List of Papers Browse by Subject Areas Author List
Abstract
Long-term visual acuity (VA) forecasting after anti-VEGF therapy is important for counseling and follow-up planning in diabetic macular edema (DME), yet remains challenging when only early post-treatment findings are available. While prior OCT-based methods mainly focus on short-term response or single-endpoint prediction, multi-horizon VA forecasting from early longitudinal data remains insufficiently under-explored.
In this study, we assembled a real-world cohort of 188 anti-VEGF–treated DME patients with paired baseline and month-1 OCT scans, along with tabular OCT-derived biomarkers and non-imaging clinical variables. Using only these early data, we formulate a multi-horizon VA forecasting problem aimed at predicting visual outcomes at 3, 6, 12, 18, and 24 months, reflecting clinically meaningful follow-up intervals.
We propose ReVA, a response-aware multimodal framework that combines baseline and month-1 OCT features with tabular variables to capture disease status and early treatment response. ReVA integrates spatial OCT attention, dependency-aware tabular encoding, and cross-modal fusion to predict patient-specific long-term VA trajectories.
The proposed framework achieves MAE = 0.1246, RMSE = 0.1621, and R^2 = 0.6064 for 24-month VA prediction, with consistent performance across all forecast horizons. Our findings show that incorporating early treatment-response signals enables clinically meaningful long-term visual acuity forecasting, supporting data-driven decision support for routine anti-VEGF management. Code and pretrained models will be released on https://github.com/nguyenpbui/ReVA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1583_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/nguyenpbui/ReVA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BuiPhu_ResponseAware_MICCAI2026,
author = { Bui, Phuoc-Nguyen AND Vo, Van-Vi AND Le, Duc-Tai AND Bum, Junghyun AND Pham, Van-Nguyen AND Kim, Kiyoung AND Yu, Seung-Young AND Choo, Hyunseung},
title = { { Response-Aware Multimodal Learning for Post-treatment Visual Acuity Forecasting } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes ReVA, a model for 24-month visual acuity prediction after anti-VEGF therapy based on baseline and month-1 OCT images combined with clinical variables. By adopting RETFound as the backbone and TCN as the prediction module, the framework is reasonably designed and shows good potential for clinical application.
- 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 model is well designed and grounded in a clinically meaningful setting. 2 The ablation studies and the interpretability analysis provide additional insight into the model’s behavior. 3 The comparisons with existing methods support the effectiveness of the proposed approach.
- 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 Some abbreviations should be defined at their first appearance. In particular, the abbreviations used in the overview figure would be clearer if they were explicitly explained in the figure caption. 2 The paper mainly integrates existing methods within a specific clinical application scenario. While this makes the framework practical, it also limits the degree of methodological novelty. 3 Regarding OCT feature modeling, the use of the difference term m−b between the two visits raises a concern about potential misalignment or registration issues across scans, which should be discussed more clearly. 4 The model uses 3-, 6-, and 12-month outcomes as auxiliary supervision, but it is unclear how missing data across follow-up time points were handled during dataset construction. In practice, strict multi-time-point data collection can substantially increase the difficulty of implementation. In addition, the paper does not report horizon-specific prediction performance, nor does it show how performance changes when auxiliary supervision is removed. 5 The study is based on single-center data, and it is unclear whether the dataset will be made publicly available. This raises concerns about reproducibility. Moreover, the manuscript does not provide sufficient information on the data distribution, which makes the reported evaluation results harder to interpret.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 lack of the data distribution introduction.
- 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 have addressed my questions.
Review #2
- Please describe the contribution of the paper
The paper’s main contribution is a response-aware multimodal framework (ReVA) that predicts long-term visual acuity trajectories by explicitly modeling early treatment response (baseline → 1 month) and fusing OCT images with clinical/tabular data for multi-horizon forecasting.
- 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-Clinically meaningful problem: Long-term VA forecasting for DME using early visits. 2-Response-aware modeling: Explicit use of early treatment change (baseline → 1 month). 3-Multimodal integration: Combines OCT images and clinical/tabular data effectively. 4-Multi-horizon formulation: Predicts full VA trajectory (3–24 months), not just a single endpoint. 5-Strong performance: Improves over prior methods (lower MAE, higher R²). 6-Use of foundation models: Leverages pretrained OCT transformer (RETFound + LoRA). 7-Interpretability: Attention maps highlight clinically relevant regions. 8-Ablation studies: Show benefits of multimodal fusion and response modeling.
- 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 method is formulated as a multi-horizon forecasting problem, the experimental results primarily report performance at 24 months. Detailed per-horizon evaluation (e.g., 3, 6, 12, 18 months) is missing, making it difficult to assess the model’s temporal consistency and robustness across prediction horizons. 2-The reported performance improvements are relatively modest, and several methods appear close numerically. Please provide statistical significance analysis across folds, including the specific test used, to better support the claimed superiority of the proposed method. 3-Limited novelty: Combines known components (ViT, LoRA, cross-attention, TCN) with modest conceptual innovation. 4-Small dataset: Only 188 patients, raising concerns about generalization. No external validation: Results are limited to a single cohort. Marginal performance gains: Improvements over strong baselines are relatively small. Insufficient statistical analysis: Limited clarity on significance testing across methods. Method complexity vs gain: Architecture is relatively complex for incremental improvements. Handling of missing data unclear: Real-world longitudinal missingness not deeply analyzed.
- 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?
It’s a useful, practical paper, but not strong enough (novelty + evaluation depth) for a clear accept.
- 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 #3
- Please describe the contribution of the paper
The authors have developed a multi-modal model that utilizes tabular data and OCT to explicitly model early post treatment changes for long range VA forecasting
- 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.
Major Strength: 1.Clear and Well defined problem definition with building blocks of the model detailed clearly 2.Figure 1 clearly represents the intention and model architecture 3.The authors have utilized TCN to encode the temporal time, while I think there is a chance of information loss (temporal effect) the cross-attention module mixes the learned feature properly.
- 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.
Fig 2, attention heatmap needed more details specially how to interpret it better. The map image should have been explained without the heatmap first then with the attention map. THis would help readers understand the relevance of this map better.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(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?
This was a clean read where the authors have explained the steps clearly. Experiments were sufficient and results suggest performance boost compared to baseline models.
- 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.
FOr me the work was already acceptable without the rebuttal. The only question I had related to the heatmap was addressed.
Review #4
- Please describe the contribution of the paper
This paper presents a study that predicts long-term visual acuity following anti-VEGF therapy for diabetic macular edema by combining OCT images with clinical data. By integrating large-scale models for each modality—such as RETFound ViT-Large and T2G-Former—with techniques like LoRA, Cross-Attention, and Temporal Learner, the authors have successfully improved prediction performance. Furthermore, the study demonstrates that using information from two visits in a complementary manner improves accuracy compared to simple single-time-point predictions. It effectively combines state-of-the-art methods, and the role and effectiveness of each method are clearly illustrated through ablation studies. The text is well-structured and easy to understand.
- 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.By effectively utilizing foundation models, the authors achieve high-accuracy visual acuity prognosis predictions using a relatively small dataset. Given the limited availability of labeled data in prognosis prediction, this approach is expected to become the standard practice. 2.The approach of combining information from multiple visits is reasonable from a clinical perspective, and the authors also explain the clinical rationale for incorporating other methods. 3.Comprehensive benchmarking against previous methods has been conducted, and ablation studies have been performed appropriately. 4.The challenge of long-term visual acuity prediction is of high clinical importance, and the clinical significance of the improved prediction accuracy achieved by this method is discussed accurately.
- 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 authors combine multiple deep learning methods, and ReVA is a framework with a very large number of parameters. It is unclear which of these parameters were fixed and which were learned from the data in this study. Since the model was trained on a small number of cases, it is believed that the number of parameters used for training should be reduced as much as possible to avoid overfitting. 2.In the abstract, the authors claim that a key advantage of this model is its ability to predict visual prognosis not only at 24 months but also at 3, 6, 12, and 18 months; however, the accuracy of these predictions at 3, 6, 12, and 18 months is not presented in the results or discussion sections, nor is the impact of these predictions on the 24-month prediction examined.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(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?
This is an excellent paper that effectively combines appropriate methods and provides clear clinical interpretation, and it appears that most of the validation was conducted appropriately. It is also interesting that the study demonstrates improved predictive accuracy with longitudinal data. This research would be of even greater value if the parameters trained using the data from this case series and the handling of the multiple endpoints highlighted as key features in the abstract were clarified.
- 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
Author Feedback
We thank the reviewers for recognizing ReVA’s clinical importance, clear formulation, benchmarking, ablations, and interpretability analysis. We address the main concerns below.
[Meta/R1/R2: Novelty and response-aware modeling] Many prior VA prediction studies focus on single-endpoint prediction or short-term treatment response, whereas ReVA targets forecasting VA at 3/6/12/18/24 months from only baseline and month-1 data. Rather than naively concatenating the two visit representations, ReVA extracts visit-level OCT/clinical features and explicitly models early treatment response via a gated delta module applied to the feature difference between month-1 and baseline. Since this delta is computed in feature space, pixel-level OCT registration is not required. Fixed scan/view ordering ensures corresponding view-level features are compared, while gating and delta dropout reduce sensitivity to noisy visit-to-visit changes. The aggregation ablations in Table 2c support the benefit of this response-aware design.
[Meta/R1/R2/R4: Multi-horizon results, missing labels, and auxiliary supervision] We will add detailed per-horizon results. Under identical patient-level 5-fold CV, ReVA achieved MAE/RMSE/R² of 0.1135±0.0019/0.1495±0.0114/0.6687±0.0429, 0.1261±0.0118/0.1640±0.0228/0.5846±0.1042, 0.1218±0.0087/0.1556±0.0175/0.6420±0.1005, 0.1283±0.0105/0.1631±0.0111/0.5914±0.0851, and 0.1246±0.0066/0.1621±0.0056/0.6064±0.0727 at 3/6/12/18/24 months, respectively. Compared with other methods, ReVA ranks second at 3 months and first at all horizons from 6 to 24 months. Patients with complete labels at all five horizons were used to ensure comparable per-horizon evaluation; the observation mask is included to support future incomplete-label settings. Removing intermediate-horizon supervision degraded 24-month performance to 0.1341±0.0057/0.1667±0.0099/0.5829±0.0827, supporting the value of trajectory-level supervision.
[Meta/R2: Clinical baseline and added architectural value] A VA-only baseline using baseline/month-1 VA obtained MAE/RMSE/R² = 0.1396±0.0122/0.1763±0.0103/0.5368±0.0742 at 24 months; ReVA reduced MAE/RMSE by 10.7%/8.1%, showing added value beyond early VA alone.
[Meta/R2/R4: Statistical testing] We performed one-sided paired Wilcoxon tests on out-of-fold patient-level absolute errors using identical 5-fold splits (α=0.05). ReVA significantly outperformed compared methods at 12/18/24 months, and the 24-month improvement over the variant without auxiliary supervision was also significant. We will report the exact p-values in the revision.
[Meta/R1/R2/R4: Dataset statistics, generalization, and sharing] We will add cohort statistics: age 61.7±11.6 years; 98 female/90 male; 97 left/91 right eyes. Baseline VA/CST are 0.499±0.248 logMAR and 425.5±97.8 μm; month-1 VA/CST are 0.583±0.263 and 339.3±73.3 μm. Lesion prevalence is IRF/SRF/PED: 176/73/3 of 188.Mean VA at 3/6/12/18/24 months is 0.572±0.262/0.577±0.263/0.566±0.266/0.545±0.259/0.562±0.262 logMAR. Although modest, the cohort spans heterogeneous DME presentations across demographics, VA/CST, and lesion, making it clinically nontrivial. However, we acknowledge the single-center limitation and will state external/multicenter validation as future work. We will release code and pretrained models; de-identified data may be shared upon reasonable academic request, subject to institutional approval.
[R2/R4: Model complexity] Although ReVA has 310.4M total parameters, only 7.1M (2.3%) are trainable. RETFound is frozen; only LoRA layers, tabular encoder, fusion module, and temporal prediction head are optimized. Thus, ReVA is parameter-efficient rather than fully fine-tuned, reducing overfitting risk in this modest cohort.
[Meta/R1/R3: Presentation and interpretability] We will define all abbreviations at first use and in Fig. 1, and revise Fig. 2 to include OCT context without and with heatmaps, with a clearer explanation of how the highlighted regions should be interpreted.
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 addresses an important and clinically meaningful problem: long-term visual acuity forecasting after anti-VEGF treatment using early OCT and clinical information. The proposed ReVA framework is practically motivated and shows promising performance, especially for 24-month visual acuity prediction. However, several key claims require stronger evidence. In particular, the paper frames the task as multi-horizon forecasting, but detailed per-horizon results for 3, 6, 12, and 18 months are not sufficiently reported. The role of auxiliary supervision from intermediate follow-up outcomes also needs to be validated through ablation. In addition, the reported performance gains are relatively modest, and the paper should provide stronger statistical testing against all major baselines. Since the tabular-only model is already close to the full multimodal model, simple clinical baselines based on baseline and month-1 VA should be included to justify the added architectural complexity. The authors should also clarify OCT alignment across visits, explain the heatmap visualization in Fig. 2 more rigorously, and provide a clearer release plan for code, pretrained models, and data.
- 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.
All reviewers who provided explicit post-rebuttal opinions support acceptance. Reviewer #1 states that the authors addressed the raised questions, and Reviewer #3 indicates that the work was already acceptable before rebuttal and that the heatmap-related concern was resolved. The rebuttal also addresses the main meta-review concerns by providing per-horizon results, auxiliary-supervision ablation, a VA-only clinical baseline, paired Wilcoxon significance testing, cohort statistics, model-complexity clarification, and a clearer release plan for code and pretrained models. Remaining limitations, including single-center validation and modest dataset size, should be acknowledged in the final version, but they do not outweigh the paper’s clinical relevance and overall positive post-rebuttal consensus. The authors should provide the full open-source implementation to improve the reproducibility and broader impact of the paper.
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.
All reviewers recommended acceptance for this paper. The authors should commit to addressing the remaining concerns in the camera-ready version. 1) more clearly report performance across all VA forecast horizons, not only the main 24-month. 2) clarify the incremental value of imaging over tabular-only predictors. 3) provide the statistical testing and uncertainty analysis. 4) provide a clear release plan for code, pretrained models, and data splits.
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 addresses a clinically important problem and presents a practical multimodal framework for forecasting long-term visual acuity after anti-VEGF treatment. The use of baseline and month-1 OCT together with clinical variables is well motivated, and the overall framework is technically coherent. The experimental results, benchmarking, ablations, and interpretability analysis are solid. The rebuttal addressed the main concerns convincingly. In particular, the authors clarified the response-aware modeling design, provided detailed per-horizon results, explained the handling of missing labels and auxiliary supervision, added statistical significance testing, and reported cohort statistics. They also clarified that only a small fraction of parameters are trainable, which helps address concerns about overfitting despite the large backbone. Some limitations remain, especially the single-center dataset and the moderate methodological novelty, but these do not outweigh the strengths of the work.
