List of Papers Browse by Subject Areas Author List
Abstract
Fungal keratitis (FK) is a vision-threatening corneal infectious disease. Among its various types, Fusarium and Aspergillus are the most common fungal genera, requiring accurate identification for treatment. In vivo confocal microscopy (IVCM) captures corneal images sequentially along the depth axis, generating structured slice sequences with depth-dependent pathological features. However, existing multiple instance learning (MIL) methods treat slices as independent and identically distributed (IID) instances, neglecting both the local inter-slice correlations produced by continuous scanning and the critical diagnostic indicator of image depth. To address this, we propose a Depth-aware Local Context Enhanced Multiple Instance Learning (DLCE-MIL) framework.
This framework incorporates physical imaging depth by integrating learnable depth embeddings directly with discriminative features extracted by DINOv2. The depth information then conditionally modulates instance representations within the latent space, providing crucial contextual information.
Secondly, motivated by the local dependencies among clustered slices observed during clinical acquisition, we design a Local Instance Context Enhancement (LICE) module. This module effectively captures short-range contextual relationships among neighboring slices in clinician-guided, irregularly acquired IVCM sequences, complementing global MIL aggregation with explicit local context modeling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6013_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/IGLICT/DLCE-MIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MenXia_DLCEMIL_MICCAI2026,
author = { Men, Xiaoyan 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 = { { DLCE-MIL: Depth-aware Local Context Enhanced Multiple Instance Learning for Patient-Level IVCM Fungal Keratitis Classification } },
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 proposes a Multiple Instance Learning (MIL) framework for patient-level classification of fungal keratitis subtypes, specifically Fusarium versus Aspergillus, from In Vivo Confocal Microscopy (IVCM) image sequences. The framework addresses two domain-specific challenges: the local inter-slice continuity inherent in IVCM acquisition and the irregular depth metadata associated with each image. To address the first, the authors propose the Local Inter-slice Context Enhancement (LICE) module, a 1D convolutional block applied over the acquisition sequence before the MIL aggregator. To address the second, they propose a depth-aware gated fusion module that integrates physical acquisition depth as a learnable additive signal into the visual features. A TransMIL backbone is used for patient-level aggregation, with DINOv2-Small as a frozen feature extractor. The framework is evaluated on a proprietary dataset of 127 patients (75 Fusarium, 52 Aspergillus) with 10,673 images acquired using a Heidelberg Retina Tomograph III (HRT-III) with Rostock Corneal Module (RCM), using 5-fold patient-level stratified cross-validation. The best configuration achieves AUC 0.8192 and accuracy 0.7932.
- 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 important and genuinely underexplored problem. Fusarium versus Aspergillus subtype classification from IVCM is a fine-grained task with direct treatment implications, and no prior work has applied a patient-level MIL framework to this specific problem. The dataset of 127 patients with culture-confirmed genus-level labels is the largest reported for this subtyping task, and the experimental design is sound with proper patient-level 5-fold stratified cross-validation preventing data leakage. The two proposed components are well motivated by the clinical acquisition structure of IVCM. The depth-aware fusion directly addresses a gap explicitly identified in prior fungal keratitis classification work as a missing capability. The ablation study in Table 2 systematically evaluates each component individually and in combination, providing clear evidence of the contribution of each module to the final performance. The comparison against both general MIL frameworks and keratitis-specific baselines is comprehensive within the scope of the paper.
- 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 paper’s core motivation is that IVCM acquisition sequences are ordered and prior MIL methods wrongly treat them as unordered sets. Towards Interpretable, Sequential Multiple Instance Learning: An Application to Clinical Imaging (Luo, Wang, and Li, PMC12919601) introduces BiSMIL, a bidirectional Transformer that directly addresses this problem by systematically incorporating the sequential nature of clinical imaging into MIL. The authors compare against ABMIL, CLAM, and TransMIL but do not include BiSMIL. Since the paper’s entire motivation rests on the sequential structure argument, this is a significant omission.
2.Reducing Variability of Multiple Instance Learning Methods for Digital Pathology (Mammadov, Le Folgoc, Hocquet, and Gori, Medical Image Computing and Computer Assisted Intervention MICCAI 2025, Lecture Notes in Computer Science, vol 15966, Springer) documents that MIL methods suffer from performance variability across runs reaching up to 10 to 15 AUC points, stemming from weight initialization, batch ordering, and learning rate. The claimed improvement of DLCE-MIL over TransMIL is only 3.4 AUC points, which falls within this documented variability range. No fixed random seeds, multi-seed averaging, or statistical significance testing is reported, making it impossible to confirm the claimed improvement is significant in real.
3.AI-Based Decision-Support System for Diagnosing Acanthamoeba Keratitis Using In Vivo Confocal Microscopy Images (Lincke, Roth, Macedo, Bergman, Lowe, and Lagali, Translational Vision Science and Technology, 2023) addresses structural challenges of IVCM sequences in the same imaging modality using Heidelberg HRT-RCM, explicitly sorting images into corneal layers before patient-level diagnosis. This paper is neither cited, nor discussed. The authors should revise their related work to accurately scope their novelty claims.
4.DINOv2-Small is used as a frozen feature extractor in all experiments, with no ablation comparing frozen versus fine-tuned configurations, no size variants tested, and no justification provided for this design choice.
5.The dataset contains 75 Fusarium versus 52 Aspergillus patients. With 5-fold cross-validation, each test fold contains approximately 10 Aspergillus patients. Per-class performance breakdown is not reported, no class-imbalance correction is applied, and no statistical significance testing is provided. The overlapping confidence intervals in Table 1 and Table 2 make it difficult to determine whether DLCE-MIL reliably outperforms TransMIL, particularly for the minority Aspergillus class.
6.The gated scalar alpha initialized to 0 is an interesting design choice, but the converged value of alpha after training is not reported. The ablation shows only aggregate metrics with no per-class breakdown showing whether depth fusion specifically helps Fusarium or Aspergillus classification. No perturbation test on depth values at test time is performed, so it cannot be confirmed the model genuinely exploits depth information.
- 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
1.The mathematical notation is inconsistent across sections where feature representations are referred to at different pipeline stages without explicit statements connecting each transformation, requiring the reader to infer the flow between modules.
2.The padding operation in Equation 9 is introduced but never defined, leaving the variable-length sequence handling ambiguous.
3.Table 2 contains an unexplained anomaly where the full model F1 score (0.7494) is lower than the depth-fusion-only configuration (0.7975), which is not adequately addressed in the text.
4.The paper contains no limitations section, which is a notable omission given the private single-institution dataset and small cohort size.
5.Several details critical for reproducibility are missing. The random seed for weight initialization, data splitting, and batch ordering is not reported. The loss function is not explicitly stated. The Lookahead optimizer’s inner hyperparameters, specifically the sync period and slow weights step size, are not provided. The strategy for model selection within training, whether the final epoch or best validation epoch weights are used, is not described. The depth value normalization is not specified, as it is unclear whether raw micrometer values or normalized values are used as input to the depth embedding. The bag size distribution across patients is not reported, and no image quality filtering or artefact exclusion procedure is described for the raw IVCM images.
6.No code or data availability statement is provided.
7.The authors are encouraged to include Towards Interpretable, Sequential Multiple Instance Learning: An Application to Clinical Imaging (Luo, Wang, and Li, PMC12919601) as an additional baseline in Table 1.
8.AI-Based Decision-Support System for Diagnosing Acanthamoeba Keratitis Using In Vivo Confocal Microscopy Images (Lincke, Roth, Macedo, Bergman, Lowe, and Lagali, Translational Vision Science and Technology, 2023) should be cited and discussed in the related work section to properly scope the novelty claims regarding IVCM sequence structure.
9.An ablation comparing frozen versus fine-tuned DINOv2 should be added given the domain gap between natural images and IVCM grayscale corneal imagery. Per-class performance broken down by Fusarium and Aspergillus should be reported in addition to aggregate metrics, as this would allow an assessment of minority class performance and directly validate the biological motivation for depth fusion. The converged value of the learnable scalar alpha should be reported, and a sensitivity analysis perturbing depth values at test time should be added to confirm the model genuinely uses depth information. The unexplained drop in F1 score between the depth-only configuration and the full model in Table 2 should be discussed more thoroughly.
10.The authors are encouraged to rerun each configuration with at least three different seeds and report mean and standard deviation across seeds rather than only across folds. This is particularly important given the known variability of MIL methods documented in Reducing Variability of Multiple Instance Learning Methods for Digital Pathology (Mammadov, Le Folgoc, Hocquet, and Gori, MICCAI 2025).
11.The setup for dataset and its ethical clearnace are not stated.
- 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 paper makes a genuine contribution to a niche but clinically meaningful problem with a well-motivated methodology and a sound experimental setup. The primary concern is that the core quantitative improvement over TransMIL is narrow at 3.4 AUC points, which falls within the documented run-to-run variability of MIL methods. The authors are strongly encouraged to provide multi-seed experiments in the rebuttal, confirming that this margin holds across different random seeds. If the margin collapses under multi-seed evaluation, the recommendation will shift to reject. If confirmed, the remaining concerns regarding missing baselines, reproducibility gaps, and absent ablations are addressable in revision.
- 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.
We thank the authors for their rebuttal. The clarification of the “Squaring” operation as a sequence reshaping step for TransMIL’s PPEG module, the admission and correction of the F1-score transcription error in Table 2, the provision of the converged α value (~0.1), and the addition of the IRB/ethics statement adequately address the core clarification-type concerns raised in our review. The multi-seed evaluation (p=0.0187) and perturbation analysis, while unverifiable at this stage, are encouraging and, if reproducible in the final manuscript, would satisfactorily address our primary statistical concern. However, several straightforwardly addressable points remain unresolved in the rebuttal: the mathematical inconsistency in feature notation across pipeline stages, the undefined padding operation in Equation 9, and the missing reproducibility details (loss function, Lookahead optimizer hyperparameters, and depth value normalization strategy), none of which require new experiments and should have been clarified here. The novelty scoping against Lincke et al. (TVST 2023) is acknowledged but not substantively argued. Contingent on these gaps being properly resolved in revision, we are willing to revise our score to Weak Accept
Review #2
- Please describe the contribution of the paper
The paper proposes DLCE-MIL, a Multiple Instance Learning (MIL) framework for patient-level classification of Fungal Keratitis (FK) subtypes from In Vivo Confocal Microscopy (IVCM) images. The key contributions include the Local Instance Context Enhancement (LICE) module, which uses 1D convolutions to capture correlations between adjacent scan slices, and a Depth-aware Fusion mechanism that integrates physical imaging depth as a learnable embedding to resolve irregularities in clinical acquisition.
- 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.Novel Formulation for IVCM: The authors effectively translate the clinical reality of non-monotonic IVCM acquisition into a structured MIL problem, moving beyond the standard “bag-of-images” assumption to model axial continuity.
2.Integration of Physical Priors: Using learnable depth embeddings to modulate latent representations is a technically sound way to leverage metadata that is often ignored in purely visual models.
3.Strong Clinical Feasibility: The use of a frozen DINOv2 backbone demonstrates a data-efficient approach suitable for medical domains where labeled samples are scarce, while still achieving state-of-the-art performance (79.32% accuracy) on a large-scale clinical dataset.
- 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.Lack of External Validation: The dataset is restricted to only two fungal genera (Fusarium and Aspergillus) from a single-center cohort. Performance is reported solely via 5-fold cross-validation on the internal dataset. Without testing on data from different institutions or different IVCM device models, the model’s robustness to domain shift remains unverified.
2.Superficial Qualitative Evidence: The paper lacks visual interpretability results (such as attention maps or saliency analysis) to confirm that the model is focusing on hyphal morphologies rather than imaging artifacts or background noise.
3.Incremental Nature of LICE: The LICE module is essentially a 1D convolutional layer added to a standard TransMIL architecture. While effective, the technical novelty of applying 1D-Conv for local context is relatively limited in the broader context of sequence modeling.
4.Sensitivity to Depth Noise: The depth fusion mechanism relies on the accuracy of clinician-entered depth values. The paper does not provide a robustness analysis or a “noisy depth” simulation to determine how much performance degrades if the physical depth metadata is imprecise.
- 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.
(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?
Although the DLCE-MIL framework effectively incorporates clinical scanning depth and local context into a patient-level classification task, its scientific rigor is limited by a single-center dataset and a lack of external validation, which leaves its generalizability to different devices or populations unproven. The experimental design is hindered by a reliance on simple probability-averaging baselines and a lack of interpretability studies (e.g., attention maps) to confirm that the model is capturing actual hyphal morphologies. Furthermore, the methodological innovation is somewhat incremental, as the LICE module provides only a basic 1D-convolutional addition to existing MIL architectures without addressing the potential for noise or inaccuracies in the clinician-entered depth metadata.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
The authors propose DLCE-MIL, a multiple instance learning framework designed for the patient-level classification of fungal keratitis from IVCM sequences. The core contributions can be summarized as follows: (1) A Local Instance Context Enhancement module that uses lightweight 1D convolutions to model the local continuity between adjacent IVCM slices, a feature often ignored by standard IID-based MIL methods. (2) A depth-aware fusion mechanism that incorporates absolute physical imaging depths as learnable embeddings to modulate instance features, effectively handling the non-monotonic nature of clinical scanning. (3) Empirical validation on a clinical dataset of 127 patients. By employing a strict patient-level 5-fold cross-validation to prevent data leakage, the model achieves an accuracy of 79.32% and an AUC of 0.8192.
- 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.
Strengths: (1) Strong clinical motivation and problem formulation. Differentiating Fusarium and Aspergillus is a real, practical clinical need for determining targeted antifungal therapy. Framing the analysis of IVCM sequences as a patient-level MIL problem is a natural and effective fit for this specific imaging modality. (2) Architecture tailored to IVCM physics. I appreciate that the authors did not simply plug in an off-the-shelf MIL model, but rather designed the network around the actual acquisition characteristics of IVCM. The LICE module breaks the standard IID assumption in traditional MIL. Using 1D convolutions to capture local tissue continuity between adjacent slices is a computationally lightweight but effective way to model short-range structures like hyphae networks. The depth-aware fusion is a clever approach to handling the non-monotonic nature of clinical scanning. Mapping physical depth to a learnable embedding to conditionally modulate the visual features makes intuitive sense. Additionally, initializing the learnable scalar α\alphaα to 0 is a nice, practical touch that allows the model to stabilize on visual features before incorporating depth constraints. (3) Solid and rigorous evaluation. The experimental setup is methodologically sound. While the dataset size is moderate, the authors correctly implement a strict patient-level 5-fold cross-validation. This is crucial for this type of sequential data to prevent slice-level data leakage, making the reported metrics reliable. Furthermore, the ablation study in table 2 is straightforward and clearly isolates the contributions of the LICE and depth fusion modules.
- 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.
Weaknesses: (1) In Section 2.3, the authors claim that initializing α=0\alpha=0α=0 improves stability under “missing depth values.” However, they explicitly state in Section 3.2 that there are zero missing depth values in their dataset. Since this was never empirically tested, the claim is entirely speculative. The authors should either remove this claim or demonstrate it by artificially masking depth values during training. (2) There is a disconnect between Figure 1 and the mathematical formulations in the text. Figure 1 explicitly illustrates a “Squaring” operation before the depth-aware fusion. While I recognize this is likely the spatial sequence reshaping step required by the TransMIL backbone (specifically to feed 1D sequences into the 2D PPEG module), the authors completely fail to mention or mathematically define this step in Sections 2.2 or 2.3.Readers unfamiliar with TransMIL’s internal mechanics will find this confusing. (3) The manuscript contains glaring errors that detract from the professional quality of the work. For instance, in Figure 1, “Module” is misspelled as “Moudel” in the LICE header. The authors must thoroughly proofread the text and figures. (4) While the strict patient-level 5-fold CV is appreciated, the standard deviations reported in Tables 1 and 2 are quite high (e.g., an F1-Score of 0.7494 ±\pm± 0.0968 for the proposed method). This instability is likely a consequence of the relatively small patient cohort (N=127). The authors need to explicitly acknowledge this high variance and discuss the limitations of their dataset size in the Discussion/Conclusion section. (5) The paper uses a private clinical dataset of 127 patients but completely lacks any statement regarding IRB approval, ethics committee review, or patient informed consent. This is a critical omission for a clinical application paper submitted to MICCAI. (6) No code or dataset link is provided. Given the custom LICE and depth-fusion implementations, providing access to the code is highly recommended to verify the methodology.
- 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?
The authors utilize a private clinical dataset comprising 10,673 IVCM images from 127 patients (Section 3.1). However, the manuscript completely lacks any ethics statement. There is no mention of whether the data collection and study protocol received approval (or exemption) from an Institutional Review Board (IRB) or a local ethics committee, nor is there any mention of patient informed consent or data anonymization procedures. As this involves human subjects and a non-public clinical dataset, providing these ethical declarations is a mandatory requirement for MICCAI. The authors must clarify the ethical approval status of this dataset.
- 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?
I am recommending a Weak Accept, but this score is strictly contingent on the authors’ response during the rebuttal phase. The scientific core of this paper is solid. The clinical motivation to differentiate fungal keratitis subtypes using IVCM is highly relevant, and the architecture is thoughtfully tailored to the physics of the imaging modality. Specifically, adapting the MIL paradigm with 1D convolutions to capture local continuity, alongside a depth-aware fusion mechanism for non-monotonic scanning, is a clever approach. Furthermore, the strict patient-level cross-validation gives me confidence in the reported metrics. However, the manuscript currently feels rushed and suffers from critical compliance and presentation issues that prevent me from giving a higher score. First, the complete absence of an IRB/ethics statement for a private clinical dataset of 127 human subjects is a major red flag that must be resolved. Second, there is a frustrating disconnect between the architecture diagram and the text. Figure 1 shows a “Squaring” operation, which I recognize is likely the spatial sequence reshaping required for the TransMIL PPEG module, but this step is never mathematically defined or mentioned in the text. The figure also contains sloppy typos like “Moudel” instead of “Module”. Finally, claiming robustness to missing depth values is purely speculative, given that the dataset contains zero missing data. I am willing to maintain or even raise my score if the authors provide the mandatory ethical declarations, clarify the mathematical omissions regarding the sequence reshaping, and tone down their unsupported claims. Failure to address these points will result in a firm rejection.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I would like to re-emphasize that the draft seems to have been prepared in haste. Although the revisions resolve my queries, there is uncertainty regarding compliance with the page restrictions.
Author Feedback
We thank the reviewers for their constructive feedback and for recognizing the clinical significance of our work and IVCM-tailored architecture.
CQ1 Statistical significance and variability (R1, R3): To address MIL run-to-run variability, we re-evaluated TransMIL and DLCE-MIL across 3 random seeds. Multi-seed AUC was 0.7682±0.0167 for TransMIL and 0.8135±0.0185 for DLCE-MIL, with significant improvement confirmed by a paired t-test (p=0.0187).
CQ2 Depth/alpha (R1, R2, R3): The learned α moved away from zero and converged to ~0.1, suggesting that depth information was utilized. A noisy-depth perturbation only slightly decreased ACC from 0.7932 to 0.7845, indicating that depth acts as a lightweight auxiliary modulation without causing over-reliance on metadata.
CQ3 “Squaring” (R1, R3): “Squaring” follows the original TransMIL terminology and refers to padding N tokens to S^2 tokens with S=ceil(sqrt(N)), followed by reshaping into a pseudo-spatial S x S grid before PPEG, not element-wise squaring. We will define this in Sec. 2.4 and correct “Moudel” to “Module”.
CQ4 Ethics/code (R1, R2, R3): We apologize for omitting the ethics statement. This retrospective study used anonymized IVCM data and complies with institutional ethical guidelines for clinical data analysis. Privacy restrictions prevent sharing raw images; however, we will open-source the core codebase upon publication, including LICE, depth fusion, and the training pipeline.
R1 F1-score: After rechecking the logs, we found a transcription error in the ablation table: the F1-score of the depth-fusion-only variant was incorrectly entered, creating the misleading impression that the full model had a lower F1-score. We apologize for this oversight and have corrected the table in the revision. The corrected table removes this apparent inconsistency.
R1 related work/baselines: We will discuss the suggested studies, including BiSMIL, MIL run-to-run variability, and IVCM-based clinical decision-support works using corneal layer or sequence information. BiSMIL targets bidirectional sequential MIL and early prediction, whereas DLCE-MIL focuses on depth-aware patient-level IVCM fungal subtype classification. We did not include BiSMIL quantitatively because it targets bidirectional sequential MIL and early prediction, which differs from our patient-level IVCM subtype classification setting, and no suitable official implementation was available for faithful adaptation.
R1 DINOv2: For the backbone concern, we additionally tested frozen DINOv2-Base and fine-tuned DINOv2-Small, which achieved ACC=0.74 and ACC=0.76, respectively; both were below frozen DINOv2-Small, supporting our choice for the limited grayscale IVCM dataset.
R2 external validation: The dataset was acquired using Heidelberg Retina Tomograph III, a mainstream IVCM device. We acknowledge the single-center/single-device limitation and will add it to the Limitations section; multicenter, multi-device validation is future work.
R2 interpretability: We added representative DINOv2 CLS-attention heatmaps for Fusarium and Aspergillus cases, showing that the model mainly attends to corneal structural regions rather than background.
R2 LICE novelty: LICE is not a generic 1D convolution operator, but as task-aware local context modeling tailored to IVCM acquisition. It explicitly captures neighboring-slice continuity before global MIL aggregation, complementing TransMIL and improving fungal subtype recognition.
R3 variance/limitations: We agree that the relatively high standard deviations are related to the small cohort size (N=127). We explicitly acknowledge this variance and dataset-size limitation in the added Limitations section.
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 reviewers recognize the clinical significance of differentiating fungal keratitis subtypes (Fusarium vs. Aspergillus) and appreciate the thoughtfully designed framework tailored to the specific physics of IVCM imaging. However, significant concerns regarding the statistical significance of the results, missing ethical compliance, and incomplete comparisons with existing literature must be addressed.
Major Issues to Address in Rebuttal
Statistical Significance and Variability: Reviewer #1 notes that the 3.4 AUC improvement over TransMIL falls within the known 10–15 point variability of MIL methods. You are strongly encouraged to rerun experiments with at least three different random seeds and report the mean and standard deviation across seeds to prove the improvement is not due to chance.
Methodological Clarifications:
- Explain the “Squaring” operation shown in Figure 1, which is currently missing from the mathematical formulation.
- Discuss the unexplained drop in F1-score in Table 2 when combining all modules compared to the depth-only configuration.
- Report the converged value of the learnable scalar $\alpha$ to confirm the model is effectively utilizing the depth metadata.
Overclaiming Robustness: The claim regarding stability with “missing depth values” is speculative as your dataset had zero missing values. Either provide a “noisy depth” simulation or remove the claim.
Moreover, a reviewer mentioned it has a Code of Ethics issue. Authors should claim if the data collection received approval (or exemption) from an IRB
- 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 authors have clearly clarified the issues raised by the reviewers. The explanations in the rebuttal look reasonable and correct to me.
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.
This paper was initially invited for rebuttal following the primary review period, where reviewers raised several concerns. After carefully evaluating the authors’ rebuttal, the reviewers’ revised feedback, and re-examining the paper, I conclude that the authors have effectively addressed the key concerns raised, and the work now meets MICCAI’s publication standards.
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 reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns the proposed methodology.