List of Papers Browse by Subject Areas Author List
Abstract
Semi-supervised medical image segmentation aims to reduce annotation costs, yet achieving anatomically consistent segmentation under limited supervision remains challenging. Most methods rely on visual representations or coarse vision–language alignment strategies that assume spatially agnostic correspondence between modalities, overlooking localization cues in visual features and textual descriptions. Consequently, pixel–text or mask–text matching yields ambiguous cross-modal alignment, particularly for anatomically adjacent or visually similar structures. We propose PC-VLG, a position-correlated vision–language graph framework for semi-supervised medical segmentation that establishes object-level positional correspondence between visual anatomy and textual semantics. A Vision–Language Graph Alignment Network injects multi-modal embeddings into Position-Correlated Cross-Modal Graph modules, where anatomical structures are modeled as graph nodes enriched with visual, textual, and spatial cues, and inter-object relationships are encoded as geometry-aware edges to enable structured and consistent cross-modal alignment beyond conventional pixel–text or mask–\linebreak text paradigms. Experiments on multi-object datasets demonstrate that PC-VLG outperforms state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6229_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/QLYCode/PC_VLG
Link to the Dataset(s)
ACDC dataset: https://www.creatis.insa-lyon.fr/Challenge/acdc/databases.html
FLARE21 dataset: https://flare.grand-challenge.org/
BibTex
@InProceedings{QiuLuy_PCVLG_MICCAI2026,
author = { Qiu, Luyi AND Dang, Min AND Pan, Feng AND Wang, Yifan AND Kong, Adams Wai-Kin},
title = { { PC-VLG: Position-Correlated Vision–Language Graph Alignment 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
PC-VLG (Position-Correlated Vision–Language Graph Alignment) is a semi-supervised medical image segmentation framework that explicitly models the positional relationships between anatomical structures in both visual and textual modalities. It constructs graph representations where nodes correspond to anatomical objects enriched with visual, textual, and spatial cues, and edges encode geometry-aware relationships. This structured approach enhances cross-modal alignment, leading to more anatomically consistent segmentation, especially in scenarios with limited labeled data.
- 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 use of object-level graphs with geometry-aware edges effectively captures spatial relationships between anatomical structures.
Moving beyond pixel or mask-level matching, the graph-based method enables more semantically and spatially consistent alignment between images and texts.
Demonstrated improvements over state-of-the-art methods on datasets like FLARE and ACDC, indicating robust and accurate segmentation performance under semi-supervised settings.
Utilizes pretrained visual priors (Med-SAM) and domain-specific textual encodings, enriching the feature representations.
- 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 graph construction and multi-module architecture increase model complexity, which may demand more computational resources and meticulous tuning.
The effectiveness hinges on the quality of coarse segmentation and prompts, which, if inaccurate, could impair graph quality and subsequent alignment.
While results are promising on certain datasets, the paper provides limited insights into the model’s robustness across diverse medical imaging modalities or unseen anatomy variations.
Despite improvements, structures with very similar appearances and close proximity might still pose challenges for precise positional graph modeling.
- 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 added complexity and reliance on accurate priors could limit scalability or ease of deployment, particularly in resource-constrained settings. Ensuring robustness across a broader spectrum of medical datasets remains an area for future exploration, but overall, the approach offers a meaningful advance in semi-supervised medical image segmentation.
- 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 PC-VLG, a semi-supervised medical image segmentation framework that integrates vision-language graph alignment with explicit positional correspondence. The method uses a LoRA-adapted visual prior branch (leveraging Med-SAM), a textual semantic prior branch encoding medical knowledge, and a Position-Correlated Cross-Modal Graph (PC-CMG) module to model object-level anatomical relationships. Experiments on FLARE and ACDC datasets demonstrate incremental improvements over existing semi-supervised 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.Technically Sound Core Idea: The graph alignment mechanism—query-guided attention for node formation, fusion of visual/positional/textual cues, and attention-entropy-based uncertainty gating—is a reasonable approach to mitigate the over-smoothing and boundary ambiguity inherent in pixel-text or global mask-text alignment.
2.Well-Executed Ablation: The incremental gains shown in the ablation study (Table 1, Fig. 4) clearly attribute performance contributions: VAP provides substantial improvement (validating the Med-SAM prior), TSP benefits small/ambiguous classes (e.g., pancreas), and PC-CMG further refines boundaries on both large and small structures.
- 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.Incomplete Comparison with Recent SOTA The most recent baseline included is BoCLIS (TMI, 2025.04). The current selection of baselines gives the impression of cherry-picking suboptimal competitors rather than a rigorous SOTA evaluation, especially on ACDC dataset, which is one of the most popular benchmark datasets. Additional recent methods should have been considered, e.g., [1,2,3,4].
2.Unclear Justification for Methodological Complexity The proposed framework is architecturally heavy. This raises a fundamental question: Is this level of complexity truly necessary for semi-supervised segmentation? A simple nnUNet trained on the same amount of labeled data could potentially match or exceed the reported performance, with dramatically lower computational and implementation overhead. The paper provides no discussion comparing against such a minimalist baseline.
3.Missing Critical Implementation Details Key components are described only at a high level, hindering reproducibility:
- The construction and source of textual knowledge (concept-definition pairs and triplets) for the specific organs in FLARE/ACDC are not specified.
- The architecture of the “redesigned mask decoder” in the VAP branch is not detailed.
- The number of graph stages, hidden dimensions, and normalization schemes in PC-CMG is omitted.
- Relation types and the embedding matrix E_r are mentioned but never defined or exemplified.
4.Limited Evidence of Practical Gains Performance improvements over the strongest baseline (BoCLIS) are marginal (e.g., ~0.6% Dice on ACDC at 10% labels). No statistical significance testing is reported. Given the substantial complexity added, the incremental gains do not convincingly justify the method’s overhead.
5.Some concern The proposed framework introduces substantial complexity, including Med-SAM encoder, UMamba backbone, dual-branch decoding, textual encoding, and multi-stage graph propagation, yet delivers less than 1% Dice improvement over the LESS SOTA baseline (BoCLIS). More importantly, most successful SSL methods preserve inference-time simplicity (collapsing to a standard U-Net/VNet), whereas this work bakes heavy computation permanently into deployment. Such over-engineering risks optimizing benchmarks at the expense of clinical practicality. While I personally admire the authors’ exploration, the work leans toward method stacking rather than offering deep insight into the semi-supervised segmentation problem.
[1] CauSSL: Causality-inspired Semi-supervised Learning for Medical Image Segmentation, ICCV 2023 [2] Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation, CVPR 2024 [3] β-FFT: Nonlinear Interpolation and Differentiated Training Strategies for Semi-Supervised Medical Image Segmentation, CVPR 2025 [4] Dual cross-image semantic consistency with self-aware pseudo labeling for semi-supervised medical image segmentation, IEEE TMI (2025.07)
- 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 core idea of position-correlated graph alignment is technically sound, and the ablation study convincingly demonstrates the contribution of each component. However, the methodological complexity is poorly justified relative to the modest performance gains, and the omission of very recent SOTA baselines weakens the comparative evaluation.
I am casting a “Weak Accept” to ensure this paper reaches the rebuttal phase, where I hope the authors will address the concerns above, particularly the philosophical question regarding complexity and community direction. A satisfactory rebuttal that includes a discussion of these issues and a commitment to adding a simple baseline comparison would strengthen my final recommendation.
- 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 PC-VLG, a position-correlated vision-language graph alignment framework for semi-supervised medical image segmentation. By integrating visual priors, textual semantics, and geometry-aware graph reasoning, it achieves anatomically consistent cross-modal alignment and delivers strong performance improvements over prior methods on FLARE and ACDC benchmarks.
- 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 tackles an important problem in semi-supervised medical image segmentation, where reducing annotation cost while preserving anatomical consistency is highly valuable. 2.The method is well motivated by clearly identifying the limitations of existing pixel-text and mask-text alignment strategies in anatomically adjacent structures.
- 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 main task of this paper is image-text alignment, and the task is semi-supervised segmentation. No explanation was given on how this alignment works for semi-supervised segmentation. 2.The overall framework mainly combines existing components, including SAM-based visual priors, BERT-based text encoding, graph reasoning, and semi-supervised consistency learning, so the methodological novelty appears somewhat incremental. 3.Although PC-VLG achieves the best reported performance, many of the gains over competitive methods are incremental. 4.Lack an ablation study on the consistency weight μ in Eq. (15), making it unclear how sensitive the method is to this hyperparameter.
- 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?
The paper addresses an important problem and proposes a reasonable vision-language graph framework for semi-supervised medical segmentation. The method is well motivated and shows consistent improvements on FLARE and ACDC. However, the overall novelty appears somewhat incremental. In addition, the experiment gains over baselines are relatively modest, and the paper lacks important ablations, such as the consistency weight μ in Eq. (15).
- 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
Author Feedback
We thank the reviewers for their valuable comments. We appreciate Reviewer #1 for recognizing the contribution to semi-supervised medical image segmentation, Reviewer #2 for acknowledging the technical soundness and ablation studies, and Reviewer #4 and the Meta-Reviewers for recognizing the work as a well-motivated contribution. Below, we address the raised concerns.
Response to Reviewers#1, #2, and #4: Motivation, Novelty. Existing SSL methods mainly rely on perturbation consistency or uncertainty regularization, while current vision-language methods are limited to pixel-/mask-text alignment without structured semantic modeling. PC-VLG models positional semantic correspondence between anatomical structures and textual concepts through graph-based visual, textual, and positional representations with relation-aware anatomical edges. Regarding the novelty concerns from Reviewers#1 and #4, this work introduces object-level positional vision-language alignment for semi-supervised segmentation through semantic correspondence between anatomical structures and textual concepts, rather than incremental refinement. As noted by Reviewer#4, the refined predictions are used for unlabeled consistency supervision in Eq.(13), improving pseudo-label quality during optimization. We acknowledge the challenge of adjacent anatomical structures with similar appearances. PC-CMG mitigates this issue through positional embeddings, textual semantics, geometry-aware graph edges, improving discrimination between nearby organs. The HD95 improvements and Fig.4 support this observation.
Response to Reviewers#1 and #2: Complexity and Practicality. To address concerns on architectural complexity, prompt sensitivity, and deployment practicality, the additional components are introduced to model object-level positional correspondence and improve pseudo-label reliability. The Med-SAM image/prompt encoders are frozen to provide anatomical priors, while PC-CMG performs object-level graph reasoning. The coarse segmentation network remains the primary backbone, and VLGR-Net provides semantic refinement and consistency supervision. Consistency-aware prompt generation and uncertainty-gated graph propagation reduce sensitivity to inaccurate coarse segmentation and prompts, addressing Reviewer#1’s concern. The vision-language graph alignment is only used during training, while inference relies solely on the coarse segmentation network. We clarify the inference cost in the revised manuscript. Regarding nnUNet, it is a fully supervised method and therefore not comparable to SSL methods.
Response to Reviewers#1, #2, and #4: Experimental Validation and Practical Gains. To address Reviewer#2’s concerns regarding recent comparisons and practical gains, our experiments already include several recent SSL methods, including BoCLIS (TMI 2025), AAUM (MIA 2024), BCP (CVPR 2023), and PatchCL (CVPR 2023). While some recent SSL approaches achieve stronger numerical performance, PC-VLG focuses on position-aware vision-language semantic correspondence for structured cross-modal semantic regularization. This perspective complements existing consistency- and uncertainty-driven SSL paradigms. Experiments show consistent improvements in structural consistency.
Response to Reviewer #2 and #4: Implementation. We expand the revised manuscript with details, including: (1) organ-specific textual descriptions derived from medical knowledge sources and anatomical relation triplets; (2) the VAP decoder and feature fusion strategy; (3) graph stages, hidden dimensions, and node/edge update operations in PC-CMG; (4) the construction of Er; (5) training details, including optimizer settings, ramp-up schedule, and prompt generation settings; (6) the time-dependent Gaussian warm-up schedule of μ with μ_max=0.1, following Luo et al., “Semi-Supervised Medical Image Segmentation via Cross Teaching between CNN and Transformer,” where the consistency term avoids unstable early-stage pseudo-supervision.
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.
I recommend provisional acceptance. The paper presents a well-motivated vision-language graph framework for semi-supervised medical image segmentation, and the reviewers generally found the position-aware graph alignment idea technically sound, with useful ablations and consistent improvements on FLARE and ACDC. The main reservations are the added system complexity, relatively modest gains over strong baselines, missing details on text-knowledge construction and graph implementation, and incomplete comparisons with very recent SSL methods. Overall, despite these weaknesses, the paper offers a meaningful and reasonably well-supported contribution, and the concerns appear addressable in revision.