Abstract

Visual field (VF) sensitivity assessment is critical for quantifying functional impairment in high myopia (HM), yet conventional perimetry is subjective, time-consuming, and affected by substantial testretest variability. Deep learning-based estimation of VF sensitivity from fundus photography has emerged as a promising alternative; however, most existing approaches directly regress all 52 pointwise sensitivities and do not explicitly consider latent defect patterns. In this work, we propose ProtoVF, a prototype-based framework for fundus-to-VF estimation in HM. ProtoVF explores the underlying low-dimensional prototypes of VF patterns and predicts their combination coefficients from fundus images to estimate the full 52-point VF sensitivity map. By reformulating VF estimation as a prototype-learning problem, ProtoVF imposes an explicit prior over plausible VF patterns and decomposes predictions into clinically interpretable prototype patterns. In five-fold cross-validation on a large real-world HM cohort, ProtoVF achieved 2.55 MAE and 4.09 RMSE, outperforming the regression baseline, particularly in eyes with severe VF loss (MD ≤ −6 dB). These findings demonstrate that this prototype-based learning enables a more accurate, robust, and interpretable solution for fundus photography-based VF estimation in HM. Our code is available at https://github.com/OptometryAI/ProtoVF.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1688_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/OptometryAI/ProtoVF

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiGuo_Prototype_MICCAI2026,
        author = { Li, Guoliang AND Yan, Zipei AND Liu, Zhengji AND Wang, Shuai AND Chun, Rachel Ka-man AND Zhou, Yongjin AND Kee, Chea-su AND Liang, Dong},
        title = { { Prototype Learning for Visual Field Estimation from Fundus Photography in High Myopia } },
        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

    This manuscript proposes ProtoVF, a prototype learning framework for estimating 52-point visual field sensitivity from fundus photography in high myopia. The core idea is to replace direct pointwise regression with prediction of coefficients over a compact set of learned visual field prototypes, with the goal of improving robustness and interpretability in a limited and noisy-label setting.

  • 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 paper addresses a clinically meaningful problem. Functional assessment in high myopia is important, and the motivation for reducing dependence on subjective perimetry is well stated. The prototype-based formulation is conceptually interesting. Recasting VF estimation as coefficient prediction over a lower-dimensional prototype space is a reasonable way to impose structure and may improve interpretability. The manuscript includes both quantitative and qualitative evaluation, as well as an ablation over prototype generation methods and prototype number. That is helpful for understanding the method’s internal 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.

    1-The novelty is somewhat incremental relative to prior fundus-to-VF estimation work in high myopia. 2-The experimental comparison is limited. The main baseline is a generic regression model, and stronger recent high-myopia-specific methods are cited but not clearly compared in a fair, detailed way. 3- (important) I couldn’t understand how the prototypes and coefficients from the fundus images are computed. Providing pseudocode for both the training and testing phases would help clarify this. 4-Clinical validation is modest. The study is single-center, from a relatively homogeneous population, and restricted to non-glaucomatous HM (i think), which narrows the clinical significance and generalizability of the findings.

  • 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

    I coudn’t figure out how the proptptypes and coeffeficnets from fundus are computed. Having a pseudo code for training steps and testing phase is needed.

  • 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 clinically relevant, generally well organized, and contains an interesting, structured prediction idea.

  • Reviewer confidence

    Very confident (4)

  • [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



Review #2

  • Please describe the contribution of the paper

    This paper proposes a new method for predicting visual field sensitivity based on fundus photographs. Unlike traditional approaches that directly predict a 52-dimensional vector from fundus photographs, this method treats the visual field sensitivity to be predicted as a linear combination of different visual field patterns. First, a prototype matrix basis is extracted from all visual field reports using algorithms such as PCA. Subsequently, deep learning methods are employed to learn the weights of the different prototype matrices, thereby predicting the sensitivity map.

  • 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 advantage of this approach lies in transforming the parameter learning of 52 distinct point locations into the learning of lower-dimensional and more interpretable weights associated with classical visual field sensitivity distribution patterns.

    A particularly interesting aspect of the paper is the detailed ablation study regarding the number prototypes for visual field reports. It clearly and intuitively demonstrates the relationship between the number of prototypes and the Mean Squared Error (MSE) across different methods for generating prototypes.

  • 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.

    a. In the visualization of prototypes, this paper utilizes AA instead of PCA, which demonstrates superior comprehensive performance. This weakens the credibility of the conclusions.

    b. Neural network learning is also a crucial component. The paper lacks a theoretical justification for the network selection. It is recommended to supplement with ablation studies to elaborate on this choice.

    c. The conclusions related to visualization should not be placed in Section 3.4 ablation study.

    d. The ablation study section in Figure 3 shows that for most methods (e.g., PCA), the model accuracy improves as the number of prototypes increases. A sufficient justification is not provided for the selection of k = 10 (the paper mentions this configuration provided a favorable balance between representational capacity and generalization. But this only applies to the AA method ).

  • 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?

    In summary, this paper proposes a new method for predicting visual field sensitivity based on fundus photographs, which possesses a certain degree of interpretability and holds innovative significance in the medical field. However, there is still room for refinement regarding the algorithm details and experimental setup.

  • 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.

    Four specific concerns were raised in Q7.After careful evaluation of the authors’ rebuttal, I conclude that none of them undermines the core contribution or validity of the paper.

    Concern a (visualization used AA instead of the best‑performing PCA). The authors clarified that visualization serves clinical pattern interpretation, not model selection. AA produces archetypes at the convex hull of the VF distribution, which resemble real, interpretable defect patterns (e.g., blind‑spot enlargement, arcuate defects). PCA components are orthogonal variance directions and are less naturally read as standalone VF maps. This explanation is fully reasonable; the paper already reports PCA as the best performer in Table 1.No credibility issue remains.

    Concern b (lack of justification for network choice / ablation). While the paper does not provide an ablation on the backbone, the choice of ResNet‑34 is standard and widely adopted.

    Concern c (visualization placed in Sec. 3.4 Ablation Study). This is a purely structural/editorial issue. It does not alter any scientific claim or result. The content itself is valuable and can be moved to a more appropriate section (e.g., Results or a separate interpretation subsection).

    Concern d (k=10 chosen without sufficient justification, seemingly only for AA). The authors correctly pointed out that Fig. 3 shows VF reconstruction error, not final fundus‑to‑VF estimation error. Increasing k reduces reconstruction error for PCA/NMF/SVD but also makes coefficient learning harder and may overfit label noise. Selecting k=10 as a compact operating point balances representational capacity and learnability. This justification applies to all methods, not only AA. The concern is resolved.



Review #3

  • Please describe the contribution of the paper

    The authors proposed a deep learning model that uses a ResNet-34 to encode color fundus images into coefficient representations of visual field (VF) prototypes. These coefficients are then used to predict all 52 test locations in the 24-2 Humphrey visual field for patients with high myopia. To construct the VF prototypes, the study explores several dimensionality reduction techniques, including archetypal analysis (AA), principal component analysis (PCA), nonnegative matrix factorization (NMF), and truncated singular value decomposition (Truncated SVD). The central motivation of this work is to leverage the relative stability and objectivity of color fundus imaging to predict visual field measurements, which are inherently subjective. This study aims to establish a relationship between structural features observed in fundus images and functional visual outcomes.

  • 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 authors clearly stated that the training and test datasets are separated at the subject level during cross-validation. This is an important methodological practice in machine learning, as it helps prevent information leakage caused by strong correlations between paired eyes from the same individual.

  • 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.

    Predicting HVF outcomes from color fundus photographs in highly myopic eyes is inherently challenging. Unlike glaucoma, high myopia often lacks clearly localized or easily interpretable retinal nerve fiber bundle defects in fundus images that can be directly associated with VF loss. In this context, OCT may provide more direct and informative structural indicators of optic nerve damage to support structure–function modeling and VF prediction. [However, high myopia can introduce substantial retinal curvature, which may degrade OCT image quality and lead to unreliable RNFL segmentation and thickness measurements, creating an additional source of uncertainty.] Anyway, relying solely on color fundus images represents a difficult problem setting and may limit the achievable prediction performance. In addition, the neural network architecture appears relatively conventional and may not reflect more recent advances in model design. Finally, some of the clinical background statements, such as ‘Deep learning-based estimation of VF sensitivity from fundus photography has emerged as a promising alternative; …’ appear somewhat overstated given the current limitations and variability in this task.

  • 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.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the study addresses an important and challenging problem, the proposed approach is limited by both the choice of imaging modality and the model design. Predicting HVF outcomes from color fundus photographs in highly myopic eyes presents inherent clinical and technical challenges, which may restrict the achievable performance. In addition, the methodological novelty appears limited, and some of the clinical claims are not sufficiently supported. These concerns reduce the overall impact and strength of this submission.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    I thank the authors for their efforts in addressing the reviewers’ comments and questions. Humphrey visual field (HVF) testing is inherently subjective and noisy, and fundus photographs acquired using different camera systems, settings, and imaging protocols may substantially affect image quality. Although the revised manuscript provides additional clarification, I remain uncertain about the robustness and generalizability of this fundus-based VF estimation approach (ProtoVFs) for highly myopic eyes in broader real-world clinical settings, particularly because the current evaluation appears to rely on images obtained from a single camera at a single center with a relatively limited number of images. Additionally, it remains unclear whether the performance differences between the linear regression model and ProtoVFs were statistically significant.



Author Feedback

We thank the reviewers for recognizing the value of our work:

  • clinical relevance, described as a “clinically meaningful problem” (R1) and an “important and challenging problem” (R3) with “innovative significance” (R2).
  • methodological value, including a “conceptually interesting” prototype formulation to “impose structure and improve interpretability” (MR/R1), “more interpretable weights” (R2), and “quantitative and qualitative evaluation” (R1) with a “detailed ablation study” (R2).

Q1: We have released anonymized GitHub code and implementation details following MICCAI reproducibility guidance.

Q2 (MR/R1/R2/R3): Novelty, comparisons, and network choice

  • ProtoVF is the first prototype-based method for VF estimation. Instead of directly regressing 52 VF sensitivities, ProtoVF predicts coefficients over a compact VF basis learned from real VF data, constraining outputs to empirically observed patterns. Thus, the contribution of ProtoVF is not a new backbone, but a structured output-space formulation for VF estimation.
  • Our primary baseline uses the same image encoder (widely used ResNet) and training protocol, differing only in output formulation, which isolates the effect of the proposed prototype parameterization. Broader network selection is complementary to ProtoVF and does not undermine this controlled conclusion.
  • Our results also show favorable performance relative to previous myopia-specific AI studies (Ref. 26, 27).

Q3 (MR/R1): Prototype and coefficient learning We clarify that ProtoVF does not extract prototypes from fundus images. Prototypes are learned only from the training VF data. As described in Section 2.2, the prototype generator derives a fixed VF basis that provides an empirically constrained output space, while the ResNet learns image-derived coefficients over this basis.

Q4 (R2): Prototype visualization and number selection

  • Among all ProtoVFs, ProtoVF-PCA achieved the best predictive performance. The visualization, however, was intended for disease-related VF pattern analysis rather than model selection. We visualized AA because its archetypes represent boundary patterns in the VF distribution and resemble realistic, interpretable VF maps. By contrast, PCA components represent orthogonal directions of variance and are less naturally interpreted as standalone VF maps.
  • Fig. 3 reports VF reconstruction accuracy, not final fundus-to-VF estimation accuracy. Larger k (more prototypes) reduces reconstruction error for NMF/PCA/SVD, but also requires the image encoder to infer more coefficients and may capture VF label noise. Thus, k = 10 was used as a compact operating point balancing representation and learnability.

Q5 (MR/R1/R3): Imaging modality and study population

  • R3’s concern reflects the challenge of fundus-based VF estimation in high myopia, but this does not mean fundus images lack VF-relevant structural signal. Prior clinical studies (Ref. 21, 24) reported associations between fundus morphology and VF in myopia, and previous AI studies (Ref. 26, 27) also support the feasibility of modeling this relationship. Our results provide further evidence that this signal can be learned.
  • OCT may provide more direct structural measurements, but it is complementary rather than a replacement for our fundus-based model. Fundus photography is relatively low-cost, fast, accessible, and widely used in screening and primary-care settings. Therefore, the fundus-only setting is clinically relevant, especially for preliminary VF-risk estimation where perimetry or OCT may be unavailable.
  • Our study is intentionally focused on high myopia, an emerging vision-threatening condition for which VF assessment is important yet often unmet and challenging (Ref. 8). However, the prototype-based output formulation is not inherently limited to high myopia and could be investigated in other diseases with structured VF loss in future work, such as glaucoma, for which characteristic VF patterns have been well documented.




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.

    This paper presents a prototype-based framework for estimating visual field sensitivity from fundus photographs in high myopia. The problem is clinically relevant, and the prototype formulation is a potentially interesting way to impose structure and improve interpretability. However, the current submission still has several important limitations. In particular, the methodological novelty appears somewhat incremental, the experimental comparisons are not sufficiently strong or comprehensive, and the description of how prototypes and coefficients are learned from fundus images is not yet sufficiently clear. In addition, the clinical validation is limited by the single-center setting and the restricted study population, which weakens the generalizability of the conclusions. The authors are encouraged to address these concerns in the rebuttal.

  • 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.

    After considering the reviews and the rebuttal, I recommend acceptance. The paper addresses a clinically relevant problem and proposes a prototype-based output formulation for estimating visual field sensitivity from fundus photographs in high myopia. The rebuttal clarifies the prototype and coefficient learning process, the choice of prototype visualization, and the selection of prototype number. The work provides a focused and interpretable proof of concept with reasonable experimental support. I therefore 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.

    ProtoVF reframes fundus-to-VF estimation in high myopia as coefficient prediction over a learned visual-field prototype basis rather than direct 52-point regression — a formulation reviewers found “conceptually interesting” and a reasonable way to impose structure and improve interpretability. The clinical motivation (a low-cost, accessible screening signal) is well stated, and subject-level data separation supports methodological rigor. The rebuttal clarified that prototypes are learned only from training VF data, resolving the central point of confusion. Remaining concerns — fundus-only limitations, single-center/single-camera evaluation, and unclear statistical significance — are fair cautions on generalizability rather than fatal flaws. On balance, a useful, reproducible, and interpretable contribution.

    For camera-ready, the authors should fold in the rebuttal clarifications (particularly that prototypes are learned only from training VF data) and the committed code release, add the training/testing pseudocode requested in review, and report statistical testing of the reported gains alongside an explicit discussion of the single-center generalizability limitations.



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.

    After considering the reviews and the authors’ rebuttal, I recommend acceptance.

    The paper tackles a clinically relevant problem and proposes an interpretable framework for visual field estimation from fundus images. The methodology is sound and the experimental results are generally convincing.

    Most of the concerns raised during the review process have been adequately addressed in the rebuttal. While questions remain regarding broader generalizability and the lack of statistical significance analysis, these are limitations of the current study rather than fundamental weaknesses of the work.

    Overall, I believe the paper makes a useful contribution to the field and will be of interest to the community.



back to top