Abstract

Intracranial aneurysms are prone to rupture, making accurate delineation crucial for rupture-risk stratification and treatment planning. While CTA is routinely used for lumen geometry, high-resolution vessel wall imaging (HR-VWI) directly visualizes aneurysm wall enhancement (AWE), providing a wall-focused biomarker that supports wall evaluation and rupture-risk assessment. However, aneurysm borders on HR-VWI are often ill-defined and easily confounded by imperfect blood suppression and parent-vessel distraction, thereby obscuring true boundary cues and compromising precise delineation. Vision–language models (VLMs) can introduce semantic priors via text guidance, yet naive dense fusion is frequently dominated by background structures, diluting weak aneurysm cues. To address these challenges, we propose Frequency-Aware Prototype Learning in VLM for intracranial aneurysm HR-VWI segmentation. Our method leverages discrete wavelet transform to decompose multi-scale features and emphasize boundary-related high-frequency cues, coupled with text-conditioned filtering to retain lesion-relevant frequency responses for query updates. Concurrently, we summarize multi-scale representations into compact prototypes, enabling selective interactions between text-guided queries and foreground–context prototypes to suppress vascular distractions and sharpen delineation. Experiments on an in-house HR-VWI aneurysm benchmark demonstrate consistent improvements over strong baselines, achieving 79.18% Dice. Size-stratified evaluation further confirms robust performance across aneurysm scales, with particularly clear gains for small lesions, supporting more reliable aneurysm quantification for clinical decision-making. Code and test set will be released at https://github.com/wwyx2722-boop/FAPL-VLM

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2555_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/wwyx2722-boop/FAPL-VLM

Link to the Dataset(s)

In-house HR-VWI aneurysm test set: https://github.com/wwyx2722-boop/FAPL-VLM

BibTex

@InProceedings{WanYix_FrequencyAware_MICCAI2026,
        author = { Wang, Yixin AND Wang, Haipeng AND Feng, Xin AND Liao, Xiangyun AND Si, Weixin},
        title = { { Frequency-Aware Prototype Learning in VLM for Intracranial Aneurysm HR-VWI Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a VLM-based framework for intracranial aneurysm segmentation in high-resolution vessel wall imaging addressing the challenges of weak boundaries and class imbalance. The core contribution lies in integrating frequency-aware feature enhancement using “discrete wavelet transform” with a prototype-guided “attention mechanism” that selectively focuses on lesion-relevant regions while suppressing background interference. The work introduces a new HR-VWI aneurysm dataset (not accessable) and demonstrates improved segmentation performance across multiple benchmarks, particularly for small lesions .

  • 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 paper addresses a clinically important and challenging problem, namely aneurysm segmentation in HR-VWI, which is more difficult yet more informative than conventional imaging. The proposed method is conceptually well-motivated, combining frequency-domain analysis with language-guided semantic priors, which reflects a thoughtful integration of complementary cues. The introduction of the dual prototype bank and contrastive gating mechanism is a notable design that explicitly models foreground–background separation, improving robustness under complex anatomical conditions. The experimental evaluation is good including comparisons with other baselines, cross-dataset validation on benchmarks, and ablation studies that demonstrate the contribution of each module. The introduction of a new dataset (not accessable) adds value to the community and supports the clinical relevance of the work.

  • 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 paper has several limitations that need to be addressed by author(s).

    1) The performance gains over strong baselines are relatively modest, raising questions about the practical significance of the proposed complexity.

    2) the evaluation relies heavily on an in-house dataset, which may not be sufficient to demonstrate robustness or generalizability in real clinical settings. Also, without sharing the dataset, reproduction of the reported results is not possible.

    3) Although the method leverages vision–language modeling, the role of textual information appears relatively simple (short prompts), and the paper does not deeply analyze how language contributes beyond standard conditioning. Text-guided model enhancement is not new and have been discussed in several earlier studies. What is the originality here?

    4) The methodological novelty is somewhat incremental, as both frequency enhancement and prototype-based learning have been explored in prior segmentation literature, and their combination may not constitute a fundamentally new paradigm.

    5) The paper lacks detailed analysis of computational cost, inference efficiency, and clinical usability, which are critical for real-world deployment.

    6) Figure quality is low and make it difficult for me to judge. For example, Fig. 3 illustrate many results in limited space which is difficult to observe differences.

  • 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?

    While the results looks interesting, originality is limited and reproduction is difficult as it mainly based on in-house dataset. No declaration on possibilities of sharing the dataset.

  • 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.

    I am satisfied with author with author’s rebuttle



Review #2

  • Please describe the contribution of the paper

    This paper proposes a VLM-based framework for Intracranial Aneurysm HR-VWI Segmentation, key contributions: 1.Text-guided Frequency Enhancement Module (TFEM) injects language-consistent high frequency cues to enhance weak boundaries

    1. Dual Prototype Bank and Contrastive Gating Mechanism(DPBM) leverages foreground and context proto types for gated, lesion-focused query refinement. 3.Introduce a new HR-VWI intracranial aneurysm dataset - HiRes-Aneurysm (HiRA).
  • 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 introduction of a new dataset (83 patients, 2,182 images) is a valuable contribution, as it supports further research in this domain. 2.The study provides a comprehensive evaluation across three datasets and multiple baselines, including detailed analysis by lesion size—particularly for small aneurysms—which is commendable. Additionally, the inclusion of thorough ablation studies effectively demonstrates the contribution of each component.

  • 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.

    I have concerns regarding the motivation and effectiveness of the two proposed components, TFEM and DPBM. Although the idea of incorporating vision-language models (VLMs) and attention mechanisms into this medical imaging task is interesting, the paper does not convincingly justify why text guidance would be beneficial for boundary (i.e., high-frequency) reconstruction and fine-tuning. Furthermore, the ablation study raises additional doubts: the variant “TFEM w/o Text” performs even worse than the baseline without any additional module, and the overall performance gains remain marginal across evaluation metrics.

  • 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 dataset contribution and comprehensive evaluation make this work a valuable addition to intracranial aneurysm HR-VWI segmentation. However, the overall impact of the paper depends on whether the aforementioned concern, particularly regarding the justification and effectiveness of the proposed components, can be adequately addressed.

  • 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.

    I recommend a Weak Accept based on the following:

    1.Statistical Validity: The authors provided paired Wilcoxon tests, proving that the marginal improvements are statistically stable and not due to random noise. 2.Data Contribution: The HIRA dataset is a significant contribution to the HR-VWI field, which suffers from extreme data scarcity. The commitment to release the code and test set further supports its value to the community. 3.Methodological Note: The rebuttal clarified that TFEM is highly dependent on text to constrain frequency noise. While this reveals a lack of intrinsic robustness in the module, it is a pragmatically effective strategy for the specific challenges of HR-VWI.

    Required Revision: The final manuscript must include the rebuttal’s analysis regarding the module’s sensitivity to text guidance in the “Limitations” section to avoid misleading future research.



Review #3

  • Please describe the contribution of the paper

    The authors propose a novel intracranial vessel segmentation pipeline using HR-VWI, with a new dataset built for this modality. HR-VWI is a different imaging technique then traditional MRA or CTA, focusing on the vessel wall and not on the lumen. Conversely, the paper still focus on lumen (volumetric) segmentation. Segmentation directly on HR-VWI could inform rupture risk assessment.

    The authors propose two main modules: DPBM uses cosine alignment to produce foreground (text aligned) and context (text “distant”) features. TEFM performs frequency aware cross attention by including frequency decompositions of the image as additional “key values”.

  • 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.
    • In general the methodology is fascinating, with ideas that align with the data modality in question. In general the paper is well written and presentation/formatting is good, pending improvements to the architectural figure clarity. The method appears to perform well against both strong baselines and recent multimodal approaches.

    • The authors promise to make code available if accepted. What about making the novel in-house data available? This would be a very valuable asset for the field if publication were possible.

  • 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.
    • Figure 2 has some undefined symbols, and the cross-attention blocks have only one input. Usually, when doing cross-attention, you want to highlight exactly who is Q and who is K V. From the manuscript text, K, V are the transformed multi-scale features? Additionally, both modules (a) and (b) output “Qout”, and it seems Qout is summed and replicated (Q=Q, K=Q, V=Q) in self-attention before the prediction head, where SAM-style segmentation works. I assume the final (NxC) “Qout” is weighting the multiscale features for the final mask. It would benefit understanding if the Figure clarity was improved for it to be self-contained, making all symbols and variables understandable without needing to look for variable definitions across the rest of the manuscript.

    • Another major weakness is the lack of statistical validation. Ablation studies show small improvements from the two proposed modules, but these may not be statistically significant. In addition, the baseline “without any module” is still very good against the other methods, which is a little bit strange; maybe the authors could clarify why this happens.

    • The authors point out that: “A clear and consistent trend can be observed across all benchmarks: methods without text guidance exhibit a pronounced performance gap compared to text-guided models, indicating that textual priors provide complementary semantic cues for medical image segmentation.” This is a strong claim that needs more data to back it up. There is no standard deviation and statistical analysis to support this claim.

  • 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 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?

    The manuscript does not contain any statement regarding ethical approval for the dataset acquisition and use in this research. Because this study utilizes human medical imaging data (HR-VWI), an explicit statement confirming Institutional Review Board (IRB) or local Ethics Committee approval is expected.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
    • The manuscript should mention cosine alignment in the DPBM text rather than simply using “cos”. MHCA is also not defined (multi-head cross-attention?).
  • 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?
    • While the underlying methodology is elegant and the clinical motivation (informing rupture risk assessment) is sound, the manuscript suffers from unclear architectural diagrams and a lack of statistical validation for its core claims. Addressing these issues would significantly strengthen the paper.
  • Reviewer confidence

    Very confident (4)

  • [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 addressed all of the weakness I and others pointed out, providing statistical analysis, clarifying ethical issues and commenting on an improved Figure.



Author Feedback

Thank you for the constructive feedback and positive remarks. We appreciate the reviewers’ recognition of our work as a “methodological contribution” (R1–R3), with a clinically important HR-VWI segmentation setting (R1/R3), a valuable HiRA dataset (R1/R2), comprehensive evaluation (R1/R2), and a modality-aligned design that performs well against strong baselines (R1/R3). Below, we address the specific concerns raised. Common Questions: Code: We will release our code upon acceptance. Dataset: First, we have to clarify HiRA has the local IRB/EC approval, and we will add the corresponding ethics statement in the revised manuscript. To support test, we will release our test dataset and plan to share an expanded dataset in future extension work. (R1 W1) Due to the difficulty of HR-VWI aneurysm annotation, the current size of HiRA is limited, which restricts the observed performance gains. We are continuously expanding HiRA for more comprehensive validation. Regarding the statistical significance of the proposed modules, case-level paired Wilcoxon tests with Holm-Bonferroni correction show that our method outperforms the no-module Baseline and the best single-module variant on all datasets/metrics (corrected p<0.05), demonstrating the effectiveness of TFEM and DPBM. TFEM with Text also outperforms TFEM w/o Text on most metrics (corrected p<0.05), confirming the benefit of text guidance. (R1 W2) Please refer to the common questions. (R1 W3) Text guidance is designed for the specific challenges of HR-VWI aneurysm segmentation, rather than serving as a simple class prompt. In HR-VWI, aneurysms are often small with weak boundaries, sparse foregrounds, and confusing high-frequency responses from artifacts, parent-vessel edges, or background textures. To address this, we use aneurysm-specific text as a semantic prior for feature selection: TFEM filters frequency-enhanced tokens toward aneurysm/wall-relevant boundary cues, while DPBM ranks prototypes to build more discriminative foreground/context banks. We performed an ablation study to verify this design. Compared with prior text-guided methods that mainly use text for general conditioning, our aneurysm-specific HR-VWI text improves Dice by +1.3, +1.1, and +1.5 points for TFEM, DPBM, and the full model, respectively, with corrected p<0.05.(R1 W4) Please see (R1 W3) (R1 W5) Our model has 215M parameters and 1.236 TFLOPs, compared with 182.16M parameters and 3.29 TFLOPs for CausalCLIPSeg, making it feasible for practical HR-VWI aneurysm analysis. (R1 W6) We will provide more clear explanations for the existing figures. For Fig. 2, we will clarify the symbols and their relationships: in DPBM, Q is the segmentation query and K/V are text-ranked foreground/context prototype banks; in TFEM, Q is the segmentation query and K/V are text-filtered frequency features. We will also specify that MHCA denotes multi-head cross-attention and replace “cos” with cosine alignment. For Fig. 3, since HiRA is designed for aneurysm analysis, we visualize prediction errors on aneurysm and vessel-wall surfaces to show the spatial relationship between aneurysm segmentation and wall assessment. This demonstrates that our method provides more reliable wall/aneurysm delineation for practical HR-VWI analysis. (R2 W1) Please see (R1 W3) (R2 W2) As discussed in (R1 W3), TFEM relies on text-guided feature selection. Without text, added frequency features become unconstrained and may introduce redundant cues from artifacts, parent-vessel edges, or background textures, thereby impacting segmentation. (R3 W1) Please see (R1 W6) (R3 W2) Please see (R1 W1) and (R1 W3) (R3 W3) Please see (R1 W1) and (R1 W3) for the significance analysis and text-guidance explanation. We will revise the description as suggested.




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.

    This paper proposes a VLM-based framework for intracranial aneurysm segmentation in HR-VWI, combining frequency-aware feature enhancement (TFEM) and a dual prototype bank (DPBM) with a new in-house dataset. The clinical motivation is sound and the evaluation covers three datasets.

    All three reviewers converge on a key worry: the improvements from the proposed modules are small and may not be statistically significant. R2 noted that “TFEM w/o Text” actually performs worse than the baseline with no added module, which raises questions about whether the text component is helping or hurting. R3 independently observed that the baseline without any proposed module is already competitive with other methods, and flagged the absence of statistical tests. R1 questioned whether the text guidance goes beyond standard conditioning and whether combining known components (frequency enhancement + prototypes) constitutes enough novelty. R3 also flagged a missing IRB/ethics statement for the in-house dataset, which the authors should address. For the rebuttal, important issues to address: 1.explain the “TFEM w/o Text” result and why it underperforms the no-module baseline, 2.provide statistical significance tests for the claimed improvements, 3.clarify data availability and ethics approval status, 4.justify why the combination of known techniques is non-trivial here. Points 1 and 2 are most important.

  • 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.

    All reviewers agreed that their concerns have been mostly addressed. I agree with R2’s comment that the new dataset and the demonstrated clinical efficacy outweigh the methodological weaknesses, and the authors are transparent about the module’s limitations, which should be encouraged for all MICCAI works to reduce overstated contributions in published work.



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 authors have well addressed the concerns of the reviewers, who have all indicated acceptance after the rebuttal.



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.

    Reviewers agreed that the work addresses a clinically important and technically challenging problem—aneurysm segmentation in HR-VWI—and introduces a valuable new dataset that helps mitigate data scarcity in this domain. They also acknowledged that the method is well-motivated, integrates complementary frequency-domain and language-guided cues, and is supported by comprehensive experiments, including cross-dataset validation and detailed ablation studies. However, they also raised serious concerns about the paper’s novelty, empirical significance, and methodological clarity: (1)​the performance gains over strong baselines are modest and the core components (frequency enhancement, prototype learning, text guidance) build largely on existing ideas, with limited demonstration of fundamentally new insights; (2) the reliance on an inaccessible in-house dataset undermines reproducibility, and the contribution of textual guidance is not convincingly isolated or analyzed; and (3) the manuscript lacks statistical validation, clarity in architectural illustrations, and discussion of computational efficiency and clinical usability. The rebuttal has successfully addressed most concerns, notably adding statistical testing, improving figure clarity, and committing to code and test-set release. Since the review scores are all positive after rebuttal, I recommend accepting this manuscript.



back to top