Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Automated diagnosis of 3D brain CT scans is essential for critical care, yet it remains challenging due to the heavy reliance on manual annotations and the limited semantic understanding of conventional models. While 2D foundation vision-language models (VLMs) have shown strong generalization, adapting their slice-level representations for 3D volumetric analysis remains challenging. In this paper, we propose Brain-Adapter, a novel dual-stream multiple instance learning (MIL) framework that leverages pre-trained 2D biomedical VLMs and raw diagnostic reports for robust scan-level multi-label classification. Specifically, we introduce a Text-Conditioned Attention (TCA) mechanism, utilizing raw diagnostic sentences as semantic queries to dynamically align visual cues with specific disease concepts. Concurrently, a parallel visual MIL stream captures global scan characteristics, supervised by structured labels extracted via a Large Language Model (LLM). To ensure representation coherence, a consistency constraint enforces synergy between the two streams. During inference, an Uncertainty-Aware Refinement (UAR) module dynamically calibrates and fuses these dual-stream predictions to resolve ambiguous cases. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art 3D models and standard MIL approaches. By eliminating the reliance on dense annotations, Brain-Adapter offers an annotation-efficient framework for 3D acute intracranial pathology analysis. Code is available at https://github.com/Scatteredrain/Brain-Adapter.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4574_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Scatteredrain/Brain-Adapter

Link to the Dataset(s)

CQ500 dataset: https://github.com/tomboulier/cq500-downloader

BibTex

@InProceedings{YiZhe_BrainAdapter_MICCAI2026,
        author = { Yi, Zhenyu AND Song, Zhiyun AND Sun, Yusong AND Liu, Zelin AND Fei, Manman AND Li, Zhenhao AND Zhao, Jiaxuan AND Han, Xu AND Zhang, Lichi},
        title = { { Brain-Adapter: A Dual-Stream Vision-Language MIL Framework for Comprehensive 3D CT Diagnosis of Acute Intracranial Pathologies } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    Adapting the 2D biomedical visual language model BiomedCLIP to 3D NCCT volumetric data; the approach does not rely on dense annotations but instead uses raw radiology reports as the supervision source.

  • 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.

    First, the problem is significant and clinically relevant; CT triage for acute intracranial lesions is inherently a high-value task. Second, the methodology flows logically, with a coherent connection between TCA, ABMIL, consistency constraints, and UAR. Third, the use of raw reports rather than densely annotated data ensures practical scalability.

  • 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 structured labels are extracted by an LLM, but their quality is not validated. The method section explicitly states that the structured supervision labels come from LLM-driven logic-set extraction and serve as the ground truth for the supervised branch. If the LLM misses labels, introduces incorrect labels, or over-standardizes the free-text impressions when extracting the 21 classes, then the supervision itself becomes unreliable. 2.The report text is used both for training supervision and semantic alignment, creating a potential shortcut through the text modality. TCA uses the raw report and its fine-grained sentences as text features, while the logic-set supervision is also derived from the same report. If the model relies too heavily on the explicit diagnostic semantics contained in the report rather than on the image itself, this may create a weak form of leakage in which the report is effectively supervising itself. 3.The comparison against stronger 3D medical classification baselines is insufficient. The only 3D models in Table 1 are ViT-B and Swin-B, which may not represent the strongest current baselines for 3D medical classification. 4.In addition, the performance gap between the trained-from-scratch and pre-trained models needs further explanation. Why do some pre-trained models not outperform their from-scratch counterparts after initialization with pre-trained weights? 5.The explanation of the ablation results is not fully consistent with the actual numbers. In Table 2, adding the consistency constraint worsens the Hamming Loss, yet the paper states that the “sequential integration of components consistently drives performance upward. ” This is not true for all metrics. 6.Attention heatmaps do not automatically imply reliable interpretability. Without quantitative localization evaluation or expert assessment from clinicians, the interpretability claim remains weak. 7.The label definition mixes pathological findings with etiological context, making the task boundary less clean. If the label set contains both imaging findings and etiological context, the task formulation becomes semantically mixed, and some labels may not be reliably inferable from the image alone.

  • 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.

    (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?

    In summary, the paper is promising, but several key issues remain insufficiently addressed, including the reliability of LLM-extracted labels, potential text-driven leakage, limited 3D baseline comparisons, and overclaimed conclusions from the ablation and interpretability analyses. Overall, the current evidence is not yet strong enough to fully support the paper’s claims.

  • 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.

    1.The proposed “Open-set Alignment” module appears conceptually closer to a combination of cross-attention and contrastive learning, rather than a genuine open-set recognition framework. While the design is reasonable and empirically motivated, the terminology may somewhat overstate the methodological novelty of this component. 2.The paper lacks sufficient implementation and engineering details to ensure reproducibility. In particular, the construction process of the fine-grained pathological descriptions is not clearly specified, and several important training settings (e.g., prompt formulation, LoRA configuration, optimization details, and logic-set extraction pipeline) remain unclear. In addition, the manuscript does not mention whether the source code will be publicly released, which further limits reproducibility and practical verification of the proposed method.



Review #2

  • Please describe the contribution of the paper

    This paper introduces Brain-Adapter, a dual-stream multiple instance learning (MIL) framework designed to adapt pre-trained 2D biomedical vision-language models (VLMs) for the 3D diagnosis of acute intracranial pathologies. The framework consists of a Text-Conditioned Attention (TCA) mechanism for open-set alignment and a parallel visual MIL stream for logic-set supervision. Furthermore, the authors propose an Uncertainty-Aware Refinement (UAR) strategy to dynamically fuse dual-stream predictions during inference. The method is evaluated on an in-house clinical dataset of 852 non-contrast head CT scans and an external CQ500 dataset for zero-shot transferability.

  • 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.

    Intuitive and Promising Methodology: The overall design of the Brain-Adapter framework is highly intuitive and demonstrates significant potential. The dual-stream architecture elegantly balances explicit semantic guidance with implicit visual evidence.

    Strong Interpretability: The method provides excellent interpretability. The Text-Conditioned Attention (TCA) mechanism can effectively localize the specific slices manifesting the lesions. Even in cases of misclassification, the ability to pinpoint the abnormal slices holds substantial clinical value.

    Innovative Inference Strategy: The Uncertainty-Aware Refinement (UAR) mechanism is a clever design that dynamically calibrates visual predictions with text-guided evidence only when the visual discriminative logic is ambiguous.

  • 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.

    Asymmetry Between Claims and Experimental Validation: The paper’s title and introduction position it as both a “MIL framework” and an “Adapter” for 2D VLMs to 3D tasks. However, the experimental design only validates the former. While comparing against other MIL aggregation methods (e.g., ABMIL, TransMIL) is necessary and sufficient for the “MIL framework” claim, the paper lacks specific architectural designs tailored for efficient fine-tuning (aside from simply applying LoRA). The framework essentially functions as a novel training paradigm rather than a dedicated adapter.

    Insufficient Baselines for VLM Adaptation Claims: Because the authors claim to solve the “open problem” of “effectively transferring their [2D VLMs] representational power to 3D volumes,” the experimental baselines are incomplete.

    • To support the “Adapter” claim, the authors should compare Brain-Adapter with other parameter-efficient fine-tuning (PEFT) methods based on the same pre-trained model (e.g., CLIP-Adapter, CoOp, VPT).
    • More importantly, the effectiveness of the adaptation should be validated on more pre-trained models, rather than solely relying on BiomedCLIP. Crucially, there are already several foundation models specifically trained for the CT modality (e.g., recent 3D Head CT and CTA foundation models). Failing to conduct experiments on these domain-specific foundation models significantly weakens the claim that this framework is a generalized solution for 3D CT diagnosis.

    Incomplete Quantitative Evaluation for Interpretability: In Figure 3, presenting only the slice aggregation weights and qualitative visualizations is insufficient. The authors should provide quantitative results regarding the accuracy of slice-level localization to substantiate the interpretability claims.

    Unsupported Claims on Scalability: Both the abstract and conclusion highlight the method’s “scalability.” However, there are no experiments in the paper to support this claim (e.g., performance comparisons under 1%, 10%, and 100% data regimes).

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I personally appreciate the proposed methodology; it is intuitive, interpretable, and holds clinical potential. However, the current manuscript suffers from a critical asymmetry between its stated contributions and its experimental validation. While it successfully demonstrates superiority over traditional MIL frameworks, it fails to provide sufficient evidence for its claims as a generalized VLM adaptation method. The lack of comparisons with other PEFT methods and the failure to validate the framework on CT-specific foundation models make the core “Adapter” contributions unconvincing. Additionally, unsupported claims regarding scalability and missing quantitative evaluations for slice localization further weaken 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.

    I thank the authors for the detailed rebuttal. The clarification of the “adapter” scope, together with the added PEFT/FM comparisons and LLM-label validation, satisfactorily addresses my main concerns. Some limitations remain, especially on localization and broader generalization, but these are understandable given the MICCAI format and rebuttal constraints. I recommend acceptance, provided that the final version carefully qualifies claims on generalization, scalability, and interpretability.



Review #3

  • Please describe the contribution of the paper

    1.Innovatively proposes a case-level multi-label diagnostic approach that leverages a pretrained vision–language model (VLM), addressing the issue that knowledge learned from 2D pretraining cannot effectively support 3D tasks.

    2.Adopts a weakly supervised training strategy that relies only on paired radiology reports, and reduces the dependence on dense manual annotations through contrastive learning.

    3.Achieves alignment between image slices and the corresponding report text via contrastive learning and the proposed Text-Conditioned Attention (TCA) module.

    4.Proposes UAR, an uncertainty-based inference enhancement method, which improves the model’s prediction accuracy.

  • 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.Author converts free-text impressions into usable weak supervision (via LLM-based structuring) and also leverages report sentences for alignment training, reducing reliance on dense or expensive manual annotations 2.By jointly training a text-alignment stream and a visual MIL stream (with a consistency constraint), author combines open-vocabulary semantic grounding with stable discriminative learning, improving robustness compared to using either one alone

    3.Author introduces text-conditioned attention to retrieve disease-relevant slice evidence and uses it to regularize training and calibrate uncertain predictions, rather than relying solely on unguided visual pooling

    4.Author provides additional zero-shot experiments

  • 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 paper does not specify which LLM is used for label extraction, and the lack of prompt details reduces reproducibility. 2.It would be helpful to report the accuracy of the LLM-extracted labels (e.g., via a small-scale manual verification). 3.The authors are encouraged to provide inference-time statistics, such as runtime/latency (and related efficiency metrics). 4.The paper does not explicitly commit to open-sourcing the code or models

  • 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?

    1.The authors successfully adapt a pretrained 2D vision–language model (VLM) to a 3D medical imaging task, demonstrating methodological novelty. 2.The paper provides extensive comparative experiments, including comparisons against both 2D-based approaches and native 3D models. 3.The additional zero-shot evaluation experiments are appreciated and satisfactory. 4.Unfortunately, the paper does not specify the exact LLM used or the prompting details, and it also does not make an explicit commitment to open-source the code/models

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    It solved my problem very well



Author Feedback

We thank the reviewers for recognizing our work as “clinically relevant” (R2), “promising” (R3), and “novel” (R4). We address key concerns below:

@R2-Q1, R4-Q1/2/3/4 LLM usage and implementation details: For LLM-based label extraction, Qwen3 is used in a constrained manner: mapping short diagnostic conclusions to predefined 0/1 taxonomy labels in a strict JSON format rather than open-ended generation. For label reliability, a neurosurgeon with >3 years of experience reviewed 100 random cases by comparing the LLM-extracted labels with the original reports, yielding 98% exact-match accuracy. For reproducibility, we will release the prompt, code upon acceptance, and report inference statistics (79.4 ms/sample latency; 1003.6 MB peak GPU memory).

@R3-Q1/2 Adapter/transfer scope: We clarify that Brain-Adapter focuses on adapting 2D VLM representations to 3D CT diagnosis via slice aggregation and dual-stream learning. Here, “adapter” denotes a dimensional and domain adaptation strategy (2D-to-3D), rather than indicating a universal PEFT module. We will highlight this definition in the revision to avoid confusion. As a preliminary study, we tested whether PEFT/VLM variants could explain the gain under the same ABMIL setting. BiomedCLIP+VPT/CLIP-Adapter scored 0.863/0.869 Micro-AUC, while our framework reached 0.887; with RadCLIP (Lu et al., TNNLS 2025), VPT/CLIP-Adapter scored 0.857/0.850, while our framework reached 0.879, supporting the advantage of our specific dual-stream design. We also appreciate the reviewer’s suggestion that turning the framework into a more modular adapter component could improve practical use, which we will explore in future work, including compatibility with alternative pre-trained models.

@R2-Q3, R3-Q2 Stronger 3D alternatives: Our work focuses on adapting existing 2D medical VLMs to 3D-aware understanding, rather than exhaustively benchmarking native 3D architectures. As a preliminary check, the recent 3D head FM, HLIP (Zhao et al., TMLR 2026), yielded 0.857 Micro-AUC vs. our 0.887.This suggests that direct 3D modeling of thick-slice CTs may dilute subtle slice-level cues, whereas 2D VLM+MIL preserves and aggregates them effectively.

@R2-Q6, R3-Q3 Interpretability evaluation (R2,R3): We thank the reviewer for highlighting the need for quantitative localization evidence. Our explainability analysis demonstrates that attention concentrates on informative middle slices, suppressing blank z-axis extremities. As a quantitative localization check, a neurosurgeon inspected 100 samples; the model achieved an 87% Top-5 accuracy in locating the manually identified key slice (out of 32 slices). This supports our slice-level localization.

@R2-Q2 Role of report text in model training: We clarify that report-derived labels and report text guidance are separated in our framework. First, logic-set labels supervise the visual-only MIL branch, while text embeddings are excluded from the classifier. Second, TCA uses text only to guide visual slice pooling during training, not to predict labels. Third, the consistency loss is designed to distill this text-guided prior into the visual branch, as supported by our ablation study. In summary, inference remains strictly CT-only, and text serves as training-time semantic guidance for evidence aggregation rather than a shortcut.

@R2-Q4 Pretrained vs. from-scratch behavior: Table 1 shows that pretraining consistently improves threshold-free metrics such as AUC/Hamming loss, while fixed thresholds may shift SEN/SPE in imbalanced multi-label prediction. We will clarify this in the revision.

@R2 R3 Manuscript clarifications: We will revise Table 2 analysis to avoid monotonic-gain claims (R2-Q5), define the taxonomy as a clinically oriented report-derived label space covering both imaging findings and diagnostic context (R2-Q7), and replace unsupported scalability claims with a narrower annotation-efficient formulation (R3-Q4).



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.

    I recommend Invite for Rebuttal. Reviewer 1 was positive overall, but questioned the quality of the LLM-extracted supervision labels and the possibility of text-driven leakage. Reviewer 2 was negative and argued that the paper’s transfer and adapter claims are broader than what the experiments actually validate. Reviewer 3 was supportive, while still noting missing implementation details and the need to verify the extracted labels more carefully. I also share these concerns, as many important details remain insufficiently described, particularly the LLM-based label extraction process, label validation, and implementation details.

  • 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 substantially improves confidence, the authors report manual validation of extracted labels, clarify that inference is CT only, provide additional PEFT and foundation-model comparisons, and add quantitative slice-localization evidence.



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.

    Reviewers #3 and #4 agreed to accept this manuscript given its merit and rebuttal, while reviewer #2 found several overclaimed points in the paper. The authors are strongly encouraged to revise their claims as per reviewer #3’s suggestions.



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 remaining concerns are about novelty framing and reproducibility. The first issue is valid, but not invalidating in my opinion. For the second concern, the authors promised to share implementation details in the project repository. I recommend acceptance, given that the authors release the code and all processing details.



back to top