Abstract

Optical coherence tomography (OCT) and its derived correlation stability (CS) maps provide high-resolution structural imaging and elastography information at the micrometer scale, which are critical for the rapid and accurate detection of malignant biliary strictures. However, acquiring intraoperative CS maps remains challenging due to the need for complex mechanical compression and the high costs associated with hardware modifications in conventional OCT systems. Moreover, existing deep learning methods for OCT-to-CS map translation often lack long-range modeling capability for recovering subtle deformation features and are limited in modeling the relative importance of different frequency components, which is essential for localizing high-frequency stiffness details. To address these limitations, we propose FA-Diff, a novel RWKV-based frequency-aware diffusion framework for elastography mapping. FA-Diff integrates a frequency-aware multi-granularity shift (FG-Shift) module and a frequency-adaptive sub-band normalization (FASN) mechanism to enhance elasticity-aware representation learning. Specifically, FG-Shift is incorporated into the RWKV architecture as a spatial-frequency adaptive scanning strategy to improve the modeling of vertical stiffness textures and enable effective global coarse-to-fine contextual aggregation. Meanwhile, FASN dynamically adjusts normalization parameters using inter-stage states to adaptively modulate wavelet sub-band coefficients. Comprehensive experiments on the clinical cholangiocarcinoma dataset demonstrate that FA-Diff consistently outperforms state-of-the-art CNN-, Transformer-, Mamba-, and RWKV-based baselines, achieving superior reconstruction accuracy and perceptual quality.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0521_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaChu_FADiff_MICCAI2026,
        author = { Zhang, Chuanhao AND Li, Yangxi AND Song, Jianping AND Wang, Yingni AND Han, Haojie AND Xiang, Canhong AND Ning, Guochen AND Chen, Fang AND Liao, Hongen},
        title = { { FA-Diff: Leveraging Frequency-Aware RWKV Diffusion for OCT to Correlation-Stability Map Translation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose FA-Diff, a frequency-aware RWKV-based diffusion framework specifically designed for translating OCT B-scans into correlation-stability (CS) maps. It introduces a frequency-aware multi-granularity shift (FG-Shift) module and a frequency-adaptive sub-band normalization (FASN) mechanism. On a clinical cholangiocarcinoma dataset, the method achieves a 22% LPIPS improvement and 24% complexity reduction.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The approach is novel and domain-aware: FG-Shift mimics physical tissue compression patterns, while FASN preserves fine-grained elasticity textures. Validation on over 10,000 paired images compares favorably against CNN, Transformer, Mamba, and RWKV baselines, with statistically significant 5-fold cross-validation results.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    First, evaluation relies solely on generic image quality metrics (PSNR, SSIM, LPIPS, MS-SSIM), lacking downstream clinical validation (e. g. , diagnostic accuracy) or expert reader studies to confirm physical fidelity. Second, the dataset originates from only 23 patients, raising concerns about diversity and generalization. Third, while FLOPs and T=4 sampling steps are reported, actual inference time or frame rate is omitted, leaving real-time intraoperative feasibility unproven.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The technical novelty and solid methodology place the paper marginally above the acceptance threshold. However, the lack of clinical task evaluation, limited patient cohort, and missing latency metrics prevent a higher score. In rebuttal, the authors should clarify inference time, discuss generalization limitations, and acknowledge the absence of clinical validation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I’m satisficed with the rebuttle.



Review #2

  • Please describe the contribution of the paper

    The authors propose FA-Diff, a frequency-aware RWKV-based diffusion framework for the translation of optical coherence tomography (OCT) images to derived correlation stability (CS) maps. They evaluated FA-Diff on an in-house dataset of paired OCT B-scans and CS maps from 23 patients, showing good performance in metrics (PSNR, SSIM, LPIPS, and MS-SSIM) compared to benchmark methods.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The major components of the proposed method FA-Diff are (1) Wavelet-Guided Stiffness Interaction (WSI) and (2) Frequency-Adaptive Sub-band Normalization (FASN). The major subcomponent of the WSI block is the Frequency-aware multi-granularity shift (FG-Shift).

    The purpose of the WSI module was to fuse the low- and high-frequency information The purpose of the FG-Shift subcomponent of the WSI module was to (1) mimic the tissue compression patterns and (2) strengthen modelling of elastic geometric structures. The purpose of the FASN module was to enable adaptive modulation and fusion of wavelet sub-bands.


    SSIM and LPIPS are presented as the quantitative metrics for structural contrast of predicted CS maps. MS-SSIM and LPIPS are presented as the quantitative metrics for structural consistency of predicted CS maps.

    The authors benchmarked their method against a comprehensive list of methods that cut across architecture types such as Swin Transformer [20], Mamba [16], a very similar RWKV architecture [8], and diffusion modelling [24].

    The presented results show that the proposed FA-Diff better preserves structural contrast and consistency (highlighted by improved SSIM, MS-SSIM, and LPIPS metrics) in CS maps it generates from OCT-B scans, compared to the CS maps generated by benchmark methods.


    Authors performed a good ablation study to show that removing either the FASN or FG-Shift modules resulted in consistent performance degradation across all quantitative performance metrics, which indicates the contributions of the modules.

    Fig. 3 is a good qualitative visualization of the ability of FA-Diff to preserve fine structural details and stiffness representation in the predicted CS Map versus benchmark methods.

    Table 3 is a good ablation study presentation of the impact of the proposed scanning scheme (FG-Shift) versus the FSO-Shift in [8] and the Zigzag in [1]

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The Frequency-Aware RWKV Diffusion (FA-Diff) method proposed in this paper is very similar to the Frequency Spatial-Aware RWKV (FS-RWKV) method in [8].

    The major method components introduced in the proposed FA-Diff, which include: (1) Wavelet-Guided Stiffness Interaction (WSI) (2) Frequency-Adaptive Sub-band Normalization (FASN) (3) the Frequency-aware multi-granularity shift (FG-Shift)

    are very similar to their counterpart components in FS-RWKV [8], which include:

    (1) Frequency Spatial-RWKV (FS-RWKV) (2) Structural Fidelity Enhancement Block (SFEB) (3) Frequency-Spatial Omnidirectional Shift (FSO-Shift)

    The major difference seems to be in the clinical application. The proposed FA-Diff is applied for OCT-B to CS map translation while FS-RWKV [8] was applied for 3T-to-7T MRI translation.


    The authors should have included Contrast-to-Noise Ratio (CNR) as the metric to quantify the structural and texture contrast in the CS maps from their method versus benchmark methods. The CNR metric seems more physically meaningful in the context of CS maps than the more generic PSNR, SSIM, MS-SSIM, and LPIPS metrics used in this paper.

    Minor issue: On page 3, authors should define the DDPM acronym.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper is the first exploration of RWKV architectures for CS map translation from OCT-B scans, showing good translation performance versus benchmark methods. However, from a method perspective, it shares several similarities with the RWKV approach in [8] with the major difference existing in the clinical application.

    The authors also miss out on the utilization of the CNR metric, which is more physically meaningful in the context of OCT and CS maps than the more generic metrics (PSNR, SSIM, LPIPS, and MS-SSIM) presented in this paper. CNR is likely to be a more effective metric to showcase the effectiveness of the proposed method versus benchmark methods for the task of OCT-B to CS map translation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors sufficiently address the concerns of the reviewers. They provided good arguments to justify how the task-specific design of FA-Diff differentiates it from the FS-RWKV model.

    Furthermore, in line with the MICCAI reviewer guidelines which allows for the presentation of additional metrics, during rebuttal, on same datasets/experiments already presented in the paper, the authors show that FA-Diff achieves a superior CNR score compared to benchmark models.



Review #3

  • Please describe the contribution of the paper

    The main contribution is the construction of a frequency-aware RWKV-based diffusion framework for OCT-to-CS map translation.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    relatively high estimation performance of a CS map.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The explanation of the research background is limited, making it difficult to determine how useful this method will be in practical applications.

  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The estimation performance is relatively high, but I hardly understand the method’s utility. Is CS the only one capable of solving OCT’s challenges? I think the explanation is needed.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all reviewers for their constructive comments. We are encouraged that the reviewers recognized the novelty and efficiency of FA-Diff, including FG-Shift and FASN for mimicking compression patterns, WSI for preserving elasticity textures, comprehensive comparisons with CNN/Transformer/Mamba/RWKV/diffusion baselines, and consistent ablation evidence. We address the remaining concerns below. 1.Clinical utility of CS maps (R3) (1)Why CS matters: OCT B-scans provide micrometer-scale morphology, but morphology alone may miss stiffness-related malignant changes. CS maps are one of the elastography imaging method providing quantitative elasticity contrast. We do not claim CS is the only solution; CS maps provide complementary biomechanical contrast to structural OCT, with prior biomedical evidence in elasticity assessment and lesion delineation [1,2]. (2)Why virtual CS generation matters: Conventional compression-based OCT elastography requires localized mechanical compression or costly hardware, which is difficult for intraoperative use. FA-Diff estimates CS maps directly from structural OCT, avoiding extra hardware modification at inference while preserving biomechanical information. This virtual conversion scheme is valuable for hepatobiliary and ophthalmic surgery, where workflow simplicity and non-contact imaging are important. We will clarify this motivation in the introduction section. [1] Zaitsev VY, et al. Proc. SPIE 9129 (2014) [2] Gubarkova EV, et al. Biomed. Opt. Express 13.5, 2859 (2022). 2.Difference from FS-RWKV (R2) While FA-Diff shares the general concept of frequency RWKV modeling with FS-RWKV, it is not an incremental adaptation. 1.FA-Diff is a diffusion model tailored for OCT-to-CS translation. 2.FG-Shift with frequency-adaptive sub-band normalization is specifically designed for OCT elastography, leveraging wavelet-guided multi-granularity window partitioning and sparse selective interaction to emphasize compression-related stiffness patterns, whereas FS-RWKV performs general frequency-spatial shifting without task-specific design (FG-Shift vs FSO-Shift) or global inter-stage modulation (FASN vs SFEB). These designs improve SSIM/MS-SSIM/PSNR/LPIPS over FS-RWKV while reducing FLOPs and parameters. 3.Evaluation with CNR and inference time (R1, R2): We agree that CNR is physically meaningful for CS maps. Using expert-defined ROI, the mean CNR values are: BCI: 1.53| Li: 1.31| Yao: 1.82| Ho: 1.47| WFI2-Net: 1.76| ResShift: 1.73| WalMaFa: 1.81| Yang: 1.79| FS-RWKV: 1.80| FA-Diff: 1.86|. Thus, beyond superior PSNR, SSIM, MS-SSIM, and LPIPS, covering fine semantic to structural representation and signal-to-noise ratio, FA-Diff also achieves the best CNR. The inference time of FA-Diff is 216 ms per frame, enabling potential real-time intraoperative application. FA-Diff provides the best trade-off between accuracy, complexity and runtime. 4.Generalization (R1, R2) All of experiments were conducted using patient-level 5-fold cross-validation on 10834 paired images, one of the largest datasets for OCE synthesis to date. While the cohort size is limited, FA-Diff consistently demonstrates robustness across folds. The revision will characterize multi-site/device validation and further classification evaluation, such as disease classification using FA-Diff-generated CS maps, as future work, to avoid overclaiming clinical readiness. Minor correction: We will define DDPM as denoising diffusion probabilistic model at its first occurrence.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Please provide justifications and clarifications for the concerns raised by the reviewers.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper proposes FA-Diff, a frequency-aware RWKV-based diffusion framework for translating OCT B-scans into correlation-stability maps for intraoperative elastography.

    Reviewers broadly agree that the application is clinically meaningful, generating CS maps directly from structural OCT without additional hardware is a practically valuable direction, and that the ablation study convincingly supports both the FG-Shift and FASN components. The comprehensive baseline comparison across CNN, Transformer, Mamba, and RWKV architectures is also appreciated.

    The shared concerns are similarity to FS-RWKV, missing clinically meaningful metrics, and limited dataset size. On similarity to FS-RWKV, the rebuttal argues that FA-Diff differs in being a diffusion model with task-specific designs for OCT elastography, and provides CNR results showing FA-Diff outperforms FS-RWKV, which R2 finds convincing. On missing metrics, the rebuttal provides CNR values (FA-Diff: 1.86, best among all methods) and inference time (216 ms per frame), resolving both concerns. The limited patient cohort (23 patients, 10,834 paired images with 5-fold cross-validation) remains a generalization concern, though the authors appropriately defer multi-site validation to future work rather than overclaiming. R3’s concern about insufficient background explanation of CS map utility is addressed in the rebuttal.

    The core contribution is sound and the rebuttal meaningfully strengthens the evidence. I recommend acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper presents solid work with novel methods for OCT-to-CS map translation and is therefore recommended for acceptance. Reviewers 1 and 2, who initially rated weak accept, have updated their assessments to “Accept,” expressing satisfaction with the authors’ rebuttal, while Reviewer 3 rated weak reject but is not certain about the review.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper received two positive and one negative reviews. The main concern from the negative review is the main background and usecase scenario of the CS map instead of methodological contribution. The rebuttal has provided a detailed explanation, which sounds reasonable to me. Therefore, I believe that the paper can be accepted.



back to top