List of Papers Browse by Subject Areas Author List
Abstract
Precise medical image segmentation is vital for modern clinical diagnosis. Integrating multiple modalities, such as vision and text, offers a promising solution. However, two main challenges remain: the scarcity of high-quality image-text pairs and the structural flaws in current visual backbones. Specifically, CNNs lack global context, Transformers are computationally expensive, and Mamba-based models often miss fine-grained local details. We propose SynMamba, a dual-branch architecture. It combines the global modeling of Visual Mamba with the local edge-awareness of CNNs. To solve the data shortage, we introduce a modality synthesis framework. This framework converts masks into hallucination-free clinical narratives to guide the model during training. By embedding clinical semantics directly into the visual backbone, SynMamba excels in both text-guided and fixed-text tasks. Experiments on COVID-CT and COVID-Xray datasets show that SynMamba accurately delineates large overlapping tissues and small isolated lesions. It sets a new state-of-the-art across both unimodal and multimodal benchmarks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3043_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{WeiXua_SynMamba_MICCAI2026,
author = { Wei, Xuan AND Li, Zihan AND Wu, Mengwei AND Zhang, Ziheng AND Wang, Zeyang AND Li, Kaiheng AND Hong, Qingqi},
title = { { SynMamba: Synergizing Visual Mamba with Synthesized Clinical Semantics for Boundary-Aware 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
The authors propose SynMamba, a text-guided dual-branch framework combining Mamba and CNN for medical image segmentation. It addresses data scarcity by generating synthetic clinical narratives to embed semantics into the visual backbone. The method excels in segmenting both large structures and small lesions, achieving superior results in unimodal and multimodal tasks on COVID-CT and COVID-Xray datasets.
- 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.
Clear and well-structured presentation: The paper is well-organized and easy to follow. The motivation, method, and experimental pipeline are presented in a logical and coherent manner, making the contributions easy to understand.
Dual-branch architecture design: The proposed dual-branch SynMamba effectively combines global context modeling (via Mamba) and local detail extraction (via CNN), which is well-suited for medical image segmentation tasks requiring both structural and boundary precision.
Effective multimodal integration via synthetic clinical narratives: The paper introduces a modality synthesis strategy that generates fluent clinical descriptions from masks and images, and couples them with visual features through an interaction bottleneck. This enables effective fusion of text and image modalities, injecting high-level clinical semantics into the visual backbone and improving cross-modal alignment.
- 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.
Lack of clarity and fairness in baseline comparisons In Section 3.2, the experimental setup for competing methods is not clearly described. It is unclear whether these baselines are (i) trained from scratch on the datasets, (ii) fine-tuned using the proposed text generation framework, or (iii) directly evaluated using pretrained weights. This ambiguity raises concerns about the fairness and validity of the comparisons, especially given that the proposed method explicitly benefits from synthesized textual supervision.
Insufficient details on LLM usage and reproducibility concerns Section 2.5 introduces the use of a “large language model (LLM)” for generating clinical narratives, but does not specify which model is used, how it is prompted, or how the rule-based components are designed. The lack of implementation details (e. g. , prompt design, generation constraints) makes the method difficult to reproduce and evaluate.
Inconsistency in dataset scale and insufficient explanation The reported dataset sizes significantly differ from those in cited works [3] and [4]. Specifically, the original datasets contain 423 AP/PA images and 349 CT scans respectively, while the paper reports 2,729 and 9,258 samples. The manuscript does not clearly explain how these numbers are obtained (e. g. , slice-level expansion, preprocessing, or aggregation), which affects reproducibility and makes it difficult to assess the validity of the experimental setup.
Unclear design and questionable necessity of the text synthesis pipeline The proposed modality synthesis pipeline (structured attributes → clinical narratives → structured attributes via back-parsing) raises concerns regarding both clarity and necessity. In Fig. 1, an Attribute Extractor is introduced, yet its role is not described in the Modality Synthesis section. It is unclear why attributes must first be explicitly extracted instead of being directly inferred by the LLM, and how this component interacts with the text generation process. Furthermore, the “Regenerate” loop shown in Fig. 1 is not explained: it is unclear how many iterations are performed, what stopping criteria are used, and how this affects computational cost and training efficiency. More fundamentally, since the pipeline ultimately maps back to structured attributes, the benefit of introducing an intermediate natural language representation is not fully justified. This design appears potentially redundant, and a more direct attribute-based modeling approach may be simpler, more efficient, and less prone to error propagation.
- 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
Potential error propagation and unclear handling of LLM hallucination: Although the paper proposes a “back-parsing mechanism” to mitigate hallucinations, the definition of “logical inconsistency” is not clearly specified. Moreover, if the LLM generates inaccurate clinical descriptions, the back-parsing step may propagate or even amplify errors. The manuscript does not provide sufficient details on how such cases are detected or corrected.
Potential risk of label leakage in text-guided setting The text prompts are generated directly from ground truth masks (e. g. , spatial location, lesion proportion). If such GT-derived text is used during testing, this may introduce label leakage, leading to overly optimistic Dice scores. The paper does not clearly state whether GT-derived information is strictly excluded at inference time.
- 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 addresses an important problem and proposes a reasonably well-designed framework with clear presentation and promising empirical results. In particular, the integration of global and local modeling and the attempt to leverage synthesized clinical narratives are interesting.
However, the overall score is mainly limited by several concerns: (1) insufficient clarity and fairness in experimental settings (especially multimodal baselines), (2) lack of reproducibility due to missing details on LLM usage and dataset construction, and (3) questionable design choices in the modality synthesis pipeline, including potential redundancy and unclear components (e.g., attribute extractor and regeneration mechanism). Additionally, there is a potential risk of label leakage in the text-guided setting.
Overall, while the idea is promising, these issues weaken the technical soundness and clarity of the work.
- 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 #2
- Please describe the contribution of the paper
The main contribution of this work is a multimodal segmentation framework that combines global-local visual modeling with synthesized clinical semantics. The paper introduces a dual-branch VSSM-CNN design, a mask-to-text synthesis strategy for semantic guidance, and a visual-text interaction module for boundary-aware segmentation. The results show that the proposed design is effective in both text-guided and image-only settings.
- 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 is well motivated and addresses a clear problem in medical segmentation. 2.The overall framework is reasonably well designed, with a clear combination of dual-branch visual modeling, synthesized semantic guidance, and boundary-aware supervision. 3.The experimental results are generally convincing, with comparisons against a reasonably broad set of baseline methods that overall support the effectiveness of the proposed method. 4.The ablation study is comprehensive and helps clarify the contribution of the main components.
- 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.While the overall framework is effective, some components build on existing ideas in multimodal segmentation and visual-text fusion, so the technical novelty lies more in the way these elements are combined and adapted to the target problem than in a single fundamentally new formulation. 2.The paper does not evaluate the model with real clinical text, making it difficult to assess the practical value of the synthesized text setting. 3.Although VSSM is motivated as an efficient global modeling tool, the paper does not provide an explicit evaluation of efficiency or computational cost.
- 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.
(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 paper addresses a clear and relevant problem in medical segmentation and presents a well-designed solution. The overall framework is coherent, the experiments are convincing, and the ablation study supports the main design choices. Although the methodological novelty is somewhat limited, the paper is technically sound, clearly written, and shows solid empirical performance.
- 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
This paper proposes SynMamba, a text-informed dual-branch framework integrating a Mamba branch for global context with a lightweight CNN for high-frequency local details. A modality synthesis framework is developed to generate hallucination-free clinical narratives which can provide semantic guidance from ground-truth target masks. SynMamba is evaluated on two COVID dataset and shows superior performance over other common pure vision and vision-language 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.
1.The motivation of combining Mamba with lightweight CNN to overcome the limits and shortcomings in existing models by emphasizing both global vision and local details is clear and well-explained. 2.The text-informed design is useful. The gated visual-text interaction injects clinical priors into visual features which helps sharpening boundary and improving precision.
- 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.Several key equations (e.g., semantic gating map, final feature) are presented without sufficient justification and derivation. Please try to illustrate these equations clearly. 2.In Chapter 2.4, the loss weights are set without providing reasons. Are these values set based on experience, or through several attempts during experiment to achieve the best performance? Briefly explain this can improve clarity. 3.In Chapter 2.5, to eliminate hallucinations, the work implements a back-parsing mechanism, which seems too simplistic to ensure fully “elimination”. And the paper mentions “discard any narratives that exhibit factual errors or logical inconsistencies”. Is this process done manually or automatically using any algorithm? What happens next after this removal process? 4.In qualitative analysis, the paper claims that SynMamba demonstrates superior performance across diverse clinical scenarios. However, it seems that the evaluation scope is narrow since two datasets in this work are both COVID-related images. This raises questions about how well the SynMamba apply to other current clinical practice and lesion areas. The generalization ability or possibility is not discussed in this paper. 5.Some vision-language baselines mentioned (e.g., CLIP, ViLT) are not specifically designed for medical segmentation domain, so they are more likely to underperform on such benchmarks. Therefore, comparing against them does not meaningfully demonstrate the superiority of the proposed method over proper segmentation baselines. 6.The paper mentions “hallucination-free”, which is a very deterministic claim. LLM may still generate unreliable text although some elimination methods are applied. Please consider justifying the strong claim and try to quantify back-parsing rejection rates and test robustness under perturbed text conditions.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
- The current text contains many short and choppy sentences that disrupt coherence. Also, consider varying transition markers—using “first” and “second” twice repeatedly feels somewhat simplistic. More sophisticated or varied phrasing would improve the overall quality and professionalism of the writing. 2.Table 2’s title has missing letter in the word “different”. Similar issue in Fig.3’s title. Please modify this issue. 3.In Page 7, the discussion about Fig. 3 is confusing. Although the figure seems to show the model’s robust performance, how “G degenerates into a uniform pass-through state” shows that the proposed model maintains strong robustness towards text input is not well-explained.
- 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 tackles an important limitation—SSMs lacking local detail and CNNs missing global context—using a dual-branch design and an intuitive visual–text bottleneck. However, several concerns exist: the limited evaluation scope, potential prompt leakage, incomplete reproducibility details, and a lack of statistical reporting. Add more appropriate clarifications and modest additional evidence will be better.
- 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 thank all reviewers for their valuable comments and recognition of our dual-branch design, multimodal integration, and comprehensive ablations. We address the comments below and will revise the paper accordingly.
1.Modality Synthesis and Hallucination Elimination (R1, R3) The Attribute Extractor in Fig. 1 is a fully rule-based module that derives anatomical localization, connected component counts, lesion occupancy ratio, and texture profiling directly from the GT mask. The LLM is used solely to verbalize these structured attributes into a natural-language clinical prompt: Qwen3-VL first rewrites the attributes into a clinical description, and is then queried again to score the consistency between the generated description and the original image–attribute pair; if a mismatch is detected, the rewriting is repeated until a faithful description is produced. Compared with raw attributes, natural-language descriptions better activate the semantic space of pretrained text encoders like, Bio_ClinicalBERT, providing richer linguistic priors for visual-text alignment. We further note that this rewriting task is relatively simple for modern VLMs, since both the structured attributes and the source image are explicitly provided, the probability of a regeneration being triggered is intrinsically low. In our experiments, after the verification step, no faulty or inconsistent samples were observed in the final synthesized dataset, confirming that the Generate-Verify loop is effective in eliminating hallucinations in practice.
2.Comparison Experiments (R1, R2, R3) All baselines were trained from scratch on the same datasets under identical splits. To eliminate any risk of text information leakage, we evaluate our model under a dummy-text input (“Medical image with potential lesion”), which contains no lesion-specific information such as location, size, or severity. The consistently strong performance under this setting demonstrates that the synthesized text serves as a training-time semantic distillation signal rather than a test-time shortcut. Regarding generalization: although both datasets are COVID-related, they cover two distinct imaging modalities (CT and X-ray), and our model achieves state-of-the-art results on both, demonstrating cross-modality robustness. Regarding the absence of real clinical text: since our core motivation is to address the scarcity of paired reports, evaluating on synthesized data is the most faithful reflection of our target scenario. Regarding baseline selection: beyond general-purpose VL models (CLIP, ViLT), our comparison includes medical domain-specific vision-language methods such as ConVIRT and TGANet, ensuring a fair and comprehensive evaluation.
3.Dataset Scale Discrepancy (R1) The reported dataset scale in our paper is correct; the issue lies in an incorrect citation. Specifically, the COVID-19-CT dataset used in our experiments is derived from three public sources, including COVID-19-CT-Seg, MosMedData, and medicalsegmentation.com/covid19.For the X-ray experiments, the dataset is actually QaTa-COV19-v2 rather than the previously cited version. We apologize for this citation mistake and will correct it accordingly.
4.Efficiency Evaluation (R2) The inference latency is roughly 50 ms per image on a single A100. 5.Novelty (R2) We acknowledge that SynMamba builds on prior work; however, the dual-branch backbone and the modality synthesis framework represent novel contributions not previously explored in medical image segmentation.
6.Clarity and Presentation (R3) We thank R3 for the detailed feedback and will address all raised issues, including equation clarity, loss weight justification, the Fig. 3 discussion, typos, and sentence fluency in the revision.
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 study proposes SynMamba, a dual-branch segmentation framework that synergizes Mamba-based global modeling and CNN local feature extraction with synthesized clinical semantics to achieve boundary-aware medical segmentation on CT and X-ray datasets. All three reviewers commend its coherent architecture design, effective vision-text interaction mechanism, complete ablation studies, and competitive quantitative and qualitative segmentation results.