Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurate differentiation between gastric adenoma and carcinoma during endoscopy is critical for clinical decision-making. Yet, this task is highly challenging due to high inter-class similarity and ambiguous boundaries between the two classes. Existing ROI-based classification methods often suffer from detection/segmentation error propagation and loss of surrounding global context. In contrast, full-image classification lacks the necessary spatial focus. Furthermore, we observe that deep neural networks gravitate towards domain-specific texture biases(e.g. bleeding, lighting artifacts), often causing models to predict based on spurious correlations instead of intrinsic morphological features. To address these limitations, we propose a novel framework, Masked Achromatic Guidance Expert (MAGE). During training, we introduce an auxiliary local expert branch trained on masked achromatic views of the neoplasm. By suppressing background context and color, this branch is forced to learn highly discriminative, purely structural features. We then employ a dual-objective distillation strategy, transferring both classification logits and spatial attention maps to provide implicit spatial supervision to the main branch that receives full WLI as input. This dual-objective distillation forces the model to ground its predictions in morphology rather than relying on shortcuts, while still retaining clinically relevant color cues. At inference time, our deployable model operates on images without annotated masks, ensuring real-time deployability . Extensive experiments on a clinical gastric endoscopy dataset show that our method significantly outperforms existing detection-based methodologies (e.g. YOLO) and classification-based methodologies (e.g. Swin-Transformer), providing not only superior classification performance but also interpretable attention maps for clinical reliability.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4527_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{JunJih_MAGE_MICCAI2026,
        author = { Jun, Jiho AND Woo, Jeongwon AND Song, Jaemin AND Nguyen, Thanh Bong AND Yeon, Dong-heon AND Kang, Donghoon AND Park, Jae-Myung AND Ko, Sung-Jea AND Uhm, Kwang-Hyun},
        title = { { MAGE: Color-Invariant and Spatial Knowledge Distillation for Gastric Neoplasm 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

    The paper proposed MAGE (Masked Achromatic Guidance Expert), a framework utilizing a training-time auxiliary expert that learns color-invariant structural cues from a ROI-focused achromatic view (inside-mask grayscale; outside-mask zero). The key contribution was a training-only masked grayscale expert that suppresses color shortcuts, and distills its structural knowledge into a deployable full-image model.

  • 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 major strengths of the paper are: 1.A novel methodology for gastric adenoma from carcinoma in upper gastrointestinal endoscopy, which is clinically important because it directly affects treatment strategy and prognosis in early gastric neoplasia 2.The Masked Achromatic Expert serves as an auxiliary branch utilizing privileged information to learn color invariant features in the neoplasm area. 3.Explainability analysis was very helpful to understand what the trained network uses for the classification.

  • 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 major weaknesses of the paper are: 1.The use only of a private dataset for training raises concerns about the generalization ability, and the biases that may occur . The authors should also use public available datasets for training and testing, in order to ensure fair comparison and reproducibility. 2.There is no information about the datasets use for the training of the compared methods. For instance it would be interesting to provide the performance of Gastronet-5M after it has been fine-tuned in the private dataset.

    1. The PICCOLO dataset seems to be colonoscopy dataset which is different from the dataset domain of trained model 4.Explanation may differ based on the domain of the training datasets .
  • 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

    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 paper proposed an important matter of classification between adenoma and carcinoma. However, the evaluation of the proposed method still needs to be clarified with more experiments .

  • 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 encourage the authors to include the missing implementation details and the augmentation comparison discussion in the camera-ready version to strengthen the paper’s rigor.



Review #2

  • Please describe the contribution of the paper

    This paper introduces MAGE (Masked Achromatic Guidance Expert), a novel training-time framework designed to foster color-invariant and morphology-focused representations for fine-grained gastric neoplasm classification. The authors propose a dual-branch architecture featuring an auxiliary expert trained on masked, achromatic views of lesions. This setup explicitly forces the model to capture discriminative structural features while systematically suppressing background noise and color-based shortcuts. Through a dual-objective distillation process—incorporating both logit-level supervision and spatial attention transfer—the expert’s structural knowledge is effectively distilled into a deployable full-image model. Notably, this design allows the final model to ground its predictions in lesion morphology without requiring segmentation masks at inference, significantly enhancing robustness, interpretability, and clinical reliability compared to conventional detection or classification 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.Principled Formulation to Suppress Color-Based Shortcuts: A key strength of this work is its explicit and well-motivated strategy to mitigate spurious color and texture shortcuts in gastric endoscopy. By introducing a “Masked Achromatic Expert” trained on grayscale, lesion-focused inputs, the framework effectively forces the auxiliary branch to prioritize intrinsic structural and morphological cues. This design directly addresses a pervasive failure mode in medical image classification—where models over-rely on unstable artifacts like bleeding or lighting—thereby enhancing the pathological relevance of the learned features.

    2.Effective Use of Privileged Information via Dual-Objective Distillation: The paper presents a sophisticated application of Knowledge Distillation using privileged information available exclusively at training time. The synergy between logit-level distillation and spatial attention transfer allows the deployable full-image model to inherit both the diagnostic precision and the localization behavior of the expert branch. Notably, this achieves implicit spatial supervision without requiring labor-intensive segmentation masks at inference, representing a technically sound and resource-efficient architectural choice.

    3.Strong Emphasis on Interpretability and Clinical Reliability: Beyond raw performance gains, the authors place significant weight on explainability, which is critical for clinical adoption. The use of quantitative attention alignment metrics (e.g., AiM, PGA, Saliency Dice) alongside qualitative Grad-CAM visualizations convincingly demonstrates that the model consistently attends to clinically relevant lesion regions. This rigorous focus on auditability and evidence-grounded reasoning significantly bolsters the framework’s trustworthiness for high-stakes deployment.

    4.Comprehensive and Methodologically Rigorous Evaluation: The experimental validation is thorough and well-designed. By including comparisons against both detection-based and classification-based baselines, as well as conducting extensive ablation studies and robustness analyses under synthetic color perturbations, the authors provide a holistic view of the model’s capabilities. Furthermore, the inclusion of cross-domain evaluation on external datasets and the reporting of standard deviations over multiple runs add substantial credibility to the empirical findings.

    5.Clinically Meaningful and Practical Deployable Design: The framework is thoughtfully engineered for real-world integration, operating on full White Light Imaging (WLI) frames without necessitating ROI annotations during inference. This clear decoupling of high-overhead training-time supervision from streamlined inference-time execution highlights the practical relevance of the approach, making it highly suitable for real-time clinical decision support systems.

  • 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.Incremental Methodological Novelty Relative to Existing Distillation Frameworks: While the proposed framework is well-engineered, its core components—knowledge distillation, spatial attention transfer, and the use of privileged information—are conceptually rooted in established literature. The logit-based distillation follows the classic formulation by Hinton et al. (2015), and the attention transfer mechanism closely aligns with the approach by Zagoruyko and Komodakis (2017). Although the “Masked Achromatic Expert” is a clever and domain-specific instantiation, the overarching novelty lies more in the effective integration of these known techniques rather than the introduction of fundamentally new learning principles.

    2.High Dependency on Pixel-Level Annotations During Training: The framework’s training phase relies heavily on high-quality lesion segmentation masks to supervise the auxiliary expert. In clinical practice, obtaining precise pixel-level annotations is labor-intensive and often cost-prohibitive, which may limit the scalability and reproducibility of the method across larger, more diverse datasets. While the authors correctly highlight that these masks are not required at inference, the training-time dependency remains a potential bottleneck for institutions with limited access to expert-level manual segmentations.

    3.Limited Scope of External Validation and Domain Diversity: Although the authors provide an external evaluation on the PICCOLO dataset, this cohort is relatively small and originates from a different endoscopic sub-domain (colonoscopy). The absence of a large-scale, multi-center gastric dataset—incorporating various imaging devices, acquisition protocols, and patient populations—weakens the claims regarding broad generalizability. Given that cross-center variability is a known challenge in endoscopic AI (as documented in major journals like Gastroenterology and Gut), a more extensive validation would be necessary to confirm real-world robustness.

    4.Narrow Comparison with State-of-the-Art Architectures: The experimental comparison focuses primarily on detection-based pipelines and standard CNN backbones. However, the study would benefit from a more direct comparison with recent end-to-end transformer-based or attention-driven multimodal models. While the choice of CNNs is justified for spatial alignment, evaluating MAGE against modern vision transformers (ViTs) or hybrid architectures trained under similar constraints would provide a more rigorous benchmark for the proposed distillation strategy.

    5.Lack of Direct Clinical Impact Assessment: While the paper offers strong quantitative metrics and interpretability analyses, its clinical utility remains inferred rather than empirically demonstrated. The study lacks a prospective evaluation, a reader study, or an assessment of how the model’s output actually influences endoscopists’ decision-making processes. Without such human-in-the-loop validation, the practical benefit for clinical workflows—beyond improved offline performance—remains an open question.

  • 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

    This paper presents a well-conceived and methodologically rigorous framework for the fine-grained classification of gastric neoplasms. The proposed MAGE framework effectively addresses a critical challenge in endoscopic AI: the tendency of deep learning models to rely on spurious color and texture shortcuts rather than intrinsic morphological features. The authors’ approach—leveraging a masked achromatic expert to foster color-invariant representations—is both technically sound and clinically intuitive, marking a significant step toward more reliable automated diagnosis.

    The primary merit of this work lies in its thoughtful integration of privileged information through dual-objective distillation. By enforcing spatial attention alignment between the lesion-focused expert and the full-image primary model, the authors achieve implicit spatial supervision without the need for manual masks at inference time. This design, coupled with an extensive evaluation including robustness tests and interpretability analyses, distinguishes the work from more conventional classification pipelines. The emphasis on auditability via quantitative attention metrics is particularly commendable and aligns well with the requirements for high-stakes clinical deployment.

    To further elevate the impact and scalability of this research, the authors might consider exploring ways to reduce the heavy reliance on pixel-level segmentation masks during the training phase. Incorporating weakly supervised or pseudo-labeling strategies could broaden the framework’s applicability to datasets where dense annotations are unavailable. Furthermore, while the current results are promising, a more extensive multi-center validation—encompassing diverse endoscopic devices and acquisition protocols—would be essential to solidify the claims of generalizability. Finally, a direct performance comparison with contemporary end-to-end vision transformers or attention-centric architectures would provide valuable context for the proposed CNN-based distillation strategy.

    Overall, this is a high-quality and carefully executed study that offers a meaningful contribution to the field of medical image analysis. The MAGE framework provides a promising foundation for developing interpretable and robust AI systems for endoscopic diagnosis. With further validation across more diverse clinical settings, this work has the potential to significantly influence how we approach feature decoupling and knowledge distillation in complex medical imaging tasks.

  • 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?

    My recommendation is primarily driven by the paper’s strong alignment between a critical clinical problem and a well-engineered methodological solution. The authors effectively address a pervasive failure mode in endoscopic image analysis—the over-reliance on spurious color and background shortcuts—through a principled training-time framework that explicitly fosters morphology-focused representation learning. The proposed MAGE framework, with its masked achromatic expert and dual-objective distillation strategy, is technically sound, clinically intuitive, and clearly articulated.

    The experimental evaluation is both comprehensive and persuasive. By incorporating comparisons against detection-based and classification-based baselines, performing extensive ablation studies, and conducting robustness analyses under synthetic color perturbations, the authors provide a holistic validation of their approach. Furthermore, the deliberate emphasis on interpretability and localization behavior—quantified through both attention metrics and visualizations—is particularly commendable, as it directly addresses the requirements for clinical reliability and practitioner trust.

    While the methodological novelty is somewhat incremental, as it builds upon established distillation and attention-transfer techniques, its integration within the specific constraints of gastric endoscopy is thoughtful and highly justified. The strengths in problem formulation, experimental rigor, and practical relevance significantly outweigh the identified limitations. Overall, this paper presents a solid and meaningful contribution that merits acceptance independent of the rebuttal process.

  • 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 agree.



Review #3

  • Please describe the contribution of the paper

    The paper introduces MAGE, a training-only dual-branch knowledge distillation framework designed to promote color-invariant, morphology-driven feature learning for gastric neoplasm classification. Its core contributions are: 1.A masked achromatic expert that leverages privileged ROI masks and grayscale inputs to suppress color/background shortcuts and enforce structural feature learning. 2.A dual-objective distillation strategy combining logit-level soft-target transfer and spatial attention map alignment, enabling the deployment-ready main branch (full RGB input) to inherit ROI-centric diagnostic focus without requiring masks at inference. 3.Comprehensive empirical validation demonstrating consistent gains in AUC/F1, improved explanation localization (AiM, PGA, saliency mDice), robustness to synthetic hue perturbations, and cross-domain generalization.

  • 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.

    Strong Clinical Motivation: The authors keenly observe that deep learning models tend to exploit color artifacts (e.g., bleeding, uneven illumination) as shortcuts, thereby overlooking critical morphological features. The strategy of enforcing structural focus through “achromatic” and “masked” inputs aligns perfectly with clinical endoscopic diagnosis principles, which emphasize micro-surface and micro-vascular patterns.

    Practical Deployment Design: While many methods leveraging segmentation masks require additional segmentation steps during inference—increasing computational costs and the risk of error propagation—MAGE ingeniously restricts mask and grayscale processing to the training phase as “privileged information.” During inference, the model operates solely on raw images. This “complex training, simple inference” paradigm holds significant engineering value for real-world applications.

    Robust Evaluation & Explainability: Beyond merely reporting accuracy, the paper provides an in-depth analysis of the underlying mechanisms of the model’s effectiveness. Quantitative Grad-CAM metrics (AiM, PGA, Saliency Dice) demonstrate that the model genuinely focuses on lesion regions rather than the background. Furthermore, the color robustness experiments (Fig. 3) are highly compelling; they directly validate the core hypothesis that reducing color dependency enhances stability—a crucial aspect often missing in similar studies.

    Cross-domain Generalization: Testing on a fundamentally different type of endoscopic data (the colonoscopy PICCOLO dataset) and achieving the highest AUC strongly indicates that the method learns universal morphological features, rather than overfitting to the specific color distributions of a single 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.The core mechanism of using grayscale, masked inputs for the expert branch risks being perceived as a sophisticated form of data augmentation rather than a distinct architectural innovation. A critical question remains unanswered: would simpler, standard photometric augmentations (e.g., random jittering of contrast, brightness, sharpness, and color) applied to the main or expert branch achieve comparable robustness? The paper lacks a direct ablation study comparing the proposed “Achromatic Expert” distillation against a baseline trained with aggressive color/texture augmentations. Furthermore, the experimental setup does not explicitly clarify whether the baseline methods (e.g., Swin-Transformer, YOLO) were trained with similar extensive augmentation strategies. Without this clarification and comparison, it is difficult to attribute the performance gains solely to the proposed distillation framework rather than to differences in regularization or data preprocessing pipelines. 2.The specific value of the loss weight α in Tables 1, 3, and 4 is not explicitly stated. Meanwhile, Table 2 and Figure 3 indicate that α = 0 (pure distillation without direct BCE supervision) yields the highest explainability and color robustness scores, yet the primary performance metrics appear to rely on a different configuration. The paper lacks a rigorous discussion on this critical hyperparameter trade-off: Does over-reliance on the expert (low α) risk inheriting the expert’s limited representational capacity or capping the performance ceiling? Conversely, does a higher α dilute the learned color-invariant features by reintroducing shortcut dependencies through direct label supervision? 3.The primary motivation of MAGE is to prevent models from exploiting spurious color correlations. However, the evaluation omits comparisons with other well-established techniques designed to address exactly this issue, such as Invariant Risk Minimization (IRM), style transfer-based domain generalization, or explicit color-invariant training objectives. Without benchmarking against these methods, it remains unclear whether the added complexity of a dual-branch distillation framework with privileged mask information is strictly necessary. It is possible that simpler augmentation-based regularization or existing de-biasing algorithms could achieve comparable robustness with significantly lower computational and annotation costs. Including these comparisons would better position the proposed method within the broader landscape of robust medical image analysis. 4.The manuscript lacks comprehensive implementation specifications required for independent reproduction. Critical training details—such as lsuch as the distillation temperature, the attention loss weight, the specific optimizer settings (e.g., weight decay, momentum), and the learning rate schedules for both the expert and main branches—are not explicitly reported. Exact data preprocessing/resolution pipelines, and hardware-specific batch accumulation settings—are omitted. Additionally, the paper does not include a code or data availability statement, nor does it commit to releasing the source code, configuration files, or trained checkpoints upon acceptance.

  • 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 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

    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?

    While the paper tackles a clinically meaningful problem with a well-motivated framework and demonstrates promising empirical results in robustness, explainability, and cross-domain generalization, several shortcomings currently prevent it from meeting MICCAI’s thresholds for technical novelty, methodological rigor, and reproducibility. 1.The central claim is that a complex dual-branch distillation framework is required to avoid color shortcuts. However, the paper fails to rule out simpler alternatives. There is no ablation study comparing the proposed method against standard, aggressive photometric data augmentations (e.g., color jitter, brightness/contrast scaling) or established de-biasing techniques like Invariant Risk Minimization (IRM). Without these baselines, it is unclear if the performance gains stem from the novel distillation mechanism or simply from effective regularization that could be achieved with much lower computational and annotation costs. 2.The distillation weight α used for Tables 1, 3, and 4 is not explicitly reported. Moreover, the ablation shows α = 0 yields superior robustness and attention alignment, yet the manuscript lacks a principled discussion on the trade-off between expert-driven morphology learning and direct label supervision. This omission weakens the theoretical grounding of the composite loss. Furthermore, essential training details (distillation temperature τ, attention weight β, specific optimizer settings) are missing. Coupled with the lack of code release, these omissions make it impossible for the community to verify the claims or reproduce the state-of-the-art results. The core idea is clinically intuitive and the evaluation suite is notably thorough, which suggests strong potential. However, the current version requires substantial revision to clarify its novelty, strengthen baseline comparisons, disclose full implementation details. With rigorous addressing of these points, the work could be competitive for future submission.

  • 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.

    While the authors have addressed superficial concerns regarding hyperparameters and implementation details, they failed to provide compelling evidence supporting the core novelty and necessity of their method. 1.The claim that the “Achromatic Expert” is superior to standard data augmentations remains unsupported by direct experimental comparison. The authors merely stated that aggressive augmentations degraded performance in “preliminary experiments” without providing data. Without comparing MAGE against a baseline trained with state-of-the-art regularization/augmentation strategies (e.g., strong color jittering, CutMix, or Mixup), it is impossible to attribute the gains to the proposed architecture rather than simple regularization effects. 2.The author’s abandonment of the IRM method is reasonable, but it neglects other related fields of generalization or debiasing methods (such as style transfer, style-invariant training, contrastive learning), which do not require environmental labels. Given that MAGE requires pixel-level masks (with high annotation costs), the failure to prove its superiority over simpler, mask-free debiasing methods is a significant flaw. 3.The trade-off analysis for α reveals that pure distillation (α=0) yields the best robustness and explainability, yet the main results rely on α=0.25 to achieve peak diagnostic accuracy. This discrepancy suggests that the “expert branch” alone is not sufficiently discriminative for the primary task, undermining the paper’s core motivation that morphological features (learned by the expert) are sufficient for diagnosis. If ground truth labels (α>0) are necessary for optimal performance, the value of the color-invariant expert is significantly diminished, raising questions about the method’s theoretical consistency.



Author Feedback

We sincerely thank all reviewers for their thoughtful feedback. We appreciate the acknowledgement of our work’s clinical relevance. We especially thank R2 for recognizing our clinical reliability. We will gladly incorporate your suggestions in the camera-ready version. [R1 & R3] Dataset Concerns Our objective is fine-grained diagnosis (adenoma vs carcinoma). To the best of our knowledge, there is no public dataset that includes gastric images with adenoma/carcinoma labels along with pixel-level segmentation masks, so using a private dataset was inevitable. Although PICCOLO is a colonoscopy dataset, gastroscopy and colonoscopy share fundamental anatomical morphologies. Testing on PICCOLO serves as a rigorous test of our proposed method’s cross-domain generalization capabilities, as mentioned in our manuscript “We test whether our training methodology …. is transferable beyond the gastric domain.” [R1] Clarification on datasets used for training We clarify that the performance of GastroNet-5M reported in the paper (Table 1) is indeed the result after fine-tuning on our private dataset. Despite fine tuning a 5M dataset pre-trained model, the performance (AUC = 0.744) fell short of MAGE (AUC = 0.786). [R3] Augmentation Concerns MAGE provides a fundamental architectural advantage over aggressive augmentation. Standard photometric augmentations alter the global statistics of the image, but the trained network still processes the full global context. Due to this, it can still learn spurious correlations from background mucosa or glare. In contrast, our “Achromatic Expert” is trained on a strictly constrained input (locally cropped and grayscaled). It is forced to rely solely on the morphological and textural patterns of the lesion itself, decoupled from global signals. Distilling this localized knowledge to the main network actively suppresses the main network’s attention to non-lesion artifacts. During preliminary experiments, we observed that applying color/texture augmentations to the baselines did not improve performance and often degraded it, providing the motivation for an explicit architectural approach. We also clarify that all baselines and our proposed method were trained using the exact same simple data augmentations (random horizontal/vertical flip), ensuring a fair comparison. We will make this clear in the revised manuscript. [R3] Comparison with IRM/Domain Generalization IRM mathematically requires training data to have explicit labels regarding the “environment” of each sample. In endoscopy, confounding artifacts are entangled within single images. Thus, it was infeasible to apply IRM without exhaustive annotations of such artifacts as different environments. [R1, R3] Reproducibility & Hyperparameters We apologize for omitting these details. distillation temperature τ = 4.0, attention weight β = 300 , Input resolution: 224x224 (all experiments) Batch size = 32 Hardware: 1x RTX 3090 Optimizer for PICCOLO experiment: AdamW (default settings) Optimizer for all other experiments: Adam (default settings) Scheduler: Cosine Annealing with max LR = 1e-4, min LR = 1e-6 (all experiments) We commit to publicly releasing the complete source code, configuration files, and pre-trained weights upon acceptance. [R3] α weight clarification and trade-off α = 0.25 was used for all experiments in Tables 1, 3, 4.As R3 correctly identified, there indeed exists a trade-off between raw diagnostic capability and robustness, where a lower α value yields stronger robustness and explainability but slightly lower diagnostic performance. This slight gain from using a nonzero α stems from the additional supervision from the ground truth labels. Yet as seen from Figure 3(b), the difference between raw diagnostic performance of α = 0 and α = 0.5 at 0 degrees hue shift is minimal. We will address this trade-off in the revised manuscript.



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.

    One confident accept by a clearly enthusiastic, and thorough reviewer and two weak rejects. The clinical relevance is acknowledged by all three reviewers, the main concerns stem from a somewhat limited evaluation and shortcomings w.r.t. novelty, rigor and reproducibility. I would like to give the authors n opportunity to alleviate these concerns at least to an acceptable level in a rebuttal round.

  • 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.

    Two post-rebuttal accepts and one reject. While Reviewer one (accept) is rather minimal in his motivation, the other two reviewers provide quite strong and elaborate motivations for their recommendation (one accept, one reject). I have to admit both make compelling arguments, but having to make a call, I propose to accept this paper, as the merits outweigh the rightfully pointed-out limitations.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The rebuttal doesn’t fully address reviewers’ concerns.



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.

    2 Accept (including 5 from R2) vs 1 Reject. Strong clinical motivation for gastric neoplasm vs carcinoma endoscopy. Comprehensive evaluation with good cross-domain generalization. The masked grayscale expert + knowledge distillation approach is novel.



back to top