List of Papers Browse by Subject Areas Author List
Abstract
Semi-supervised medical image segmentation has emerged as a dominant research problem in medical image analysis, mitigating annotation scarcity by leveraging consistency regularization on unlabeled data. However, existing approaches operate predominantly via visual pattern matching, relying heavily on pixel-level similarities. This visual-centric dependency often falters in clinical scenarios characterized by the visual-semantic mismatch, where visually similar lesions warrant distinct diagnostic conclusions, thus failing to capture the underlying diagnostic logic used by experts. To address this, we move beyond visual cues and propose CERS (CoT-Enhanced Reasoning Segmentation), a framework that integrates Chain-of-Thought (CoT) reasoning to distinguish pathologically distinct cases. Specifically, we construct a knowledge pool enriched with linguistic reasoning descriptions generated by large language models (LLMs). A semantic-aware reference selection strategy is introduced to identify historical evidence, filtering candidates first by morphology, and then refining them via CoT consistency to eliminate hard negatives. Furthermore, a multi-scale coordinate attention module (MCAM) is designed to effectively fuse this reasoning-derived context into the decoding process. Extensive experiments demonstrate the superiority of CERS against state-of-the-art approaches, particularly in resolving boundary ambiguities and semantic inconsistencies. The code is available at https://github.com/cymasuna/CERS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1650_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/cymasuna/CERS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheYum_Beyond_MICCAI2026,
author = { Chen, Yuming AND Xie, Yuxin AND Zhou, Tao AND Zhou, Yi},
title = { { Beyond Visual Cues: CoT-Enhanced Reasoning for Semi-supervised Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.This paper proposes the CoT-Enhanced Reasoning Segmentation (CERS) framework, introducing Chain of Thought (CoT) reasoning into semi-supervised medical image segmentation, aiming to address the visual semantic mismatch problem caused by traditional methods relying solely on pixel-level “visual pattern matching” 2.The designed multi-scale coordinate attention module (MCAM) can model feature responses along both horizontal and vertical directions, which helps protect complex boundary structures during the decoding process without introducing excessive parameter burdens.
- 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 author proposed the CERS framework, which was the first to incorporate the Chain-of-Thought (CoT) reasoning generated by large language models into the segmentation task. This design aims to address the “visual-semantic mismatch” problem in clinical settings. By introducing diagnostic logic, the model no longer merely performs “shape matching”, but rather mimics the “clinical judgment” of doctors. 2.The author utilized LLM to extract multi-dimensional reasoning descriptions regarding texture, boundary, context, and summary. These structured texts were transformed into a “Knowledge Pool”, serving as external evidence in semi-supervised learning. This approach makes the implicit expert knowledge in labeled data explicit, enabling unlabeled data to be assisted in segmentation through the retrieval of logically consistent cases. 3.Extensive tests were conducted on three different modalities of public datasets (CT, X-ray, MRI), and the results showed that its performance was superior to the existing SOTA methods under different annotation ratios.
- 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 same experimental result is presented differently in different tables. The IoU of the author’s method on the QaTa-COV19 (50%) dataset is marked as 71.42 in Table 1, but it changes to 66.87 in Table 2.In Table 2, 66.87 is not the best result, yet it is bold.
-
The framework’s success is heavily tied to the quality of the CoT descriptions generated by the frozen GPT-5.2 model. While a manual review of 100 cases showing high clinical accuracy, the paper lacks a quantitative sensitivity analysis demonstrating. It is indistinct that how the model behaves when the LLM generates noisy or incorrect reasoning.
-
The ablation study for the decoder compares the proposed MCAM against standard U-Net and Swin U-Net decoders. If an internal ablation analysis of MCAM could be conducted, it would be of greater reference value.
4.The header “strateges” in Table 2 should be changed to “strategies”.
-
- 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 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.
(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 method is innovative.The author proposed the CERS framework, which was the first to incorporate the Chain-of-Thought (CoT) reasoning generated by large language models into the segmentation task. The logical expression of the article is quite good. However, there is a problem in the experimental part that the same data varies greatly across tables.
- 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
This paper proposes CERS, a CoT-enhanced framework for semi-supervised medical image segmentation. The method augments conventional visual learning with LLM-generated segmentation-aware reasoning, constructs a labeled knowledge pool, performs coarse-to-fine image–CoT retrieval, and fuses the retrieved semantic cues through the proposed MCAM and a dual-decoder consistency scheme. The main contribution is to show that reasoning-aware retrieval can help semi-supervised segmentation move beyond purely visual matching, especially in cases with boundary ambiguity or visual-semantic mismatch.
- 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 addresses a meaningful limitation of existing semi-supervised segmentation methods, namely their reliance on pixel-level visual similarity in clinically ambiguous cases. 2.The proposed pipeline is technically coherent: CoT generation, labeled knowledge pool construction, two-stage retrieval, and reasoning-aware fusion are well connected and clearly motivated. 3.The empirical evaluation is strong overall, with consistent gains across three public datasets, two labeled-data ratios, qualitative comparisons, and informative ablations on retrieval design and decoder choice.
- 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 validation of the LLM-generated CoTs is still limited: the paper reports expert review on 100 sampled cases with >95% clinical accuracy, but does not provide enough detail on sampling across datasets, reviewer protocol, or agreement statistics. 2.The provenance and quality of the auxiliary text are not fully characterized. The manuscript uses different text sources across datasets, which makes it difficult to assess how robust the method is when metadata are sparse or noisy.
- 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.
(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 recommend weak acceptance because the paper tackles an important limitation in semi-supervised medical segmentation, proposes a coherent reasoning-aware framework, and demonstrates strong empirical performance across multiple datasets and label ratios. The method is well motivated, technically sound, and supported by useful ablations. My main reservations concern the limited validation details for the generated CoTs and the heterogeneity of the text sources, but these are clarification issues rather than major weaknesses.
- 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 authors have fully addressed all my core concerns with concrete evidence: they provided complete details of the two-radiologist CoT validation protocol and added a rigorous robustness experiment demonstrating strong resilience to sparse auxiliary text. All remaining issues are minor editorial fixes. I recommend acceptance.
Review #3
- Please describe the contribution of the paper
A semi-supervised medical image segmentation method was designed with the aid of CoT.
- 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.
- Three public datasets are used for experiments.
- The figures are tidy and easy to understand.
- 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.
- Unclear motivation. The authors said “a critical challenge remains that clinical scenarios often exhibit a visual-semantic mismatch, where images with high visual similarity warrant distinct diagnostic conclusions”. However, Segmentation is primarily a localization task, which focuses on identifying the spatial extent of a region. It does not directly equate to a clinical diagnosis.
- Insufficient novelty. This work seems to merely integrate CoT into standard frameworks for image-text alignment, retrieval, and image segmentation. The technical contribution appears marginal as it lacks significant algorithmic innovation.
- The number of references cited in the description of the current status and research significance is somewhat insufficient.
- DuCiSC and LeFeD were proposed in 2025 and 2024, respectively. However, why did their performance lag behind UCMT proposed in 2023? Are there any mistakes with the exerimental results?
- 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 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.
(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?
I have concerns regarding the novelty, motivation, and some experimental results of this paper, but I would like to prefer an opportunity for explanation.
- 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.
My concerns have been dealt with properly.
Author Feedback
We thank all reviewers for their constructive suggestions.
To R1: Q1&4: Typos A1: The value 66.87 in Table 2 should be 71.42, consistent with Table 1.It was a table transcription error. We will check all the typos in the revision.
Q2: Sensitivity analysis of CoT A2: To address your concern, we randomly replaced 30% of the words in the CoT with “
" to simulate noisy reasoning (Dice -1.57%). Moreover, we substituted GPT-5.2 with Qwen2.5-VL 72B to generate lower-quality CoTs (Dice -1.14%). CERS degrades gracefully under noisy reasoning because retrieval is jointly constrained by both visual morphology and CoT consistency. The image-based coarse retrieval still provides stable structural priors even when textual reasoning becomes imperfect. Q3: Ablation Study of MCAM A3: We replaced sub-components with 1×1 convolutions. Removing Coordinate Attention caused a 1.52% Dice drop, and removing the Multi-scale Branch caused 0.95%, confirming both components’ effectiveness.
To R2: Q1: Experts reviews on CoT A1: The 100 reviewed cases were randomly selected across datasets, evaluated by two board-certified radiologists (>10 years of experience). A CoT was counted as incorrect only when both reviewers confirmed, resulting in 4 errors out of 100 cases. Full CoTs and texts for all datasets are now available at the anonymous link.
Q2: Robustness to weak metadata A2: For MosMedData+ and QaTa-COV19, we adopt text annotations from LViT describing lesion type, number, and distribution. These annotations have been consistently used and validated in subsequent works [ARSeg, MICCAI; TeViA, TMI]. For BRISC 2025, we generate structured descriptions from metadata and graphical features following LViT style, including modality, lesion type, number, and lesion size. We further evaluated robustness on MosMedData+ using 20% original text, 40% incomplete text (randomly removing half of the words), and 40% empty text. The performance drop (Dice -0.62%) indicates robustness to sparse or degraded textual information.
To R3: Q1: Motivation A1: We agree that segmentation is a localization task rather than a diagnostic task. Our intention is not to equate segmentation with diagnosis, but to highlight that purely visual matching can incorrectly activate semantically irrelevant regions under ambiguous medical scenarios. Our CoT is designed to mimic radiological reading logic for resolving ambiguous localization rather than performing diagnosis itself. In Fig. 3 (Case 2), most baselines incorrectly segment a calcified region as COVID-19 infection due to visual similarity. By using CoT-derived morphological and contextual cues, CERS retrieves semantically consistent references and suppresses attention to irrelevant regions. We will further clarify this motivation in the revision.
Q2: Novelty A2: Our contribution is not simply adding CoT into a standard segmentation pipeline. The key novelty lies in using reasoning as a retrieval consistency signal to filter visually similar but semantically inconsistent hard negatives. Our proposed coarse-to-fine retrieval design, together with the MCAM designed to adapt retrieval-derived semantic signals, is specifically tailored to improve target grounding under ambiguous medical contexts. Ablation studies confirm complementary improvements beyond a naive module combination.
Q3: References A3: We will expand more references, including MedSeg-R, to better contextualize our contributions.
Q4: Experimental mistakes A4: We confirm there are no experimental errors. LeFeD uses a heavily modified V-Net, making it infeasible to replace with ConvNeXt. Therefore, the performance difference relative to UCMT can likely be attributed to both architectural and methodological factors. For DuCiSC, the gap is most pronounced on QaTa-COV19.We repeated this evaluation three times (mean Dice: 71.44, variance: 0.23), confirming stable results. The lower performance mainly reflects DuCiSC’s weaker generalization on different datasets.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers generally find the paper timely and promising, with a coherent CERS framework that integrates LLM-generated CoT descriptions, image-CoT retrieval, and MCAM for semi-supervised medical image segmentation, supported by encouraging results across three datasets. However, the current reviews raise important issues that should be clarified in the rebuttal: the motivation and novelty relative to existing image-text/retrieval-based segmentation methods, the reliability and validation protocol for the generated CoTs, the provenance and robustness of auxiliary text, and apparent inconsistencies in reported experimental results across tables.
- 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 addressed reviewers’ concerns reasonably well. The idea is interesting and the empirical evidence is promising, but the technical novelty is somewhat incremental in that it combines reasoning, retrieval, image-text alignment, and segmentation components, and the final paper should present the CoT validation, text-source provenance, robustness checks, MCAM ablation, and corrected tables clearly.
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.
After reading the rebuttal, I recommend Accept. CERS uses LLM-generated reasoning descriptions as a retrieval space for semi-supervised segmentation, rather than using text only as an auxiliary prompt. The pipeline is easy to separate into its main steps: CoT generation for labeled cases, knowledge-pool construction, coarse-to-fine retrieval for unlabeled cases, and feature fusion through MCAM with a dual-decoder consistency loss. The results on CT, X-ray, and MRI datasets show consistent gains over the reported semi-supervised baselines.
The initial concerns were mainly about reliability of the text signal and experimental reporting. The paper depends on generated CoTs, so it was important to know whether these descriptions were clinically plausible and whether the method would fail when the auxiliary text was sparse or noisy. The rebuttal gives the missing details of the two-radiologist validation protocol and adds a robustness test for sparse auxiliary text. It also addresses the table inconsistency and clarifies the baseline results. These responses make the empirical case easier to trust.
I still think the final version should be careful in how it describes “reasoning.” The experiments show that retrieved structured descriptions help segmentation, but they do not prove that the model performs clinical reasoning in the human sense. The provenance of the auxiliary text and the validation protocol should also be documented clearly. These are presentation and reporting issues rather than reasons to reject. Overall, the method is well scoped, the gains are consistent, and the rebuttal resolves the main concerns, so I support acceptance.