Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

The retina offers a noninvasive window into neurodegenerative disease, capturing subtle structural patterns associated with a risk of future cognitive decline. Vision–language alignment frameworks such as REVEAL have shown that pairing retinal fundus images with structured clinical risk narratives improves early prediction of Alzheimer’s disease (AD). A key design choice in these approaches is the use of phenotypic grouping, where individuals with similar risk profiles are treated as multi-positive pairs during contrastive learning. However, existing methods operationalize phenotypic similarity as a discrete construct, relying on hard group assignments that impose rigid supervision and decouple group formation from representation learning. We propose a continuous formulation of phenotypic structure within contrastive learning. Rather than assigning samples to fixed clusters, we model inter-subject similarity as a differentiable weighting function derived from intra-modality embedding similarities in both retinal images and risk profiles. These weights define soft multi-positive relationships through a continuous aggregation operator, enabling graded supervision that reflects the spectrum nature of disease risk. We further introduce a soft-target contrastive objective that jointly learns cross-modal alignment and phenotypic structure in an end-to-end manner. Evaluated on UK Biobank retinal imaging data for incident AD prediction, the proposed framework consistently outperforms discrete group-based contrastive learning and standard vision–language baselines. By treating phenotypic similarity as a learnable, continuous signal rather than a fixed grouping rule, our approach provides a principled and robust foundation for population-scale neurodegenerative risk modeling from multi-modal retinal and clinical data. Code is available at: \url{https://github.com/lab-smile/REVEALpp}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2649_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lab-smile/REVEALPP

Link to the Dataset(s)

UK-Biobank Dataset: https://www.ukbiobank.ac.uk UK Biobank is a massive, long-term biomedical database housing de-identified genetic, biological, and health-related records from 500,000 volunteers across the United Kingdom. Established to improve the prevention, diagnosis, and treatment of serious illnesses, it serves as an open-access resource for global health research.

BibTex

@InProceedings{MeiEth_REVEAL_MICCAI2026,
        author = { Meidinger, Ethan AND Leem, Seowung AND Zhao, Zeyun AND Fang, Ruogu},
        title = { { REVEAL++: Differentiable Phenotypic Grouping for Vision–Language Retinal Modeling of Alzheimer’s Disease Risk } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

The paper proposes a soft Differentiable Phenotypic Grouping method within a contrastive learning framework and applies it to the published REVEAL for Alzheimer’s disease (AD) risk prediction. Results show that the proposed method achieved the best performance
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.The proposed method extends the REVEAL with a soft Phenotypic Grouping within contrastive learning and shows an advantage in AD risk prediction
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.Clarity of the overall framework: The overall framework is not clearly presented. The proposed method appears to be built upon REVEAL, but this prior work is not properly cited in the main text. Although the REVEAL paper is included in the reference list, it is not explicitly referenced where relevant, which affects clarity and attribution. 2.Task definition: The task is not clearly defined. In particular, it is unclear what form the Alzheimer’s disease (AD) risk prediction takes (e. g. , classification, regression, or risk score estimation). 3.Clarity of notation: The definition of symbols needs to be clarified. For example, the meaning of \tau in Section 2.3 is not clearly explained, which may lead to confusion.
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?

Clarity of writing, Completeness of the Paper
Reviewer confidence

Confident but not absolutely certain (3)
[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.

Although the authors claim to have corrected some errors and revised certain definitions, I do not think these changes adequately address the overall concerns through a minor revision, especially those related to the overall framework. Therefore, I maintain my final recommendation of rejection.

Review #2

Please describe the contribution of the paper

The paper proposes REVEAL++, a vision-language alignment framework for Alzheimer’s disease (AD) risk prediction. Based on the prior work REVEAL, REVEAL++ introduces a key innovation of substituting the original hard threshold-based phenotypic grouping with continuous phenotypic weighting in the group-aware contrastive learning framework. REVEAL++ applies a soft-target contrastive objective based on phenotypic similarity to enable multi-positive contrastive supervision so as to reflect the spectrum-like nature of AD risk and improve the prediction performance. The method is evaluated on a clinical Alzheimer’s dataset and is compared with traditional SVM baselines and vision-language baselines. AUROC, balanced accuracy, F1-score and MCC are reported in the results to evaluate the effectiveness of the proposed method. The proposed REVEAL++ consistently outperforms baseline methods in all the reported metrics and achieves new state-of-the-art performance.
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) The paper presents a clear and well-motivated improvement over the prior method, where the hard threshold-based phenotypic grouping strategy in REVEAL is effectively redesigned into continuous phenotypic weighting considering the limitations of introducing hard artificial boundaries, resulting in improved performance. (2) The proposed modifications are clearly described, and implementation details and hyperparameter settings are sufficiently documented, facilitating reproducibility. (3) The experimental results show consistent performance gains over a range of visual-language baselines and the prior method REVEAL, indicating that the proposed modifications are practically effective.
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 overall methodological novelty is somewhat limited, as the proposed method (REVEAL++) is only an modification of the prior work (REVEAL), with relatively minor changes to the original design. (2) The paper only reports performance of the modified model compared to the prior work with a single additional result without more comprehensive analysis. The lack of deeper experimental studies, for example, including comparisons with image-based CNN/Transformer baselines or studies on the contribution of each modality, may limit the overall novelty and the strength of the empirical evidence supporting the proposed improvement.
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?

REVEAL++ presents a well-motivated improvement over REVEAL by replacing a hard threshold-based phenotypic grouping strategy with continuous phenotypic weighting, which is intuitive and empirically effective. The method is clearly described with sufficient implementation details, and shows consistent performance gains over several visual-language baselines and REVEAL, suggesting practical usefulness.

However, the overall novelty is incremental, and the experimental evaluation is limited in scope, lacking deeper comprehensive comparisons and ablation studies.

Overall, despite these limitations, the method demonstrates consistent empirical improvements and reasonable design clarity. With additional analysis and clearer justification, the paper would meet the criteria for acceptance.
Reviewer confidence

Somewhat confident (2)
[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 paper presents REVEAL++, a well-motivated extension of REVEAL that improves retinal image-based AD risk prediction by replacing discrete group assignments with a differentiable phenotypic weighting mechanism. This modification better captures the continuous nature of preclinical AD progression and consistently improves performance over REVEAL and multiple vision-language baselines. While the novelty builds on prior work, the authors distinguish inherited components from new contributions and provide thorough ablation analyses within the page limit. Additional clarifications, including task definition, notation, and evaluation protocol, further strengthen understanding and reproducibility will be supported via code release. Overall, the paper makes a significant methodological contribution to early AD prediction from retinal data and meets the standard for MICCAI acceptance.

Review #3

Please describe the contribution of the paper

The paper introduces a differentiable phenotypic weighting framework for vision–language modelling of Alzheimer’s disease (AD) risk from retinal images and associated clinical narratives. Building on prior work (REVEAL) the key idea is to replace hard phenotypic group assignments with a continuous, learnable formulation that captures inter-subject similarity.

The method constructs a soft phenotypic weighting matrix that encodes pairwise similarity between subjects using sigmoid-based gating functions with learnable sharpness. This matrix is then incorporated into a multi-positive contrastive learning objective, allowing sampled pairs to contribute to the loss as positives or negatives with varying degrees of strength, rather than through hard assignment.

The approach is evaluated on a large-scale UK Biobank cohort (~40k subjects), where it demonstrates improved AD risk prediction performance over prior vision–language baselines, including REVEAL, achieving state-of-the-art results within this 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 presents a clear and conceptually sound methodological contribution. The proposed approach is well motivated and addresses a genuine limitation of the GACL approach. The formulation is elegant, integrates naturally into standard contrastive objectives, and appears novel in this context.

The paper is nicely structured and the methodology is easy to follow, supported by a strong visual overview in Fig. 1.The combination of intra-modality similarity to define phenotypic relationships and cross-modality similarity for alignment is well thought through.

The experimental evaluation is thorough and convincing. The use of a large-scale dataset (UK Biobank) strengthens the empirical claims, and the inclusion of strong baselines, including prior work from the same line of research, provides a fair comparison. Running experiments across multiple random seeds further supports the robustness of the results.

Empirically, the method consistently outperforms baselines and improves upon the prior REVEAL framework, suggesting that the proposed weighting mechanism provides meaningful gains.

Finally, the paper is generally well written and clearly structured, with only minor issues in notation and clarity.
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 most significant issue is the lack of detail regarding the AD prediction task itself. Key information such as disease prevalence, label definition, and task setup is missing from the main paper. While these details are available in the supplementary material (as prior work), they should be briefly summarised in the main manuscript to ensure the results are interpretable and self-contained.

There are also some minor clarity issues in the methodology. The distinction between encoders and projection layers is not fully clear, and the notation appears slightly inconsistent across sections. Specifically, Section 2.3 should be clarified: $E_I$ and $E_T$ are described as encoders, and the projection layers described as separate components. However, it would seem from the rest of the description that $E_I$ and $E_T$ include the projection layers. Please make this explicit, confirming that $\mathbf{z}^{I}p \in \mathbb{R}^{d}$ and $\mathbf{z}^{T}_p \in \mathbb{R}^{d}$, if that’s indeed the case. There’s also no need to redefine $\hat{\mathbf{z}}{p}^{I}$ and $\hat{\mathbf{z}}_{p}^{T}$ in Section 2.4, as they’re already introduced in Section 2.3.Additionally, Section 3.2 discusses how the phenotypic similarity thresholds were selected, but the details are quite vague. Please also be clear that this refers to $\tau_F$ and $\tau_T$ from Section 2.4 / Equation 4, if indeed it does.

Another key concern is reproducibility. The paper does not provide an anonymised code repository or a commitment to release it upon acceptance, which limits transparency and the ability to verify results.

More broadly, while the method achieves state-of-the-art performance, the absolute performance remains modest (AUC ~0.68), and the paper does not sufficiently contextualise this in terms of task difficulty or clinical relevance.
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

Overall, I thoroughly enjoyed reading this paper. I think it could be a strong contribution that would attract interest from the MICCAI community, subject to concerns being addressed about the reproducibility of the methodology.
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 methodological contribution is both elegant and well-motivated, and addresses a clear limitation in prior work. The paper is also clearly presented, with a coherent narrative and a method that is easy to follow despite its novelty. Furthermore, the experimental evaluation is rigorous and lends credibility to the claims of SOTA performance. I have erred towards weak acceptance due to the lack of anonymised source code (or a commitment to release it upon acceptance) and lack of detail about the AD prediction task.
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 am satisfied that the authors rebuttal satisfactorily addresses my minor concerns with the presentation of the paper. Having already felt the paper was worthy of acceptance, my decision to accept remains the same.

Author Feedback

We thank the reviewers for their constructive feedback. We are encouraged that reviewers recognized REVEAL++ as a “clear and well-motivated improvement” (R2), a “conceptually sound methodological contribution” with an “elegant” formulation (R3), and a method that addresses a genuine limitation of hard GACL by replacing discrete grouping with continuous phenotypic weighting (R2, R3, Meta). We also appreciate that reviewers found the results consistently improved over REVEAL and other vision-language baselines (R2, R3, Meta). Q1: (R1, R3, Meta) What exactly is the AD risk prediction task? A1: The submitted experiments formulate incident AD prediction as a binary classification task. Participants free of AD at baseline retinal imaging are labeled positive if they subsequently develop AD during follow-up, while controls do not meet the AD definition during follow-up. Evaluation is performed using subject-level splits and classification metrics: AUROC, balanced accuracy, F1, and MCC. We will revise Sec. 3.1/Table 1 to include label construction, case/control counts, prevalence, split structure, and evaluation protocol. Q2: (R1, R2, Meta) What is inherited from REVEAL, and what is new? A2: REVEAL++ is intentionally built on REVEAL. The inherited components are the retinal image-clinical narrative alignment setting, RETFound/GatorTron encoders, and the motivation for group-aware contrastive learning. The new contribution is replacing REVEAL’s binary thresholded group membership with a differentiable phenotypic weighting matrix and a soft-target multi-positive contrastive objective. This changes the supervision from hard positive/negative assignment to graded pairwise phenotypic similarity, which better matches the continuum-like nature of preclinical AD progression. We will explicitly cite REVEAL where inherited components are introduced and distinguish between inherited and new components in Sec. 2.Q3: (R1, R3, Meta) Can the notation be clarified? A3: In the submitted method, image/text backbones and projection heads are separate components. We will revise the notation so E_I and E_T denote the image and text backbone encoders, while P_I and P_T denote the projection heads. We will remove redundant definitions of normalized embeddings. To avoid overloading tau, we will denote contrastive temperature as T_c = exp(-s), reserving tau_F and tau_T for the phenotypic similarity thresholds in Eq. 4.We will also clarify that tau_F and tau_T are tuned as hyperparameters using Optuna over ranges initialized from empirical intra-modality cosine similarity distributions. Q4: (R2, Meta) Are there sufficient ablations and baselines? A4: We appreciate the request for deeper ablations and modality analysis. Since MICCAI rebuttal guidelines do not allow new experimental results in the rebuttal, we cannot introduce additional ablation results here. We clarify that the submitted Table 2 already includes a component comparison between “Ours (no GACL),” “Ours (with GACL),” and “Ours (Phenotypic Weighted),” directly isolating the effect of replacing hard grouping with continuous weighting. This table also compares against retinal and biomedical vision-language baselines, including RETFound+GatorTron, RETCLIP, BiomedCLIP, PMC-CLIP, and KeepFIT-CFP. Q5: (R3, Meta) How should the modest AUROC be interpreted clinically? A5: Our claim is not that an AUROC of around 0.68 is sufficient for standalone clinical deployment. Rather, the results show improved representation learning for a difficult, low-prevalence, long-horizon incident AD prediction task (mean time to diagnosis: 8.68 years) using scalable retinal and clinical risk-factor data. Q6: (R3) How will reproducibility be addressed? A6: We will release an anonymized code repository upon acceptance. UK Biobank raw data cannot be redistributed due to data-use restrictions, but we will provide documentation enabling authorized UK Biobank users to reproduce the submitted experiments.

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.
- The paper proposes a soft differentiable phenotypic grouping strategy for vision-language AD risk prediction. Replacing hard phenotype grouping in REVEAL with continuous phenotypic weighting is a well-motivated and elegant idea, and the reported results show consistent gains over REVEAL and other vision-language baselines.
- However, the novelty is still incremental because the method is mainly an extension of the prior REVEAL framework.
- The task definition and experimental setup need clearer presentation. The paper should explicitly state what “AD risk prediction” means, including the label definition, disease prevalence, and whether the task is classification, regression, or risk score estimation. Some notation and methodological details, such as the role of thresholds and projection layers, also need clarification.
- The empirical results are promising, but additional analysis would strengthen the claim. The authors should provide deeper ablations, comparisons with image-only CNN/Transformer baselines, modality contribution analysis, and clearer discussion of the modest absolute AUC and clinical relevance.
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 rebuttal satisfactorily clarifies the task definition, the distinction between inherited REVEAL components and the new differentiable phenotypic weighting, notation, threshold selection, clinical interpretation of the modest AUROC, and reproducibility plans. Despite incremental novelty, the consistent empirical gains and positive post-rebuttal consensus support 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.

After considering the reviews and rebuttal, I recommend acceptance. The paper proposes REVEAL++, a differentiable phenotypic grouping strategy for vision-language retinal modeling of AD risk. The main contribution is to replace hard threshold-based grouping in REVEAL with continuous phenotypic weighting and a soft-target multi-positive contrastive objective. Although this is an incremental extension of prior REVEAL work, the idea is well motivated, technically clean, and empirically shows consistent gains over REVEAL and other vision-language baselines.

The rebuttal addresses several important concerns. The authors clarified that the task is binary incident AD prediction, explained what is inherited from REVEAL versus what is newly introduced, clarified notation and threshold definitions, and appropriately contextualized the modest AUROC as representation learning for a difficult long-horizon risk prediction task rather than standalone clinical deployment. Some limitations remain, including limited novelty, modest absolute performance, and the need for clearer task details and code release in the final version. However, the overall methodological contribution and post-rebuttal reviewer support justify acceptance.

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 rebuttal, I recommend acceptance. The paper presents REVEAL++ for retinal image-based AD risk prediction, with a differentiable phenotypic weighting mechanism to better model continuous preclinical AD progression. The rebuttal addressed the main concerns raised during review. The paper shows consistent improvements over REVEAL and several vision-language baselines, and the authors clarified the task definition, notation, evaluation protocol, and new contributions. I therefore recommend acceptance.

back to top

REVEAL++: Differentiable Phenotypic Grouping for Vision–Language Retinal Modeling of Alzheimer’s Disease Risk

Author(s):