List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation of infected regions in radiological imagery is fundamental for the diagnosis and treatment of pulmonary infectious diseases. While recent advancements in multi-modal learning have introduced clinical reports as textual guidance for medical image segmentation, existing early- and late-fusion frameworks predominantly treat textual input as a monolithic global feature. Such approaches neglect the distinctive semantic roles of individual text tokens–specifically infection type, quantity, and location–resulting in diffuse, non-specific guidance. This lack of granularity leads to feature redundancy and suboptimal performance in complex clinical scenarios where diverse lesion morphologies require precise grounding. To address these limitations, we propose a novel Role-Conditioned Alignmnet model (RCA) for medical image referring segmentation. RCA effectively integrates text-guided cues into visual features by explicitly exploiting the unique functional characteristics of different text tokens to enhance cross-modal feature alignment. Furthermore, we introduce a Boundary Refinement Module (BRM) that leverages edge-aware information to precisely delineate fuzzy boundaries where lesions overlap with healthy tissue. Extensive experiments conducted on the QaTa-COV19 and MosMedData+ benchmarks demonstrate the effectiveness of our proposed method, which achieves state-of-the-art results and significantly outperforms current unimodal and multimodal approaches. The code is publicly available at https://github.com/lik1996/RCA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2035_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lik1996/RCA
Link to the Dataset(s)
QaTa-COV19: https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset
MosMedData+: https://medicalsegmentation.com/covid19/
BibTex
@InProceedings{LiKun_Learning_MICCAI2026,
author = { Li, Kun AND Wang, Fu AND Zhou, Feixiang AND Zhao, He AND Xu, Xiaowei AND Meng, Yanda AND Zhao, Yitian AND Zheng, Yalin},
title = { { Learning Role-Conditioned Alignment for Medical Image Referring Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes RCA, a role-conditioned alignment framework for medical image referring segmentation that explicitly decomposes clinical text into three functional roles—lesion semantics, quantity, and anatomical location—and aligns them with visual features via a dedicated Role-Conditioned Block (RCB). The RCB combines a frequency-enhanced upsampling module (FEUP), three parallel role-specific interaction mechanisms (energy-based semantic alignment, quantity-gated lesion queries, and coordinate-scored location attention), and a fusion step; a simple edge-aware Boundary Refinement Module (BRM) further sharpens lesion borders. Experiments on QaTa-COV19 and MosMedData+ show consistent improvements over unimodal and recent multimodal baselines, yielding new SOTA Dice and mIoU scores with ablations attributing gains to each role and the BRM.
- 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, consistent gains across two public MIRS benchmarks with comprehensive baselines (unimodal and multimodal), including recent state-of-the-art approaches.
- Overall pipeline, module motivations, and ablation design are clearly communicated with an informative figure and structured method description.
- Addresses a central limitation in MIRS-lack of token-level, role-aware grounding-leading to improved spatial disambiguation in complex pulmonary infection cases.
- 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 method relies critically on partitioning tokens into roles, but the paper does not describe how this role assignment is performed (rule-based parsing, NER, a learned classifier, or dataset-provided structure?).
-
The BRM uses a Sobel operator to extract edges and modulate features. FEUP already uses a Laplacian high-pass filter for a similar purpose. Both are edge-enhancement mechanisms applied at different stages. The paper doesn’t discuss this overlap or justify why both are needed
-
No analysis of robustness to noisy/incomplete/misaligned text, which is realistic in clinical workflows; performance under partial or erroneous role tokens is unknown.
-
The paper claims the three mechanisms are tailored to their respective roles, but provides no visual or interventional evidence that the attention actually specializes by role. A figure showing role-specific attention maps across several test images, ideally paired with a text-component swap experiment, would substantially strengthen this central claim.
-
The global textual gate (Eq. 3) — global average pooling over text, followed by linear projection, sigmoid, and channel-wise multiplication with visual features — is functionally identical to Squeeze-and-Excitation channel attention [Hu et al. , 2018] or FiLM-style feature-wise modulation [Perez et al. , 2018], simply conditioned on text rather than self-derived features. The paper presents this as a novel “textual gate” without citing or acknowledging this well-established lineage.
-
Minor writing issues. “Alignmnet” typo in the abstract. “the the” duplication on page 5.
-
- 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.
(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 proposes a sensible idea — role-partitioned cross-modal attention for MIRS — and supports it with clean ablations and strong results on two benchmarks, with a backbone ablation that strengthens attribution of gains to the method itself. However, the central contribution is under-validated in important ways: the role-partitioning procedure is not operationally specified, the global textual gate is standard SE/FiLM-style conditioning presented as novel without citation, and the specific mechanism–role pairings are not validated by a swap experiment or attention visualization. Table 2 further shows that location dominates, which is consistent with but does not uniquely support the paper’s “synergy of roles” narrative. Additional issues include overlap between FEUP and BRM without justification, and no robustness analysis despite the method’s dependence on correct role partitioning.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper argues that in medical image referring segmentation, clinical text should not be treated as a single global prompt. Instead, it proposes a Role-Conditioned Alignment (RCA) framework that explicitly decomposes textual information into semantic, quantity, and location roles, and aligns each role with visual features through dedicated interaction mechanisms. It also designs a boundary refinement module(BRM) that integrates edge-aware priors to enhance structural delineation in complex morphological scenarios. The approach achieves SOTA results on QaTa-COV19 and MosMedData+, outperforming prior multimodal MIRS baselines including FMISeg.
- 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.Clear presentation. The paper is fairly easy to follow. The motivation, framework, and experiments are clearly organized, and the qualitative results are aligned with the intended effect of the method, namely cleaner predictions and improved boundary quality.
2.Well-motivated problem formulation. The paper identifies a meaningful limitation of prior MIRS methods, namely that most of them treat the input report as a monolithic global feature and do not explicitly distinguish between different token roles such as lesion semantics, quantity, and anatomical location. This is a sensible observation in the medical setting, where these cues are often naturally structured in clinical descriptions. The proposed RCA framework is reasonably coherent and well aligned with the task setting.
3.Relatively solid ablation study analysis. The paper provides a reasonably informative ablation study by evaluating FEUP, each role-conditioned branch, their fusion, and BRM separately, which makes it easier to understand where the improvement comes from. The results support the central claim that semantic, quantity, and location cues are complementary, with the location branch contributing the largest single gain. In addition, the paper compares different visual and language backbone combinations, which suggests that the proposed design is not entirely tied to one specific encoder pair.
- 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 method is tailored to the current benchmark setting, where referring expressions are relatively structured and naturally decomposable into semantic, quantity, and location cues. Although the paper shows robustness across backbone choices, it does not provide evidence of broader generalization across organs, diseases, datasets, or more diverse report styles. As a result, the external validity of the proposed role-conditioned decomposition remains unclear.
2.The presentation of Figure 1 could be more professional and academically polished. In particular, the use of decorative icon-style elements (e.g., the coin stack and map marker) makes the architecture diagram look somewhat informal and less consistent with the visual style typically expected in computer vision or medical imaging papers, it’s more like a NLP-paper-style figure. If such visual symbols are intentionally retained, their technical meaning should be explicitly explained in the figure caption or main text.
3.A minor weakness is that the experimental comparison is not fully comprehensive. The paper’s “state-of-the-art” claim may be incomplete. Given that both MMI-UNet[1] and MIT-SAM[2] are open-source and highly relevant methods on QaTa-COV19 and MosMedData+, it would strengthen the paper to include direct comparisons with them. The Dice of MIT-SAM is 82.08%, higher than this paper. This would make the empirical evaluation more complete and help better support the claimed state-of-the-art position.
[1] Bui, P.N., Le, D.T., Choo, H.: Visual-textual matching attention for lesion segmentation in chest images. In: Linguraru, M.G., Dou, Q., Feragen, A., Giannarou, S., Glocker, B., Lekadir, K., Schnabel, J.A. (eds.) Medical Image Computing and Computer Assisted Intervention – MICCAI 2024.pp. 702–711.Springer Nature Switzerland, Cham (2024) [2] Zhou X, Yan L, Ding R, et al. MIT-SAM: Medical image-text SAM with mutually enhanced heterogeneous features fusion for medical image segmentation[J]. IEEE Journal of Biomedical and Health Informatics, 2025.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 gave this paper a weak accept recommendation. The main reasons are that the paper is clear, well motivated, and empirically reasonably strong. I think the paper identifies a sensible problem in prior medical image referring segmentation methods, namely the lack of explicit modeling of different textual roles, and proposes a coherent framework to address this. I also appreciated that the paper is easy to follow and includes a reasonably informative ablation study, which helps support the claimed contribution. But I would still encourage the authors to address the above concerns during the rebuttal stage, particularly regarding broader generalization, the meaning of Fig.1 icons, and comparisons with recent open-source baselines.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper introduces the Role-Conditioned Alignment (RCA) framework, which represents a paradigm shift from traditional Medical Image Referring Segmentation (MIRS) methods that treat text as a monolithic global feature. Specifically, the RCA framework partitions clinical text into three functional subsets—lesion semantics, quantity, and anatomical location—and implements tailored interaction mechanisms for each. Furthermore, a Boundary Refinement Module (BRM) is designed to enhance segmentation precision in regions where lesions and healthy tissues overlap.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper introduces the RCA framework, which breaks through the limitations of traditional MIRS mechanisms that treat text as a single global feature. By partitioning the text into three functional subsets and designing distinct interaction mechanisms for each, it achieves a more fine-grained cross-modal feature alignment. 2.A Frequency-Enhanced Upsampling (FEUP) strategy is introduced, which effectively overcomes the problem of high-frequency information loss commonly caused by standard upsampling processes. 3.The Boundary Refinement Module (BRM) is incorporated into the architecture, which effectively leverages edge-aware information to precisely delineate fuzzy boundaries in areas where lesions overlap with healthy tissues. 4.The paper provides a comprehensive comparison with various existing unimodal (e.g., UNet++) and multimodal methods (e.g., LviT). The experimental results demonstrate that the proposed RCA method significantly improves lesion prediction using textual descriptions and effectively addresses spatial ambiguity in X-ray and CT imaging.
- 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 lacks technical details regarding text processing. Although it states that the input text is partitioned into three functional subsets, it fails to provide a clear explanation or algorithm for how the raw clinical reports are decomposed and mapped into these specific categories. 2.There is a notable absence of experimental evidence concerning the computational overhead. The Role-Conditioned Block (RCB) introduces three parallel attention mechanisms, yet the paper does not discuss or provide data on whether this increased complexity significantly impacts inference latency, GPU memory footprint, or overall running costs in practical applications.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper introduces the RCA framework, significantly enhancing the performance of traditional MIRS by achieving fine-grained cross-modal alignment. Additionally, the incorporation of frequency-enhanced up-sampling effectively addresses the loss of high-frequency information inherent in conventional up-sampling methods. It would be better if the paper analyzes the practical computational overhead of the framework.
- 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
Author Feedback
We sincerely thank the reviewers for their time and valuable feedback. The positive responses greatly encourage the medical multimodal research community. We will address the key comments, incorporate necessary revisions, and correct all typos in the camera-ready version.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This submission proposes RCA, a role-conditioned alignment framework for medical image referring segmentation that decomposes clinical text into lesion semantics, quantity, and anatomical location, and reports strong results on QaTa-COV19 and MosMedData+ with generally clear presentation and useful ablations. The reviewers are broadly positive about the motivation, empirical gains, and role-aware design, supporting a provisional accept recommendation. However, authors are encouraged to address several issues in the rebuttal and final version: most importantly, the paper does not specify how raw text is partitioned into the three roles, the claimed role-specific mechanisms are not directly validated through attention visualizations or intervention/swap experiments, and the state-of-the-art claim may be incomplete without comparisons to relevant recent baselines such as MMI-UNet and MIT-SAM.
