List of Papers Browse by Subject Areas Author List
Abstract
Multi-view imaging, such as mammography and chest radiography, is a standard component of clinical practice. However, medical images are often unregistered and contain view-specific artifacts or irrelevant background cues that can obscure diagnostically relevant findings. Many existing methods directly fuse per-view representations, allowing such irrelevant content to contaminate the fused embedding and reducing robustness under varying view configurations. We propose OTCHA, a confidence-aware latent hub token alignment module based on optimal transport (OT) that refines patch tokens before fusion for multi-view classification. OTCHA introduces a set of learnable latent hub tokens shared across views. For each view, we compute an OT plan between patch tokens and hub tokens that jointly considers feature similarity and geometry, and augment the OT formulation with token-conditional dustbins to enable partial matching and discard irrelevant tokens. The resulting transport plan provides token-wise matching confidence, which gates hub-mediated message passing and weights a novel optimal-transport-based representation alignment loss to stabilize refinement. Experiments on three multi-view medical image datasets demonstrate consistent improvements over competing baselines across diverse anatomies and view configurations. Our code is available at https://github.com/labhai/OTCHA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4914_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/labhai/OTCHA
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YanJiw_OTCHA_MICCAI2026,
author = { Yang, Jiwoong AND Chung, Haejun AND Jang, Ikbeom},
title = { { OTCHA: Optimal Transport-Driven Confidence-Aware Latent Hub Alignment for Multi-view Medical Image Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper studies multi-view medical image classification under unregistered conditions and proposes OTCHA, a refinement-before-fusion framework that aligns patch tokens to shared latent hubs via optimal transport with token-conditional partial matching. The transport plan is further used to estimate token-level confidence and guide subsequent feature refinement. The method is technically sound and achieves consistent improvements across multiple datasets. However, it largely builds on existing components such as optimal transport, hub-based representations, and confidence weighting, with limited conceptual novelty. Moreover, key assumptions regarding the effectiveness of token filtering and the quality of cross-view alignment are not thoroughly validated. Overall, the work is solid but moderately incremental.
- 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.Addresses a realistic and important problem in multi-view medical imaging, particularly handling unregistered views and view-specific noise.
2.Proposes a structured refinement-before-fusion framework with a coherent use of optimal transport for token-level alignment.
3.Introduces token-level partial matching and confidence estimation to filter irrelevant information.
4.Demonstrates consistent performance improvements across multiple datasets.
- 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 novelty is limited, as the method mainly integrates existing elements such as optimal transport, latent hub representations, and confidence-aware weighting into a unified framework, rather than introducing a fundamentally new modeling perspective.
2.The effectiveness of token filtering is not directly validated. Although the paper claims that token-conditional dustbins allow irrelevant tokens to be discarded, it does not provide explicit analysis showing that the filtered tokens indeed correspond to background or artifacts.
3.The interpretability claim is only weakly supported. Fig. 2 provides an illustrative hub assignment map, but there is no quantitative or controlled evidence showing that the learned correspondences consistently reflect true anatomical or pathological matching across views.
4.The current evaluation is restricted to fixed-view settings. The paper itself notes this limitation in the conclusion, so the robustness of the method under missing-view or variable-view scenarios remains unclear.
- 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.
(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?
The paper presents a technically sound and practically useful method with consistent improvements on multiple datasets for a realistic multi-view medical imaging problem. However, the novelty is moderate as it mainly integrates existing components, and some key claims such as token filtering and interpretability are not fully validated. The evaluation is also limited to fixed-view settings. Addressing these concerns would further strengthen the paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author’s response have addressed my concerns.
Review #2
- Please describe the contribution of the paper
This paper proposes OTCHA, a refinement-before-fusion module for multi-view medical image classification. The method introduces learnable latent hub tokens shared across views, computes feature-and-geometry fused optimal transport plans between view tokens and hub tokens, uses token-conditional dustbins for partial matching, and derives token-level confidence from the transport plan to gate message passing and weight a new OTRA alignment loss. The method is evaluated on VinDr-Mammo, MURA, and CheXpert.
- 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.
-
Good balance of novelty and practicality. Compared with standard cross-view attention, the use of token-conditional dustbins is particularly appealing because it enables partial matching rather than forcing every token to participate. That directly addresses a real issue in multi-view medical imaging, where background and noncorresponding content often contaminate fusion. The confidence-weighted OTRA loss is also a sensible complement to stabilize the refinement process.
-
Strong empirical evaluation across datasets and anatomies. The method is tested on three public datasets spanning mammography, musculoskeletal radiographs, and chest X-rays, with consistent improvements over a reasonably broad set of baselines. The study-level VinDr-Mammo gains are especially meaningful because the method seems to help most when more views must be integrated.
-
Useful ablations and qualitative analysis. Table 2 does a good job of isolating the contributions of OTCHA, OTRA, geometry, and dustbin design. The qualitative visualization on page 8 also supports the paper’s claim that the method suppresses background regions and improves cross-view localization consistency.
-
- 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.
- Incremental relative to recent cross-view alignment literature. The paper is technically solid, but the novelty is somewhat compositional. It combines known ingredients, OT matching, latent tokens/hubs, partial matching via dustbins, and refinement-before-fusion, in a thoughtful way, but none of these components is individually groundbreaking. The contribution is strongest as an engineering synthesis for multi-view medical imaging rather than a major conceptual leap. This is not fatal, but it does limit enthusiasm.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This is a solid paper with a clear technical contribution, strong experimental validation, and good presentation. While the method is somewhat incremental at the component level, the integration is effective and well matched to the multi-view medical imaging problem. The empirical consistency across three datasets makes the paper convincing. I would support acceptance independent of rebuttal.
- 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
This paper proposes OTCHA, a refinement-before-fusion module for multi-view medical image classification. The key idea is to introduce a set of learnable latent hub tokens shared across views, then use optimal transport between view-specific patch tokens and these hub tokens to refine features before multi-view fusion. The OT formulation combines feature similarity and a geometry-aware prior, and includes token-conditional dustbins for partial matching so that irrelevant or view-specific tokens can be rejected rather than forced into alignment. The transport plan also yields confidence scores that are used to gate message passing and weight an auxiliary alignment loss, called OTRA. The method is evaluated on VinDr-Mammo, MURA, and CheXpert, where it reports consistent AUROC improvements over prior multi-view 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 paper is well motivated. The authors identify a real issue in multi-view medical imaging: views are often unregistered and contain background or view-specific artifacts, so naive fusion can contaminate the shared representation. This motivation is clear and clinically sensible.
- The method is technically coherent. The OT formulation, geometry-aware cost, dustbin mechanism, confidence derivation, and hub-mediated message passing fit together cleanly as one design rather than feeling like unrelated add-ons. The paper is also clearly written, and the method section is easy to follow.
- 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 comparison against XFMamba is not fully convincing because the reported XFMamba performance in Table 1 is substantially lower than the performance reported in the original XFMamba paper. For example, OTCHA reports XFMamba at 0.889 on MURA and 0.906 on CheXpert, whereas the original XFMamba paper reports notably higher AUROC values on these datasets. This discrepancy may be due to differences in preprocessing, splits, backbone variants, or retraining protocol, but the paper does not clearly explain the source of the gap. As a result, it is difficult to determine whether the comparison is fair and whether the claimed improvement over XFMamba is truly meaningful.
- The absolute gains over the strongest baselines are modest. On some datasets the margins are small, for example 0.906 to 0.908 on CheXpert and 0.889 to 0.895 on MURA. This makes the contribution feel somewhat incremental empirically, especially given the added OT machinery and auxiliary loss.
- Computational cost is not fully analyzed. The paper says the parameter overhead is marginal, but OT is run with Sinkhorn and then recomputed after hub update, which adds nontrivial compute beyond parameter count alone. There is no runtime, memory, or training/inference cost comparison, so “efficient” is not fully demonstrated.
-
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper is well motivated and technically solid. The proposed OTCHA module is a coherent refinement-before-fusion design for multi-view medical image classification, and the idea of using optimal transport with token-conditional dustbins and a confidence-weighted alignment loss is interesting. However, I scored the paper slightly below the acceptance threshold for several reasons.
- First, although the gains are consistent, they are relatively modest against the strongest baselines, which makes the empirical impact feel somewhat incremental given the added methodological complexity.
- Second, the efficiency tradeoff is not well analyzed. While the parameter overhead is small, the method relies on Sinkhorn-based OT and recomputes transport during refinement, which likely adds nontrivial runtime and memory cost, yet no computational analysis is provided.
- I also have some concern about the fairness and transparency of the baseline comparisons. In particular, some reported baseline numbers, such as XFMamba, appear lower than those reported in the original baseline paper, but the manuscript does not clearly explain whether all baselines were reimplemented and retrained under a unified setup or why these discrepancies arise. This makes the strength of the comparison harder to assess.
- 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.
After reading the rebuttal, I am satisfied that the authors have adequately addressed my main concerns. Therefore, I recommend acceptance.
Author Feedback
We thank all reviewers for recognizing OTCHA’s coherence and practical value, and address the main concerns below. (R1, R2) Contribution: OTCHA introduces a pre-fusion reliability modeling perspective for unregistered multi-view medical image classification. Unlike methods that directly aggregate dense view tokens, OTCHA recasts fusion as reliability-guided token participation through a shared latent hub. This reframes unregistered multi-view fusion from dense aggregation to reliability-guided refinement, which our ablations (Tab. 2) confirm are not reducible to any single component. (R3) Baseline comparison: We agree this needs clarification. Tab. 1 reports benchmark results under a unified evaluation protocol: all baselines were retrained with the same patient-level splits, sample construction, label definitions, preprocessing, and 5-run evaluation, using official model implementations where available. Since some patients have multiple studies, we split by patient before sample construction to prevent data leakage among studies from the same patient. The discrepancy from XFMamba’s paper reflects different split/sample-construction protocols. For MURA, we sampled 2 views once per study (11056/1270/1098); XFMamba’s official code uses pair-level sampling, so studies with >2 views yield multiple pairs (35185/3158/3283). For CheXpert, we constructed samples after patient-level splitting (23577/3923/3913); XFMamba uses random sample-level splitting (23559/3926/3928). We will clarify this in revision. (R3) Computational cost: Due to limited space, the computational cost was insufficiently discussed. We measured per-sample latency and peak memory under the paper settings (inference/training). Relative to MV-HFMD (base; 2-view 1.08/3.47ms, 0.96/6.64GB; 4-view 2.30/7.19ms, 1.63/12.93GB), OTCHA requires 1.65/4.75ms, 0.97/6.85GB (2-view) and 3.42/9.22ms, 1.65/13.34GB (4-view), respectively. Thus, our module adds only 0.57-2.03ms in absolute latency and 1-3% peak memory, suggesting limited added overhead. We will revise the efficiency statement. (R3) Modest gain: While absolute gains on MURA/CheXpert are modest, they are statistically significant over XFMamba (Tab. 1), and OTCHA achieves the best AUROC in all settings. Notably, the gain is most pronounced in the challenging 4-view VinDr-Mammo study-level, where OTCHA significantly outperforms MV-HFMD. With the cost analysis above and the ablations in Tab. 2, this indicates a favorable accuracy-complexity trade-off. (R1) Effectiveness of token filtering: On the VinDr-Mammo test set, we analyzed inference-time OT assignments. Each token’s unit mass is split between hubs and a dustbin, so dustbin mass serves as a rejection score. Treating zero-intensity patches as background (11.9% of all 196K tokens), dustbin mass alone detected background at AUROC 0.781, and the top-20% highest-dustbin tokens are 2.37× enriched in background over random. This supports preferentially filtering background/low-information tokens. (R1) Interpretability: We quantitatively evaluated pathological correspondence on VinDr-Mammo cases with lesion bounding boxes in both paired views. For each lesion patch in one view, we ranked paired-view patches by hub-assignment connectivity and computed AUROC using lesion/non-lesion patches as positive/negative. Same-patient pairs achieved 0.642 on average; patient-level shuffled pairs, where no true lesion correspondence exists, yielded to 0.522 (Wilcoxon p<1e-9). This provides supporting evidence for preferential lesion-level cross-view alignment. (R1, AC) Missing/variable-view: We agree this needs clarification. “Fixed-view” refers to the benchmark protocol, not an architectural constraint; OTCHA supports variable-view inputs (Tab. 1). In a controlled missing-view scenario analysis, OTCHA remained functional under partially available views, although performance showed a modest decrease compared with the full 4-view setting; broader missing-view evaluation remains future work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper addresses a realistic multi-view medical image classification problem and proposes a technically coherent refinement-before-fusion framework based on optimal transport, latent hubs, token-conditional dustbins, and confidence-aware feature refinement. The reviews are close to the acceptance boundary, and the reviewers generally acknowledge the motivation, clarity, and consistent improvements across VinDr-Mammo, MURA, and CheXpert. However, several points need clarification before a final decision, especially the fairness of baseline comparisons, including the lower reported XFMamba results compared with the original paper, and whether all baselines were retrained under the same protocol. The authors should also provide runtime, memory, and inference-cost analysis, since Sinkhorn-based OT may add nontrivial computational overhead beyond parameter count. The rebuttal should further address the modest gains over strong baselines, directly validate token filtering and hub alignment, and discuss robustness under missing-view or variable-view settings.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal adequately clarifies the baseline protocol, computational overhead, and the intended contribution of reliability-guided token participation before fusion. Although the method remains moderately incremental and some diagnostic analyses are provided only at rebuttal stage, the consistent post-rebuttal reviewer support and favorable accuracy–complexity trade-off make the paper acceptable as a weak accept.
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.
The authors have provided solid and detailed clarifications in their rebuttal, satisfactorily addressing the core empirical and technical questions raised during the review process. Specifically, the authors have properly justified the discrepancies regarding the reported XFMamba performance. Crucially, they clarified that all baseline models were evaluated under a strictly unified protocol—employing identical patient-level splits, sample construction strategies, label definitions, and preprocessing pipelines, alongside 5-run repeated evaluations—while utilizing official code repositories for all publicly available baselines to guarantee an absolutely unbiased comparison. Furthermore, the authors delivered a thorough quantitative analysis detailing the inference and training latency per single sample as well as the peak memory footprint, effectively mitigating concerns regarding the computational tax imposed by the Sinkhorn-based optimal transport mechanism. Finally, the rebuttal provided strong empirical evidence directly validating the token filtering and hub alignment modules, while demonstrating the framework’s robust adaptability to scenarios involving missing or variable view counts. While the reviewers’ formal ratings remained unchanged post-rebuttal, the AC independently assesses that the authors’ technical explanations and rigorous cross-validation have fully resolved the primary skepticism, successfully elevating the manuscript to the publication threshold. Therefore, the AC recommends the acceptance of this paper.
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 rebuttal has addressed the main concerns from the reviewers and all the reviewers are supportive to accept this paper.
