List of Papers Browse by Subject Areas Author List
Abstract
Leveraging diagnostic reports as auxiliary guidance has emerged as a promising solution to data scarcity in medical referring image segmentation. However, existing methods often suffer from modality heterogeneity due to shallow interaction strategies and scale mismatch between pre-trained encoders and task-specific decoders. In this paper, we propose MedTSC-Net, a systematic and optimized framework designed to bridge these semantic and scale gaps through a harmonized encoding-decoding paradigm. Specifically, we introduce a Visual-Linguistic Harmonizer (VLH) that projects image and text representations into a shared latent space, mitigating semantic shifts at the early encoding stage. To ensure precise lesion delineation, a Channel-Spatial Attention (CASA) module is devised to model complementary text-guided dependencies across multiple dimensions. Furthermore, we address the receptive field inconsistency via a Receptive Field Alignment (RFA) mechanism, which recalibrates feature scales to enhance boundary localization. Extensive experiments on the QaTa-COV19 and MosMedData+ datasets demonstrate that MedTSC-Net significantly outperforms state-of-the-art methods, achieving Dice scores of 91.89% and 80.64%, respectively. Notably, our framework exhibits superior robustness in few-shot scenarios, highlighting its potential for clinical applications with limited annotations. Our code is publicly available at https://github.com/DYcheng-tech/MedTSCNet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0644_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/DYcheng-tech/MedTSCNet
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheZha_MedTSCNet_MICCAI2026,
author = { Chen, Zhaomin AND Cheng, Danyang AND Ge, Yisu AND Zhang, Guodao AND Huang, Hui AND Chen, Huiling},
title = { { MedTSC-Net: Bridging Semantic and Scale Gaps in Text-Guided Medical Image 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
MedTSC-Net proposes three targeted modules — VLH for early cross-modal alignment, CASA for text-guided channel-spatial attention, and RFA for receptive field recalibration — to bridge the semantic and scale gaps in text-guided medical image segmentation, achieving state-of-the-art performance especially under limited annotation.
- 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.
Strengths: MedTSC-Net presents a well-motivated framework addressing real limitations in text-guided medical segmentation. The three proposed modules (VLH, CASA, RFA) are logically cohesive and target distinct failure modes. Results on both datasets are compelling, and the few-shot analysis (87.99% Dice with only 5% data) is particularly convincing for clinical relevance.
- 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.
Weaknesses: The evaluation is limited to COVID-19 datasets, raising concerns about generalizability to other anatomies or modalities. The architectural complexity introduced by three additional modules lacks thorough computational cost analysis (FLOPs, inference time). Additionally, the VLH alignment depth K=12 and other hyperparameter choices appear empirically set without sensitivity analysis, and the qualitative text-manipulation experiments (Fig. 4) remain anecdotal without quantitative grounding.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
MedTSC-Net presents a technically sound framework with well-motivated module designs and consistent performance gains over strong baselines across two imaging modalities, with the few-shot robustness being particularly compelling for clinical relevance. However, the evaluation is confined solely to COVID-19 lung segmentation, limiting generalizability claims, and no computational cost analysis is provided. The VLH design also bears conceptual similarity to existing adapter-based alignment methods without sufficiently distinguishing its novelty. These limitations collectively support a weak accept — a meaningful incremental contribution, but broader evaluation and sharper novelty positioning are needed to warrant higher confidence.
- 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.
Although the methodological novelty is somewhat incremental, the authors sufficiently addressed the reviewers’ concerns through strong empirical validation, fair multimodal comparisons, clinically credible annotations, and efficient deployment performance, collectively justifying acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes MedTSC-Net, a text-guided medical image segmentation framework addressing semantic and scale mismatches between modalities. It introduces a Visual-Linguistic Harmonizer for early alignment, CASA for multi-dimensional attention, and RFA for receptive field calibration. Experiments show good performance, particularly under limited data 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.This paper addresses an important problem in medical visual language segmentation: shallow interaction strategies and scale mismatch between pre-trained encoders and task-specific decoders. 2.The proposed Visual-Linguistic Harmonizer introduces early-stage cross-modal alignment, which helps reduce semantic discrepancies between image and text representations. 3.Experimental results demonstrate consistent improvements over baselines 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 CASA module builds upon standard channel and spatial attention mechanisms, and its contribution is largely incremental rather than conceptually innovative. 2.Lack analysis of computational complexity and inference efficiency. 3.Not provide sufficient analysis of how alignment depth (K in VLH) affects performance.
- 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.
(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 well-motivated framework addressing semantic and scale gaps with empirical improvements and few-shot performance. However, the overall novelty is moderate, with components largely built on existing designs. But overall, I am positive about 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 authors have addressed my concerns, I recommend accept.
Review #3
- Please describe the contribution of the paper
This paper presents MedTSC-Net, a text-based medical image segmentation framework, which enhances the segmentation effect through text.
- 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 suggests that the use of text can be employed to guide and enhance the effect of image segmentation.
2.The method proposed in the article achieved the best results in both the comparative experiments and the visualizations.
3.The author proposed an appropriate module for the model based on the newly added text guidance.
- 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 author did not specify the source of the texts in the dataset, whether it matches actual clinical practice, and whether there are any unfair comparisons.
2.The text in the dataset used during training has marked the positions and quantities of the divisions, which may lead to unfair comparisons and prevent the direct demonstration of the superiority of the algorithm.
3.The paper fails to adequately demonstrate what unique mechanisms the proposed module approach has added. Currently, it seems more like a combination of improvements to existing modules rather than a clear breakthrough in a new methodology.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The article lacks a certain degree of originality. 2.The use of such text guidance raises suspicions of unfairness.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
After carefully reading the rebuttal, I maintain my final recommendation of Reject. The authors clarified the source of the textual annotations and the comparison protocol, but my main concerns are not fully resolved. The text prompts contain strong information such as severity, lesion location, and quantity, and the paper still does not convincingly demonstrate that the performance gains come from the proposed architecture rather than from highly informative textual guidance. In addition, the methodological novelty remains limited. The rebuttal itself frames the contribution as a systematic solution rather than a fundamentally new method, and the choice of key hyperparameters is still largely empirical. The additional generalization evidence is also weak and insufficient to support broad applicability beyond the main COVID-19 setting.
Author Feedback
Q1: Methodological Novelty (R1, R2, R3) A2: Though lacking foundational math innovations, our core contribution is a systematic clinical solution addressing three referring segmentation bottlenecks: (1) inadequate multi-modal alignment; (2) over-reliance on local text, ignoring global contexts (e.g., severity); and (3) encoder-decoder capability gaps. By aligning multi-scale visual features with global semantics, we bridge these gaps. Few-shot results confirm robustness under data scarcity, proving real-world utility. Q2: Hyperparameter Sensitivity (R1, R2) A2: We set K as a hyperparameter to flexibly accommodate different pre-trained text encoder scales. During development with our 12-layer BERT, we explored numerous Harmonizer setups (e.g., varying application stages, skip-connections, and K values). The optimal configuration empirically proved to be covering all BERT layers and the final two vision encoder stages, determining our choice of K=12.Q3: Computational Complexity and Inference Efficiency (R1 R2) A3: MedTSC-Net is highly efficient. Compared to Ariadne [11], trainable parameters increase by only 4.7% (+2M) to 45M. Requiring 15.7 GFLOPs/image, it achieves a 45ms average latency on one RTX 3090, satisfying real-time clinical needs. Q4: Generalization Capability (R1) A4: Our primary focus on COVID-19 stems from the current scarcity of high-quality, text-annotated medical segmentation benchmarks. Nevertheless, our evaluation spans two distinct imaging modalities: chest X-rays (QaTa-COV19) and CT slices (MosMedData+). MedTSC-Net consistently improves performance on both, demonstrating its robustness against imaging variations. Furthermore, our framework relies on no organ-specific priors or hand-crafted anatomical rules. Although omitted from the main text due to space limits, we conducted a preliminary lightweight validation on the MoNuSeg pathology dataset (nuclei segmentation). MedTSC-Net achieved an 81.54% Dice score (+0.55% over Ariadne’s 80.99%), clearly validating its strong generalization potential beyond lung infection scenarios. Q5: Quantitative Basis for Text Perturbation Experiments (R1) A5: Omitted for space, our quantitative text perturbation analysis involved removing specific details (severity, quantity, location) or using a generic “lung infection” prompt. The model’s Dice score ranged from 91.89% (optimal) to 88.34% (lower bound). Crucially, even this 88.34% lower bound outperforms pure-vision baselines. This proves that embedding text priors during training does more than just enable cross-modal reasoning; it fundamentally strengthens the visual branch’s own representation learning capability. Q6: Source and Nature of Textual Data (R3) A6: The public datasets we used (QaTa-COV19 and MosMedData+) feature non-arbitrary, authoritative text annotations directly adopted from the pioneering LViT [10]. Ensuring high clinical reliability, these texts were manually annotated and cross-validated by two senior medical experts from UT Southwestern Medical Center. Encompassing infection severity, specific locations, and lesion quantity, this structural design faithfully replicates real-world radiology reports and physician referral notes, perfectly aligning our evaluation with actual clinical practice. Q7: Fairness of Text-Guided Comparisons (R3) A7: We respectfully clarify that utilizing detailed text prompts is the defining formulation of referring segmentation, not “label leakage” compared to traditional blind segmentation. Following standard protocols (LViT [10], Ariadne [11]), all evaluated multimodal baselines received identical image-text pairs. MedTSC-Net achieves SOTA performance under absolute information parity, proving the inherent superiority of our multi-modal interactive architecture rather than an unfair information advantage. Pure-vision methods were included solely to establish a performance lower bound and quantify the benefits of text priors, a standard practice in this field.
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.
Reviewers found MedTSC-Net well motivated, clearly presented, and empirically strong, with consistent gains on QaTa-COV19 and MosMedData+ and compelling few-shot results. However, the overall reviews remain mixed due to concerns about limited generalizability beyond COVID lung data, moderate novelty of the VLH/CASA/RFA components, missing computational cost and hyperparameter sensitivity analysis, and especially the provenance and fairness of the text guidance, which may encode lesion location/count information and affect comparisons.
- 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 several concrete concerns. However, across the reviews there is also broad agreement that the method is more incremental than foundational. The final paper should temper claims of broad generalization and clarify the role and clinical nature of the text prompts.
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 paper addresses a relevant problem in text-guided medical image segmentation, namely how to better align visual features with clinical textual descriptions under limited annotation. The proposed MedTSC-Net is well motivated, and its three modules are organized around clear technical goals: early visual-linguistic alignment, text-guided attention, and receptive-field adjustment. The experimental results on QaTa-COV19 and MosMedData+ are consistently strong, and the few-shot results are particularly valuable given the annotation burden in medical imaging. The reviewers raised valid concerns about the moderate novelty of the individual modules, the limited evaluation beyond COVID-related lung data, computational cost, and the fairness of text guidance. The rebuttal addresses these issues to a reasonable extent. In particular, the authors clarify the provenance of the textual annotations, explain that all multimodal baselines use the same image-text pairs, provide efficiency numbers, and add evidence from text perturbation and preliminary non-lung validation. While the method is still more of a careful system-level integration than a fundamentally new formulation, the empirical evidence is strong, the problem setting is clinically meaningful, and two reviewers moved to or maintained acceptance after rebuttal. Overall, I believe the paper makes a solid and useful contribution to text-guided medical segmentation. I would support acceptance, while encouraging the authors to further temper broad generalization claims and include the added efficiency and text-guidance analyses in the final version.
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 paper proposes MedTSC-Net, which is a text-guided medical image segmentation framework addressing semantic and scale mismatches between modalities. Overall, the method is a well-motivated framework with empirical improvements and few-shot performance, although the methodology novelty is moderate. Authors addressed most concerns of R1 and R2, such as computational complexity and some ablation settings. Hence, recommend acceptance of this work.