List of Papers Browse by Subject Areas Author List
Abstract
Infectious keratitis is a leading cause of corneal blindness worldwide, and in vivo confocal microscopy (IVCM) has become a key non-invasive diagnostic tool for this condition. However, existing methods predominantly rely on single-frame classification, ignoring pathogen-specific infection depths and frequently causing misdiagnosis. Furthermore, inherent imaging degradations like low signal-to-noise ratios and poor contrast severely compromise classification accuracy. To address these challenges, we propose ReDIVE, a Retinex-enhanced Depth-Informed Vertical Embedding framework for sequence-level infectious keratitis classification. ReDIVE integrates a customized, differentiable multi-scale Retinex preprocessing module, a robust encoder-decoder network, and a specialized depth embedding module. This approach leverages clinical prior knowledge to enhance sequence analysis while significantly improving image quality. Experiments on a large-scale dataset demonstrate that ReDIVE outperforms state-of-the-art method
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5257_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{LiaAnx_ReDIVE_MICCAI2026,
author = { Liao, Anxu AND Zhang, Ling-Xiao AND Zhang, Fang-Lue AND Zhou, Jie AND Qi, Mengying AND Chen, Fen AND Zeng, Qingyan AND Gao, Lin AND Chen, Yiqiang},
title = { { ReDIVE: A Two-Stage Hybrid Network for Keratitis Diagnosis via Retinex-based Enhancement and Depth-Informed Vertical Embedding } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents ReDIVE for sequence-level infectious keratitis diagnosis using in vivo confocal microscopy (IVCM) images. The work introduces a customized differentiable multi-scale Retinex (CDMSR) module for image enhancement and a hybrid depth-informed feature embedding strategy that fuses visual features with absolute and relative depth information, which is then processed by a Vision Transformer for sequence classification. Experiments on a private dataset demonstrate that the proposed method outperforms existing state-of-the-art approaches.
- 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 motivation is clear and clinically meaningful. Integrating depth information into sequence-level infectious keratitis classification is a novel and promising idea. 2.The design of depth embedding and the strategy of treating frame-level features as distinct tokens provide valuable insights for sequence-based medical image analysis.
- 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.As a core contribution, the CDMSR module lacks clear details: the computation of loss L_S1 (including L_tv and L_ent) and its design rationale are not explained. The definition of I_in is ambiguous; if it denotes the raw input, the normalized L1 loss between I’_raw and I_in may conflict with the goal of image enhancement. 2.The implementation is not well aligned with the core idea of depth-aware learning. It would be more reasonable to jointly optimize image feature encoding (Stage 2) and sequence classification (Stage 3), rather than using separate stages. The paper does not clarify whether every frame in the IVCM sequence is diagnostically informative; single-frame optimization may learn misleading features if some frames are non-informative. 3.The fairness of comparison is questionable: it is unclear whether baseline methods use single-frame or sequence inputs. The strong performance of the “w/o CDMSR & D_rel & D_abs” variant raises concerns that gains may come mainly from larger model parameters.
- 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.
- 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
The paper states that “relative depth encodes inter-frame difference ∆d into a 54-dimensional representation (L × 6, L = 9)”. Here, ∆d should be ∆D (a L×L matrix as defined in Fig. 1), and this notation inconsistency should be corrected.
- 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?
Key implementation details of the core CDMSR module are missing, the training strategy is not well justified, and the experimental comparison lacks fairness justification.
- 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 is a clinically meaningful paper that addresses an important limitation of prior IVCM-based keratitis diagnosis methods. The motivation is clear, the framework is coherent, and the empirical results are very strong. In particular, the idea of explicitly combining sequence-level modeling with pathogen-related depth priors is well justified and experimentally supported.
- 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 proposed framework is logically organized. The CDMSR module addresses a real imaging problem in IVCM, namely low contrast and non-uniform illumination. The hybrid feature embedding then combines local image features with absolute and relative depth information, and the transformer models inter-frame dependencies for sequence classification. These components are conceptually aligned with the stated clinical and technical goals. The framework diagram is also clear and helpful. In addition, the statistical analysis of depth distributions across pathogen categories is valuable, since it provides empirical support for the proposed depth-informed design rather than relying only on intuition. The reported Kruskal–Wallis and Mann–Whitney test results strengthen the paper’s core claim.
- 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 provides a fair amount of training detail, but given the complexity of the full pipeline, reproducibility remains a concern. The reported dataset is large and clinically meaningful, which is a strength, but the evaluation is still confined to a single-center in-house collection. The authors themselves acknowledge this limitation in the conclusion. Because the method makes strong claims about establishing a new paradigm for IVCM-based diagnosis, broader validation would be desirable.
- 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.
(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 paper has strong motivation, a well-designed framework, and excellent empirical performance. It would be even stronger with better reproducibility support, clearer baseline adaptation details, and broader validation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The authors present a two-stage model for classifying infectious keratitis from IVCM imaging. Their pipeline initially employs a custom image enhancement algorithm, followed by a ConvNeXT V2 base model with depth context, and finally a ViT that operates on a fused feature sequence. They show their model achieves high AUC (0.9948), outperforming several baselines.
- 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 proposed methodology is well motivated. The idea of aggregating information across images in a sequence is important for classifying IVCM images since sequences will show complementary information from the same disease. Similarly, depth context can help with differentiating between variability in anatomy versus disease.
Their model achieves remarkably high accuracy and AUC, outperforming several baselines.
- 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.
While the number of images in the dataset is high, the actual number of scans is relatively limited; the test set only has 3-4 patients in each category. While each patient has many images, these images will be correlated as they will show overlapping or adjacent anatomy and the same disease. Having so few patients may not represent the variability of disease presentation. Also, reporting the number of frames is somewhat deceptive as the proposed algorithm aggregates information across each sequence, so the number of sequences would be a better representation of the data.
While the authors put a lot of emphasis on their image enhancement method and the depth context, they have relatively little impact on the results. The ablation study shows that without these, the AUC is still high (0.9924) and adding in CDMSR and Drel actually result in reduced AUC (0.9905). While the model with all components has the highest performance, the marginal differences and lack of statistical testing or measure of variation call into question whether this difference is meaningful.
Most of the benefit in this study seems to be from aggregating images in a sequence. As mentioned above, the image enhancement and depth context have marginal impact on results, so the only apparent difference between the proposed method and the baseline ConvNeXT is the ViT to aggregate over the sequence. While this procedure does show superior results, the baseline methods all seem to be operating on each image independently, so they aren’t performing the same task. In order to truly evaluate the proposed algorithm, the baselines should perform a similar aggregation. Otherwise, they are mainly demonstrating that looking at a full sequence is better that evaluating from a single image.
- 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 is important to emphasize utilizing a full sequence rather than individual images when classifying IVCM. However, there are some limitations in the evaluation and it is unclear whether some of the emphasized technical innovations actually impacted results.
- 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 sincerely thank all reviewers for their insightful comments, constructive suggestions, and recognition of our work. Their detailed feedback has helped us refine the presentation and strengthen the manuscript. Below we address the main concerns raised, and all suggested clarifications will be incorporated into the final manuscript.
A. CDMSR Design and Loss (R1-W1) Due to page limitations, we condensed the loss explanation in the manuscript. Our L_S1 (Eq. 7) is inspired by the Retinex optimization framework of Zhao et al. [20], which also addresses IVCM enhancement. The loss balances three objectives through carefully weighted terms: 1) L_TV spatially smooths the predicted parameter maps, preventing abrupt artifacts. This translates the illumination smoothness prior (e.g., ||∇ L|| in [20]) into direct regularization on modulation parameters. (2) L_content preserves structural consistency between the enhanced output and the min-max stretched input I_in, analogous to the fidelity term ||R*L - S||_2 in [20], while suppressing noise from over-enhancement. (3) L_ent maximizes local texture gradients to amplify subtle pathological structures, inspired by reflectance gradient regularization. All loss weights were determined empirically. Together, these terms enable effective and stable enhancement. We will expand this explanation in Sec. 2.2.Additionally, we will correct the notation Δd to ΔD (the L x L relative depth matrix) in Sec. 2.3, consistent with Fig. 1. B. Training Strategy (R1-W2) Stage 2 pre-trains the ConvNeXt backbone on enhanced frames to learn robust single-frame features; Stage 3 then introduces the ViT for sequence classification, after which both stages are jointly fine-tuned (see Sec. 3.2). This separate-then-joint strategy stabilizes the heterogeneous objectives of frame-level encoding and temporal modeling, avoiding potential gradient conflicts from direct joint training. Regarding frame informativeness: during dataset curation, we filtered low-information images through gold-standard culture and multi-clinician annotation. Additionally, the ViT self-attention mechanism naturally learns to assign higher weights to diagnostically informative frames.
C. Comparison Fairness (R1-W3, R3-W3) Among existing methods for infectious keratitis classification, only Li et al. [7] adopted sequence-level classification, but their training code is not publicly available and proved difficult to reproduce. All other baselines employ single-frame classification. Our method naturally leverages depth priors to construct the relative depth matrix, making sequence input an inherent and well-justified design choice. We fully agree with the reviewer that sequence-level diagnosis offers meaningful benefits for this task, which holds valuable implications for future clinical deployment.
D. Ablation Gains (R3-W2) We acknowledge that over the already high baseline (AUC 0.9924), absolute improvements appear incremental. However, accuracy increases by +1.67% (93.75 -> 95.42) and F1 by +1.62%. In a clinical screening context, such improvements translate to a meaningful reduction in misdiagnosis, which directly impacts patient care.
E. Dataset and Generalizability (R2-W2, R3-W1) We fully agree that multi-center validation is the essential next step and have acknowledged this limitation in our conclusion. Nevertheless, our single-center dataset is among the only ones of its kind with depth metadata and rigorous gold-standard labels. While the test set contains 15 patients, the 2,366 frames provide dense sampling of disease presentations across all four categories.
F. Reproducibility (R1, R2) We have provided a clear and complete description of the model architecture and design in the manuscript. The source code will be publicly released once it is fully organized.
We are grateful for the reviewers’ constructive feedback, which has meaningfully strengthened this work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
There is consensus among the reviewers that there is clear clinical motivation for the work, resulting in a well-motivated and clinically meaningful approach. The authors present well-structured framework, including sequence modelling and the incorporation of depth information with strong supporting analysis. The empirical performance is consistently strong with improvements over baselines and high accuracy.
While several weaknesses were raised, particularly in terms of the marginal contribution of some components, these do not undermine the core contribution of the work. Given the strong empirical evidence and contribution with clinical relevance, the paper warrants early acceptance.