional statement. In the conditional-mean analysis, let Bblindη(p)=E[Npη∣Zpη],Bauxη(p)=E[Npη∣Ya]B_{\mathrm{blind}}^{\eta}(p)=\mathbb{E}[N_p^{\eta}\mid Z_p^{\eta}],\quad B_{\mathrm{aux}}^{\eta}(p)=\mathbb{E}[N_p^{\eta}\mid Y^{a}], When ∣Bauxη(p)∣≤∣Bblindη(p)∣\left|B_{\mathrm{aux}}^{\eta}(p)\right|\le\left|B_{\mathrm{blind}}^{\eta}(p)\right|, fusion satisfies ∣Bdualη(p)∣≤αpη∣Bblindη(p)∣+(1−αpη)∣Bauxη(p)∣≤∣Bblindη(p)∣.\begin{aligned}\left|B_{\mathrm{dual}}^{\eta}(p)\right|&\le\alpha_p^{\eta}\left|B_{\mathrm{blind}}^{\eta}(p)\right|+(1-\alpha_p^{\eta})\left|B_{\mathrm{aux}}^{\eta}(p)\right|\\&\le\left|B_{\mathrm{blind}}^{\eta}(p)\right|.\end{aligned} Under this condition, the fused noise-bias magnitude does not exceed the blind-spot counterpart; strict reduction follows when the inequality is strict and αpη<1\alpha_p^{\eta}<1. To evaluate this condition, we estimated the conditional means of the same shuffled target noise from the blind-spot context and auxiliary view. The scatter plots show smaller auxiliary bias magnitudes for 84.1% of evaluated voxels under Rician noise (the image in the second panel of Fig.8(b)). These results empirically support the condition for most evaluated voxels. The noise estimation bias results and denoising ablations in Fig.7 provide complementary evidence of the aggregate benefit.

(2) Consistency regularization. We agree that its benefit depends on the relative biases of the two predictions. The revised quadratic analysis in Eqs. (16)-(18) makes this dependence explicit: consistency produces a weighted combination of their errors and can improve the estimate when the complementary error is smaller or the errors partially cancel. We present this as a conditional explanation, supported by the ablation results (Fig 5.(a)), rather than an unconditional guarantee, therefore original Eq.(21), now Eq.(18) is empirically validated.

(3) Alignment with the training target. We revised the original Eq. (2), now Eq. (3), to explicitly account for shuffled supervision: f∗(Zi,qη)−Si,q=E[Ni,qη∣Zi,qη]+(E[Si,q+Δi,q∣Zi,qη]−Si,q).f^{*}(Z_{i,q}^{\eta})-S_{i,q}=\mathbb{E}[N_{i,q}^{\eta}\mid Z_{i,q}^{\eta}]+\left(\mathbb{E}[S_{i,q}+\Delta_{i,q}\mid Z_{i,q}^{\eta}]-S_{i,q}\right). This decomposition separates the noise-related contribution from the clean-signal prediction error, including shuffle-induced structural mismatch. Section 3.2 limits this mismatch through FA-guided shuffling, while Section 3.3 investigates cross-view guidance and consistency. The revised analysis thus aligns with the shuffled supervision target in Eq. (15) (formerly Eq. (13)). We also simplified the notation for clarity.

Q2: FA is estimated from the noisy diffusion images using all available directions. Its role is limited to defining a binary mask for selecting regions where shuffling is applied; FA values are not used as supervision targets. To assess sensitivity to FA estimation errors, we compared the masks under Rician and nc-chi noise at 6%, 8%, and 10% (Fig.6(a)). Percentile-based selection preserves the overall spatial patterns of the masks across these settings, with some local variations at higher noise levels, supporting the stability of the region-selection procedure.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.