Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

Recent vision–language models for medical image segmentation leverage radiology reports as auxiliary supervision. However, existing methods inject text cues uniformly across decoder stages, overlooking the hierarchical semantic demands of different visual abstraction levels. This limits precise alignment between global context and fine-grained lesion representations, particularly for diffuse abnormalities. To address this issue, we propose Hierarchical Prompt Selector (HiProS), a vision–language segmentation framework that performs stage-conditioned textual token selection for local–global alignment. HiProS introduces a Stage-wise Prompt Selector that explicitly selects and suppresses textual tokens according to the abstraction level of each decoder stage, rather than propagating all text cues through uniform attention. The selected tokens are then integrated with multi-scale visual features via a Fusion Alignment Module, enabling structured cross-modal interaction at each decoding stage. Importantly, HiProS operates with automatically generated textual cues, avoiding reliance on manually curated clinical reports and enabling scalability where expert annotations are unavailable. Experiments on QaTa-COV19 and MosMedData+ demonstrate consistent improvements over strong vision–language baselines, with competitive results. Notably, HiProS shows pronounced gains in segmenting diffuse lesions on MosMedData+, indicating that stage-conditioned textual selection mitigates semantic misalignment in fine-grained segmentation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4897_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)

QaTa-COV19: https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset MosMedData+: https://www.kaggle.com/datasets/maedemaftouni/covid19-ct-scan-lesion-segmentation-dataset

BibTex

@InProceedings{KimDa_HiProS_MICCAI2026,
        author = { Kim, Da-Hee AND Chae, Jongbin AND Lee, Seungyeon AND Lee, Dong-Gyu},
        title = { { HiProS: A Hierarchical Prompt Selection for Stage-wise Vision–Language Alignment in 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

This paper proposes a hierarchical prompt selection framework, HiProS, for medical image segmentation. Specifically, it aligns fine-grained observations from the Findings section of medical reports with shallow layers of the image decoder, while aligning high-level diagnostic summaries from the Impressions section with deeper layers. In addition, the method employs a hard-gating mechanism to select informative text tokens, thereby enabling more precise local-to-global multimodal semantic alignment.
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 is insightful and clinically intuitive. Mapping the clinical reasoning process, from detailed observations to summarized diagnosis, onto the feature hierarchy of deep networks, from shallow to deep representations, is an elegant and well-motivated design. 2.The use of learnable, stage-specific hard gating with a straight-through estimator is well designed. It provides a more explicit and transparent alternative to conventional unified cross-attention by selectively preserving or suppressing text tokens at different stages. 3.The hierarchical alignment strategy is well matched to the structure of radiology reports. It addresses a known limitation in prior vision-language segmentation methods, namely the repeated use of a single shared text representation across all decoder stages.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

1.The method combines MiniGPT-Med for generating Findings/Impressions, BioClinicalBERT for text encoding, ConvNeXt-Tiny for image encoding, and a training objective built on Dice, cross-entropy, and boundary loss. Taken together, the framework seems closer to a reasonable integration of existing components than to a fundamentally new methodological advance. 2.The FAM module appears technically sound but not particularly novel. Structurally, it resembles a fairly standard implementation of stage-specific text-aware cross-attention followed by residual refinement. 3.The claim of improved scalability via automatically generated reports is not fully convincing. In practice, vision-language models such as MiniGPT-Med or LLaVA-Med are known to suffer from hallucinations. If the report generator produces incorrect Findings for a given image, forcing the segmentation network to align with such erroneous text may substantially interfere with visual feature learning and degrade performance. This issue is not sufficiently discussed. 4.The paper does not provide sufficient information about the prompt templates used for report generation, nor does it clearly describe what safeguards were adopted to prevent information leakage during the generation process. 5.It does not report the additional cost of text generation, the training or inference time, or the parameter/FLOP overhead introduced by SPS and FAM. 6.Minor issue: the figures are weak. All three figures in the paper are relatively uninformative, do not highlight the key ideas effectively, and are not visually well designed.
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 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 presents an interesting and clinically intuitive framework for vision-language medical image segmentation. The hierarchical alignment design is well motivated and appears reasonable overall. That said, the current version would be further strengthened by a clearer discussion of the method’s novelty, the reliability of automatically generated reports, and its practical computational cost.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #2

Please describe the contribution of the paper

The authors present HiProS, a language-guided medical image segmentation framework designed to align specific segments of a radiology report with corresponding decoder stages. The pipeline automatically generates hierarchical reports, encodes them, and employs a Stage-wise Prompt Selector (SPS) to filter text tokens based on their cosine similarity to pooled visual features at each stage. Finally, a Fusion Alignment Module (FAM) integrates the selected text tokens with the image features.

The primary contribution of this work lies in the explicit token selection mechanism tied to decoder stage semantics, coupled with the hierarchical split. However, the broader design space of stage-aware cross-modal alignment is already well-explored. Existing baseline models (such as TGCAM, MAdapter, RecLMIS, and FMISeg) utilize similar concepts, including interaction of multi-level features, word-of-interest prediction, and progressive cross-attention. Consequently, HiProS is best characterized as a clean, explicit, and effective variant within an established lineage of multi-modal segmentation research.
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.

Architectural design: The foundational intuition of routing global report summaries to coarser semantic stages and detailed descriptions to finer stages is logical. The authors successfully translate this concept into a concrete, stage-conditioned gating mechanism rather than relying on generic cross-attention.

Robust empirical evaluation: The framework demonstrates solid performance, particularly on the MosMedData+ dataset, outperforming established baselines. The comparisons are fair and meaningful, as the authors note that key baselines (TGCAM, MAdapter, FMISeg) share the identical backbone.

Comprehensive ablation studies: The experimental validation of the architecture is thorough by ablating major components. This provides compelling evidence that both the hierarchical text split and the fusion module actively contribute to the model’s success.

Practical clinical applicability: The framework is well-motivated for real-world scenarios lacking clinician-authored reports. Utilizing automatically generated hierarchical reports is a good systems-level contribution, allowing the method to be applied to standard image-only segmentation 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.

Incremental conceptual novelty: The field has largely moved beyond uniform text injection, utilizing various cross-attention and region-of-interest mechanisms. While the explicit stage-wise hard gating of text tokens is a promising refinement, it represents a moderate, incremental step rather than a paradigm shift in cross-modal alignment.

Lack of interpretability: The core novelty of the token routing is under-validated because the paper does not visualize or analyze the routed tokens. It can be important to see which specific words are selected at each stage, how the thresholds evolve, and whether anatomical terms genuinely map to fine stages while summary terms map to coarse stages. Without this analysis, it is difficult to verify if the routing mechanism functions as theorized.

Heuristic stage assignment: The mapping of report sections to decoder stages is manually prescribed (Impressions to the highest abstraction, Findings to lower stages). While the swap ablation proves this specific mapping is optimal among binary choices, the paper lacks comparisons to learned routing policies, soft routing, or similar, leaving the framework feeling overly engineered rather than data-driven.

Narrow evaluation scope: The experimental validation is restricted to two 2D pulmonary infection datasets, which belong to a closely related lesion family. While sufficient to demonstrate promise, this limited scope fails to establish HiProS as a broadly generalizable medical segmentation framework.

Confounding variables in performance gains: It is unclear how much of the performance improvement stems from the novel architecture versus the high-quality generated text. Because the full model uses newly generated, structured reports while most baselines use raw text, the attribution of the performance gains is somewhat entangled.
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?

This manuscript presents a solid framework with a practical approach to generating and utilizing hierarchical text for medical segmentation. The empirical results are good on challenging datasets, and the architectural design is backed by ablation studies that isolate the contributions of the proposed modules.

However, the conceptual advance remains incremental relative to recent multi-modal segmentation literature. The paper is primarily held back by its limited evaluation scope and the lack of qualitative validation for its central claim.

To warrant a stronger score, the authors may directly validate the token routing mechanism by demonstrating exactly which text tokens are selected at each decoder stage and proving that this routing aligns with the theorized local-global hierarchy.

In its current form, the submission represents a practically useful, well-executed variant of existing methods. The overall quality exceeds the MICCAI acceptance bar, resulting in a recommendation of Weak Accept.
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.

Rebuttal decently addressed the concerns, I will keep my initial judgement of accept.

Review #3

Please describe the contribution of the paper

This paper proposes HiProS, a hierarchical vision-language framework for medical image segmentation. The key idea is to exploit the hierarchical structure of clinical text by separating Findings and Impressions, and injecting them into different decoder stages according to visual abstraction levels. The method includes a Stage-wise Prompt Selector (SPS) for token selection and suppression, and a Fusion Alignment Module (FAM) for cross-modal fusion at each stage. The paper reports strong results on QaTa-COV19 and MosMedData+, with comprehensive comparisons and ablations.
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 clearly written and well organized, with a coherent overall story. The paper does a good job of maintaining a clear writing from the problem motivation to the proposed stage-wise alignment design. 2.The experiment results are strong, and the experiments are relatively complete. The improvements on MosMedData+ are particularly notable. The qualitative results also suggest better handling of diffuse or subtle lesions.
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.A key concern is novelty. The central intuition of aligning Findings and Impression with different visual abstraction levels has already been explored in prior work such as IMITATE[1], which performs hierarchical report-aware vision-language alignment during pre-training. This is an important missing citation given the strong conceptual similarity. And the authors should either provide a more explicit comparison to IMITATE, or clearly justify why decoder-level integration for segmentation constitutes a non-trivial and distinct contribution beyond prior hierarchical VLP designs.

2.The paper does not provide sufficient support for the key design choice of assigning Findings to lower stages and Impressions to higher stages. While this mapping may be intuitively plausible from a semantic perspective, it is much less clear whether such a correspondence should also hold at the level of text embeddings used for cross-modal fusion. In other words, even if Findings are more fine-grained and Impressions more global, the paper does not establish why their encoded representations should naturally align better with lower-level versus higher-level decoder features. If this design choice is supported by prior work, the relevant literature should be cited explicitly; otherwise, stronger experimental validation or analysis would be needed.

[1] Liu, C., Cheng, S., Shi, M., Shah, A., Bai, W. and Arcucci, R., 2024.Imitate: Clinical prior guided hierarchical vision-language pre-training. IEEE Transactions on Medical Imaging, 44(1), pp.519-529.
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.

(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 gave this paper a weak reject recommendation. The paper is well written, easy to follow, and reports strong empirical results with a reasonably complete experimental section. However, I have three main concerns. First, the novelty appears limited, as the main idea of aligning different textual semantic levels with different visual abstraction levels seems related to prior hierarchical vision-language work, while the proposed modules themselves appear largely incremental. Second, the paper misses a very important citation to IMITATE. Third, the key design choice of assigning Findings to lower decoder stages and Impressions to higher stages lacks sufficient theoretical or empirical support or prior work citation.
Reviewer confidence

Confident but not absolutely certain (3)
[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.

The paper is well written and empirically strong, but my two main concerns remain unresolved.

First, the novelty is still not sufficiently convincing. The central intuition, namely exploiting the Findings/Impressions hierarchy and aligning it with different levels of visual abstraction, appears already substantially explored in IMITATE. The rebuttal acknowledges the missing citation, but mainly distinguishes HiProS by saying that it applies this idea in a segmentation decoder rather than in a pre-training framework. In my view, this is not enough to establish a sufficiently distinct conceptual contribution. The paper still seems closer to a segmentation-specific adaptation of an existing idea than to a genuinely new one.

Second, the key design choice of assigning Findings to lower stages and Impressions to higher stages remains only partially justified. The rebuttal provides some qualitative token-level examples and refers to the existing ablation in Table 2 as partial support, but I am still not convinced that this correspondence is adequately established at the level of encoded text representations used for fusion. Overall, the design is understandable, but still largely heuristic.

For these reasons, I do not think the rebuttal sufficiently resolves the concerns that are central to my evaluation, and I therefore maintain my reject decision.

Author Feedback

We thank the AC and reviewers for their constructive comments and valuable feedback. We considered all concerns and respond below: Novelty (R1, R2, R4): HiProS incorporates existing vision-language (VL) components, but its contribution is not limited to a new backbone or attention operator. The key contribution lies in a stage-aware text utilization strategy aligning hierarchical report semantics with decoder abstraction levels. Unlike prior cross-modal fusion methods using the same text representation across decoder stages, HiProS uses SPS to select stage-relevant tokens and FAM to align only the selected tokens with spatial visual features. Although cross-attention itself is standard, FAM operates on SPS-selected stage-specific tokens rather than full-text representations. Table 4 shows that this routing-aware fusion outperforms simple feature fusion. Prior Work (R4): We thank the reviewer for pointing out the missing citation to IMITATE and will include it in the revision. IMITATE performs hierarchical VL pre-training by aligning multi-level visual features with Findings/Impressions. However, it does not address decoder-stage refinement for dense lesion segmentation or stage-conditioned token routing. In contrast, HiProS injects hierarchical report semantics into the segmentation decoder, where report tokens are selectively routed across decoder stages instead of fixed section-level representations. Token routing (R2, R4): As shown in the report examples of Fig. 1, SPS routes localized anatomical semantics to lower decoder stages and global diagnostic semantics to higher stages. Lower stages selected “right lower lobe consolidation,” “left hilar effusion,” and “perihilar region,” while higher stages selected “pneumonia” and “moderate severity,” consistent with the intended stage-wise semantic allocation. Stage allocation (R2, R4): Prior radiology studies [9, 15] indicate that Findings contain localized anatomical observations, whereas Impressions summarize diagnostic interpretations. This aligns with decoder hierarchies, where lower stages preserve localization-sensitive spatial details while higher stages encode abstract contextual semantics. HiProS assigns Findings to lower stages and Impressions to higher stages as a clinically motivated inductive bias, with Table 2 showing that both “Swap” and “Full Text” underperform this design. SPS adaptively selects stage-relevant tokens via image-text similarity and learnable thresholds. Reliability and information leakage (R1): We agree that VLM-generated reports may contain hallucinations. HiProS uses generated reports as auxiliary semantic guidance rather than ground-truth supervision. SPS mitigates unreliable semantics by selecting visually relevant tokens based on stage-specific image-text similarity. Table 1 shows that HiProS‡ maintains stable performance with reports from another medical VLM (LLaVA-Med), suggesting robustness to different report generators. Report generation used only the input image and a fixed prompt, without ground-truth masks or lesion labels. We will clarify the prompt template in the revision. Computational cost (R1): MiniGPT-Med report generation requires 7.95 sec/report. SPS adds +1.1M Params, while FAM adds +17.9M Params and +4.6 GFLOPs. HiProS uses 48.4M Params and 13.9 GFLOPs, with a training time of approximately 1 hour and an inference time of 7.27 ms/image. Evaluation scope (R2): We acknowledge that broader validation would strengthen generality. Our experiments cover two pulmonary infection datasets across X-ray and CT modalities, with both diffuse and localized lesions, providing partial evidence of generalizability. We plan to evaluate HiProS on additional datasets in future work. Performance attribution (R2): We evaluated HiProS† using the same LViT-provided text as prior studies instead of generated reports (Table 1). HiProS† still outperformed existing baseline models, suggesting gains are not solely attributable to generated text.

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 found HiProS to be a clearly written and practically motivated framework for hierarchical vision- language medical image segmentation, with strong empirical results and useful ablations on QaTa-COV19 and MosMedData+. The proposed stage-wise use of Findings and Impressions, together with token selection and fusion modules, is viewed as intuitive and effective. However, reviewers raised consistent concerns about limited novelty relative to prior hierarchical vision-language alignment work, especially IMITATE; insufficient validation that the selected tokens truly follow the claimed local-to-global hierarchy; possible confounding from generated report quality and limited evaluation on only two related pulmonary infection datasets. The rebuttal should directly address these points, particularly the distinction from prior work and token-routing evidence.
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 main unresolved issue is conceptual novelty: the reviewers generally view HiProS as an effective and well-executed combination of existing vision-language and segmentation components, while Reviewer 4 remains unconvinced that decoder-level stage-wise routing is sufficiently distinct from prior hierarchical report-aware alignment ideas. The rebuttal provides a plausible distinction, but not one that fully eliminates the concern. A second remaining concern is that the Findings-to-lower-stages and Impressions-to-higher-stages assignment is supported by intuition, examples, and ablations, but still lacks a deeper analysis of the learned token routing or of alternatives such as learned or soft routing. The camera-ready version should explicitly cite and discuss the closely related hierarchical prior work including but not limited to Imitate.

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.

HiProS is a clearly written and practically motivated framework for text-guided medical segmentation. Its main contribution is a stage-wise strategy that routes Findings/Impressions and selected tokens to decoder levels, with SPS and FAM enabling more explicit local–global alignment. The submitted results on QaTa-COV19 and MosMedData+ are strong, and the ablations support the value of hierarchical text, token selection, and fusion. The rebuttal addresses key concerns by clarifying the distinction from IMITATE: HiProS applies hierarchical report semantics to decoder-stage dense segmentation with token-level routing, rather than hierarchical VLP pretraining. This distinction is sufficient, though the novelty remains incremental. The authors also clarify generated-report reliability, information leakage, computational cost, and performance attribution using the same LViT text. Limitations remain: evaluation is restricted to pulmonary infection datasets, token-routing evidence is mostly qualitative, and generated reports may hallucinate. These should be discussed, but the empirical quality and practical contribution justify acceptance.

Meta-review #3

After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

Reject
Please justify your recommendation.

This paper proposes HiProS, a hierarchical prompt selection framework that aligns Findings and Impressions with different decoder stages for vision-language medical image segmentation. The paper is well written and the empirical results are strong.

However, after rebuttal, the fundamental novelty concern remains unresolved. Reviewer #4 convincingly argues that the core idea of exploiting the Findings/Impressions hierarchy and aligning it with visual abstraction levels was already substantially explored in IMITATE. The authors’ distinction rests primarily on applying this idea at the decoder level for segmentation rather than in pre-training, which is an implementation-level adaptation rather than a clearly new conceptual contribution. The assignment of Findings to shallow layers and Impressions to deep layers, while intuitive, also lacks rigorous quantitative validation at the level of encoded text representations.

With Reviewer #2 supporting acceptance, Reviewer #1 remaining cautiously positive but not explicitly endorsing acceptance, and Reviewer #4 recommending rejection on substantive novelty grounds, the evidence does not clearly favor acceptance. I recommend reject. The authors are encouraged to develop a more clearly differentiated conceptual advance over IMITATE and to provide stronger quantitative token-routing evidence for the hierarchical assignment before resubmission.

back to top

HiProS: A Hierarchical Prompt Selection for Stage-wise Vision–Language Alignment in Medical Image Segmentation

Author(s):