Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Deep learning has been introduced into optical coherence tomography angiography (OCTA) imaging, which is a non-invasive technique for visualizing vascular structures. Intralipid injection has shown promise in improving blood cell scattering for better OCTA imaging. However, administering intralipid to human subjects for imaging purposes may raise ethical concerns. To address this challenge, we acquire intralipid-enhanced OCTA in rats and introduce cross-domain learning to address the domain shifts. Specifically, we collect data from eyes of anesthetized rats to obtain motion-free data and introduce a noise-guided self-training framework to bridge the domain gaps between rats and primates. Additionally, an en face enhancement loss is incorporated to further refine en face vectors during adaptation. Compared with other classical and fully supervised OCTA imaging algorithms, our method improves B-scan denoising performance by 53.1% and 65.0% on CNR and BRISQUE in human subjects respectively, while enhancing vessel contrast in en face images.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3594_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{YanBin_Improving_MICCAI2025,
        author = { Yang, Bingyu and Tan, Bingyao and Gu, Zaiwang and Schmetterer, Leopold and Li, Huiqi and Cheng, Jun},
        title = { { Improving OCTA Imaging through Cross-Domain Adaptation: A Noise-Guided Framework Using Intralipid-Enhanced Rat Data } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15966},
        month = {September},
        page = {284 -- 294}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper tackles the challenge of reconstructing high-quality OCTA images from OCT scans with a reduced number of repeated acquisitions. Traditional OCTA requires multiple OCT readings (or repeats) at each slow-axis position, which is time-consuming and sensitive to motion. To reduce the number of repeats, intralipid injections serve as as signal enhancement for animals, but is not readily viable for human use. To address this, the authors propose a cross-domain self-training framework. A model is first trained on high-quality OCTA data from rat retinas enhanced via intralipid, and then adapted to primates (humans and monkeys) using unsupervised domain adaptation with a teacher-student approach. This enables improved OCTA reconstruction from limited-scan OCT data. Experimental results show superior image quality compared to baseline 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.

    S1. [Original use of data] The use of a high-quality OCTA dataset from rat retinas as a source domain, followed by unsupervised domain adaptation to human and monkey data, is a well-motivated and relevant strategy. In contrast to standard transfer learning or fine-tuning approaches, which would be limited by the lack of reliable ground-truth in human data, the use of a self-supervised teacher-student framework provides a more suitable solution. This approach effectively addresses the domain gap without relying on noisy or impractical ground-truth.

    S2. [Well-designed method] The proposed method is clearly motivated and carefully designed. It avoids unnecessary complexity, and each component of the pipeline is justified within the context of the problem. Figure 1a provides a clear and intuitive diagram of the teacher-student architecture, helping to convey the structure of the method. The use of noise injection in the student branch, the role of en-face projections in the loss function, and the design of the teacher-student framework are all well explained. The inclusion of an ablation study further supports the rationale behind the design choices.

  • 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.

    W1. [Inconclusive metrics] The experimental results are inconclusive due to the lack of a reliable reference ground-truth and the metrics used. In the absence of ground-truth OCTA images for the target domain, the authors rely solely on no-reference image quality metrics such as contrast-to-noise ratio (CNR) and a custom vessel contrast (VC) score. While these metrics may reflect certain aspects of visual quality (e.g., contrast and sharpness), they are insensitive to the anatomical fidelity of the reconstructed images. For example, a simple image processing algorithm (such as increasing contrast or applying Gaussian smoothing) could improve these metrics without preserving true vascular structures. Moreover, a model could hallucinate vessels or omit important features while still scoring highly on these metrics. As such, these measures are insufficient on their own to support the claimed effectiveness of the proposed method.

    Previous work, such as [11], has used high-repetition OCT scans (e.g., 48 repeated B-scans per location) to construct high-quality OCTA volumes. It is unclear to me why a similar approach could not have been used to acquire reference data for at least a subset of the target domain (e.g., monkeys). This would have enabled a more meaningful evaluation of anatomical accuracy and provided a way to assess whether the domain adaptation process introduces dangerous hallucinations or misleading features.

    W2. Minor weaknesses worth addressing.

    W2.1. Including a figure that shows examples of OCT scans and corresponding OCTA images from the source domain would be very helpful. This would illustrate the domain gap and provide a visual reference for what high-quality OCTA images should look like compared to the obtained target-domain reconstructions.

    W2.2. The paper states that 14,308 OCTA volumes were used from the source domain, but does not specify how many individual animals these volumes came from. Similarly, information about the number of subjects in the target domain is not provided. Including these details would help assess dataset diversity and reduce ambiguity.

    W2.3. The method penalizes the differences between ê and z using the Pearson correlation in Equation (5). It is unclear to me why this measure was preferred over more common alternatives like the L2 norm. A brief justification for this choice would improve transparency and interpretability.

    W2.4. The paper treats monkey and human retinas as part of the same target domain, but does not provide any justification or visual comparison to support this assumption. While it is reasonable to assume some anatomical similarity, the fact that evaluation metrics differ noticeably between monkey and human subjects suggests there may be non-negligible domain differences. A visual or statistical comparison of the two would help clarify whether they can be reliably grouped together for training and evaluation.

    W2.5. In Section 2.2, the authors state that the teacher and student networks “are trained with the well-trained weights.” This phrasing is unclear and somewhat awkward. Presumably, this refers to the weights trained in the source domain (i.e., on the rat data), but the terminology should be clarified to avoid confusion.

    W2.6. It is not clearly stated how the monkey and human data are used during the self-training phase. I assume that the 1500 monkey and 2025 human training images are jointly used as a single dataset for domain adaptation, but this is never explicitly explained. Providing a clear breakdown of how each part of the dataset is used at each stage of the method would improve transparency and reproducibility.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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.

  • 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/2025/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 — could be rejected, dependent on rebuttal

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

    I am on the fence with this paper. On the one hand, the problem is relevant and the proposed solution is elegant and well-motivated. The use of a high-quality animal dataset combined with a self-supervised domain adaptation approach is an interesting and thoughtful strategy for OCTA imaging.

    However, the experimental validation is a significant concern. The paper relies solely on no-reference image quality metrics, which, while helpful for assessing visual clarity, provide little insight into anatomical fidelity or clinical utility. Higher scores on these metrics do not necessarily imply more realistic or trustworthy reconstructions. In a medical context, this is critical: improved appearance does not equate to improved diagnostic value. Relying solely on such metrics and visual inspection makes it difficult to assess the quality of the method and its actual contribution.

    I believe that a more rigorous evaluation or stronger justification is needed before the contribution can be clearly established. At this stage, I would value clarification of the concerns raised in W1. Clarifying the points listed under W2 would also be helpful and strengthen the manuscript, though they are of lower priority. The proposed direction is promising, and I encourage the authors to further strengthen the evaluation to fully demonstrate its value.

  • 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.

    I thank the authors for their thoughtful rebuttal, which partially addressed my primary concern regarding the reliance on non-reference image quality metrics. In particular, the authors conducted an additional experiment using a simplified 2D version of their method, applied to 2D scans reconstructed from 1000 repetitions (serving as ground truth) and 3 repetitions as input. In this setting, they demonstrated noticeable improvements over OMAG using standard reference-based metrics PSNR and SSIM.

    While this experiment is limited to a single target domain and reports only quantitative (not qualitative) results, I appreciate the effort and am willing to give the benefit of the doubt. I strongly encourage the authors to include this experiment, or a refined version of it, in the final paper to enhance its robustness and impact.

    Overall, I find the proposed method interesting and relevant, and I believe it will be of value to the MICCAI community.



Review #2

  • Please describe the contribution of the paper

    This study proposes a cross-domain deep learning approach for denoising OCTA images. The authors used a rat model with intralipid injection and anesthesia to stabilize the rats and improve image quality, which served as the source domain. A student–teacher framework was employed for semi-supervised learning, where different noise levels were used to generate pseudo-ground truth in the target domain. The target domain consisted of monkey and human OCTA images. The proposed method aimed to improve image quality in both B-scans and en-face images. The image dimensions differed across the datasets, and high-quality OCTA images (using intralipid injection) were only available in the source domain.

  • 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.
    1. Cross-domain learning is a meaningful topic in medical imaging, especially when high-quality ground truth data in human studies is limited or difficult to obtain.
    2. OCTA denoising is a relevant and clinically important problem in ophthalmic imaging.
  • 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.
    1. Although the authors used multiple image-based quality metrics (i.e., CNR, BRI, ENT, and VC), there was no definitive or “true” ground truth of the OCTA images for the microvasculature in this study.
    2. Minor formatting issues are present in the references (e.g., oct, octa, and 3d), though these do not affect the core contribution.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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 provide sufficient information for reproducibility.

  • 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/2025/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 — could be accepted, dependent on rebuttal

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

    I found the idea of cross-domain learning particularly interesting, as it addressed a key challenge in medical imaging and machine learning. In this study, it was not practical to use intralipid injection to improve OCTA image quality in human subjects, so the authors used animal models to support model training and then adapted the model to monkey and human images. The use of animal data to guide learning was well-justified in this context. The proposed semi-supervised framework, based on student–teacher models trained with different noise levels, was also reasonable. However, the evaluation mainly focused on improvements in image quality, without validating against the true microvasculature structure in OCTA. This limited the clinical relevance of the results. For future work, the authors could consider comparing their method with OCTA images acquired from advanced SS-OCT systems, which provide higher quality and could serve as a more reliable reference. Another possible direction would be to test the proposed method on repeated OCTA scans to assess the repeatability of enhanced microvasculature. It would also be valuable to evaluate the method in diseased cases, such as those with microvasculature dropout or significant vascular remodeling.

  • 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 am satisfied with the authors’ responses to the reviewers’ comments.



Review #3

  • Please describe the contribution of the paper

    The authors have offered an intralipid-enhanced OCTA in rats and introduce cross-domain learning to address the domain shifts to human subjects. Authors have validated their method by comparing with other classical fully supervised OCTA imaging algorithms to show improvement in B-scan denoising performance while enhancing vessel contrast.

  • 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.
    • This article introduces a type of innovation in the form of knowledge transfer from one model to another and uses it for a suitable use case.
    • Using domain adaptation when target domain has high noise intensity.
  • 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 paper should explain why it is ethically infeasible to apply to human subjects.
    • Using no-reference evaluation metrics.
    • The experimental results should be represented more effectively, particularly with respect to some quantitative data.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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 provide sufficient information for reproducibility.

  • 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/2025/en/REVIEWER-GUIDELINES.html
    • The concept of comprising a student network and a teacher network sharing the same architecture as the source domain, is not explained clearly.
  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    The use of transfer learning from a network for rat to a network for humans is somewhat novel and worth publishing.

  • Reviewer confidence

    Somewhat confident (2)

  • [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

  1. No-reference evaluation metrics (R1 R2 R3) We thank the reviewers for the valuable comments. We fully agree with the reviewers that full-reference metrics compared with golden ground truth is important. An ideal ground truth for B-scans should present clear vessels in the ROI without noise in non-ROI regions, with en face displaying clear vascular structures. Fluorescein angiography imaging is often used by clinicians for visual comparison. However, it is difficult to obtain well aligned data for quantitative measurement. Averaging from highly overlapping scans (e.g., 48 frames per position) is an alternative approach. But this requires hardware modifications in the machine, while the existing 3D OCTA machine (including ours) can only sample a few times (e.g., 4) at the same position for human in 3D, which may produce poor-quality ground truth. Meanwhile, we have another machine that can conduct 1000 overlapping scans in 2D mode for monkey. We use first 3 scans as input and synthesize the ground truth using averaging from 1000 scans for 2D imaging. Our method, w/o the enface enhancement loss (not applicable for 2D), applied to 234 different locations (26 different monkey eyes) achieved a PSNR of 31.6933 and SSIM of 0.7134, surpassing traditional OMAG, which achieved a PSNR of 24.5673 and an SSIM of 0.4759. However, as this machine cannot scan in 3D to generate en face OCTA images, we did not include these results in our original submission. As non-blind metrics often show consistent results with full-reference metrics, many researchers often use non-blind metrics, e.g., Marcel et al. compute enhancement factor similar to CNR in “Investigation of artifacts in retinal and choroidal OCT angiography with a contrast agent”, Biomed. Opt. Express 9, 1020-1040 (2018).

  2. Intralipid injection for human (R1) Current experiments with intralipid injection for OCTA have primarily been conducted in animals only. It is not used for human eyes due to safety concerns as it may interact with ocular tissues unpredictably. We propose to apply that on rats and transfer the knowledge to human to avoid potential risks.

  3. Effectively showing quantitative results (R1) We will include explanations for full-reference metrics in the final version.

  4. Network structure (R1) Using the same network architecture in the target domain is to transfer the weights from the source domain.

  5. Formatting issue (R2) We will revise the formatting issues (e.g., oct, octa, and 3d) in the references.

  6. Source domain figure (R3) Thank you for your constructive questions. We will include a comparison of B-scan images from different domains in the final submission.

  7. Number of experiment subjects (R3) Table 1 describes OCTA B-scan images, not volumes. The source domain includes data from 28 different rat eyes. The monkey data were collected from two different eyes with one for adaptation and one for test. The human data include 9 different subjects for adaptation and one for testing. We will provide a more detailed description of these data in the final version.

  8. Pearson correlation and L2 loss (R3) Pearson correlation assesses the linear relationship between two variables, regardless of their absolute scale. L2 measures the absolute difference between two vectors, making it sensitive to magnitude. In general, OCTA B-scan images with higher noise intensity tend to have higher mean values in their lateral en face vectors. As denoising reduces noise in non-ROI regions, the mean of lateral en face vectors decreases. To avoid the effect of magnitude shift, we use Pearson correlation as the loss function.

  9. Monkey and human adaptation (R3) Monkey and human data were adapted separately. We will add more detailed descriptions of the experiment settings.

  10. Other minor issues (R3) We will revise the sentence “They are initialized with the well-trained weights” to “They are initialized with the weights trained in the source domain (i.e., intralipid rats)”.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top