Abstract

Vision-Language Models such as CLIP enable effective few-shot medical anomaly detection (AD) via strong image-text semantic alignment. However, their globally contrastive pretraining lacks explicit spatial supervision, limiting precise lesion localization. In contrast, Vision Foundation Models (VFMs) such as DINO learn spatially coherent patch representations via self-distillation and local-to-global consistency, better capturing fine-grained anatomical structures. Leveraging this complementarity, we propose Spatial-FAD, a spatial-aware few-shot medical AD framework that improves lesion localization by combining VFM spatial priors with CLIP semantics. Specifically, we introduce a VFM-enhanced adapter that injects a structural affinity prior derived from DINO into CLIP features. This structure-guided refinement encourages visual embeddings to better adhere to lesion boundaries while maintaining semantic alignment. To address the loss of spatial detail from patchification and the limited input resolution of CLIP, we adopt a sliding-window aggregation strategy. This generates high-resolution, spatially dense embeddings to further enhance localization granularity. Moreover, we introduce a prototype-enhanced support memory scheme to efficiently exploit the few-shot support set. This module stores compact prototypes for normal and abnormal patterns, reducing memory costs while boosting performance by fusing patch-to-prototype and image-text similarities. Extensive experiments on three benchmark datasets, including Liver CT, Retinal OCT, and Brain MRI, demonstrate that Spatial-FAD significantly outperforms state-of-the-art methods, especially in lesion segmentation. Notably, in the 4-shot scenario, our method achieves an average improvement of over 11.4% in Dice score and 1.8% in AUC. Code is available at: https://github.com/JuzhengMiao/Spatial-FAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3052_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/JuzhengMiao/Spatial-FAD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MiaJuz_Bridging_MICCAI2026,
        author = { Miao, Juzheng AND Yuan, Yuchen AND Chen, Cheng AND Heng, Pheng-Ann},
        title = { { Bridging Vision Foundation Model Priors with CLIP for Spatial-aware Few-shot Anomaly Detection in Medical Images } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a VFM-enhanced adapter with structural affinity prior (VFMEA) and Sliding-window Aggregation(SWA) strategy as two key contributions. VFMEA is proposed to combine CLIP’s semantic alignment with DINO-based structural priors. SWA is introduced to alleviate the loss of fine spatial details caused by CLIP’s limited input resolution and pachification process. Furthermore, the paper also includes a prototype-enhanced support memory scheme to better utilize few-shot support examples during inference.

  • 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 well organized, and the proposed method is well motivated. 2.This paper addresses an important limitation of CLIP-based few-shot medical anomaly detection, namely weak spatial grounding despite strong semantic discrimination. 3.The proposed method shows consistent performance gains on three public datasets across multiple shot settings.

  • 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.Limited novelty: While the overall framework is reasonable and effective, the novelty of the proposed component appears moderate. The Sliding-Window Aggregation is conceptually close to prior window-based CLIP methods such as [1]. The authors should clarify more clearly what is conceptually novel beyond adapting this idea to the current medical setting. In addition, fusing CLIP and DINO features such as [2] have been popularly proposed recently. Authors should clarify their technical novelty compared to other existing methods. 2.The practical trade-offs are not discussed enough. Since the proposed method relies on overlapping local windows and additional support-memory comparison, more explicit discussion of inference cost, efficiency, and scalability would strengthen the practical significance of the method. [1] Jeong, Jongheon, et al. “WinCLIP: Zero-/Few-shot anomaly classification and segmentation. ” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [2] Bi, Xiuli, et al. “SSR: Semantic and Spatial Rectification for CLIP-based Weakly Supervised Segmentation. “ Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40.No. 4.2026.

  • 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 paper is solid overall, but some weaknesses regarding novelty and practical significance could be further clarified.

  • 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 appreciate the authors’ clarifications and additional analyses. The response addresses my main concerns regarding the distinction from SSR, the role of SWA, and the efficiency/scalability of the proposed method. However, the core novelty still appears moderate, and some key analyses should ideally have been included in the original manuscript. Therefore, I will maintain my original score.



Review #2

  • Please describe the contribution of the paper

    This paper integrates DINOv3 spatial priors into CLIP-based few-shot medical anomaly detection to address the spatial coarseness of CLIP’s patch-level representations. Combined with a sliding-window aggregation strategy for higher-resolution feature extraction and a prototype-based support memory scheme, the proposed framework demonstrates improved lesion localization across three medical imaging 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.Clear problem motivation. The paper identifies a concrete and well-known limitation of CLIP-based anomaly detection — poor spatial grounding — and directly targets it with each proposed component, giving the work a coherent narrative.

    2.Strong empirical gains. The improvements are substantial, particularly on LiverCT where Spatial-FAD outperforms the next best method by over 19% Dice across all shot settings, suggesting the approach addresses a genuine bottleneck.

    3.Comprehensive evaluation. Experiments span three modalities (CT, OCT, MRI) and four shot settings, providing a thorough empirical picture across diverse clinical scenarios.

  • 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.Missing evaluation of DINOv3 as a standalone anomaly detector. The paper leverages DINOv3’s spatially coherent patch representations as a structural prior but never evaluates it directly on the anomaly detection task. Given its boundary-aware local-to-global training, it is non-trivial whether DINOv3 alone is already effective for lesion localization via patch-level matching or affinity-based scoring. Such an evaluation would provide important insight into how well DINOv3 transfers to medical anomaly detection in its own right.

    2.Sliding-window aggregation is not applied to competing methods. Sliding-window-based feature extraction has been explored in prior anomaly detection literature [1], yet this connection is not discussed. In this work, SWA is the largest single contributor to performance (+17% Dice), but it is applied only to Spatial-FAD and not to competing CLIP-based methods such as MVFA or MadCLIP. Since SWA is a generally applicable strategy for improving spatial resolution, its exclusive use makes it difficult to determine whether the observed performance gains arise from the full Spatial-FAD design or primarily from this resolution enhancement. Applying SWA to strong baselines would help clarify the contribution of the proposed components.

    3.Figure 3(b) and 3(c) are not clearly labeled. The caption does not specify which panel corresponds to normal versus abnormal prototypes. Based on x-axis ranges and implementation details this can be inferred, but should be explicitly stated to avoid ambiguity.

    4.Prototype-enhanced support memory is insufficiently justified. The three-score fusion uses a fixed 1/3 average with no learned weighting, assuming equal reliability across signals of clearly different quality. More critically, since MVFA uses only normal support images, the gain from PSM could be attributed entirely to incorporating abnormal examples rather than to the prototype design itself. A controlled ablation against a dense memory bank that also includes abnormal images is needed to isolate the prototype contribution.

    [1] Dong, Haoyu, et al. “SWSSL: Sliding window-based self-supervised learning for anomaly detection in high-resolution images.” IEEE Transactions on Medical Imaging 42.12 (2023): 3860-3870.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper presents a well-motivated combination of DINOv3 spatial priors with CLIP-based few-shot anomaly detection, and the empirical gains in lesion localization are convincing across three modalities. However, several concerns need to be addressed before the contribution can be fully assessed: the sliding-window component should be applied to competing methods to rule out confounding, DINOv3 should be evaluated standalone to clarify its individual contribution, and the prototype fusion design requires a more controlled ablation. If these experiments confirm that the DINO-CLIP integration is the primary driver of improvement, the contribution would be well justified.

  • 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 Spatial-FAD, a few-shot medical anomaly detection framework that improves lesion localization by combining CLIP-based semantic alignment with spatial priors from a vision foundation model. The method integrates a VFM-enhanced adapter to inject DINO-based structural affinity into CLIP features, a sliding-window aggregation strategy to produce higher-resolution spatial representations, and a prototype-enhanced support memory to better utilize few-shot examples during inference. Through this design, the paper addresses the known limitation of CLIP-based approaches in spatial grounding and demonstrates improved segmentation performance across multiple datasets and shot settings.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper addresses a clear and important limitation of existing CLIP-based few-shot anomaly detection methods, namely strong image-level discrimination but weak spatial localization.
    • The overall method is coherent and well motivated. The VFM prior, sliding-window aggregation, and prototype-based support memory are all aligned with the goal of improving lesion segmentation quality
    • The empirical results are strong, especially in terms of Dice, and the improvements are consistent across three datasets and multiple shot settings.
    • The qualitative results support the quantitative findings, showing clearer boundaries and more accurate localization.
  • 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 largest gain in the ablation appears to come from the sliding-window aggregation strategy, whereas the title and framing emphasize bridging VFM priors with CLIP. This makes the attribution of the main contribution somewhat less clear.
    • The paper does not provide enough discussion of the computational and memory overhead introduced by sliding-window inference, which is important for judging both practical value and fairness of comparison.
    • The paper would be strengthened by additional analysis showing why the proposed DINO-based affinity prior is preferable to simpler alternatives for spatial refinement or feature smoothing.
    • The main empirical gain is clearly in segmentation quality, while the AUC improvements are more modest and not uniformly best across all dataset/shot combinations. The claims around classification performance should therefore be phrased somewhat more carefully.
    • The ablation study is only conducted on LiverCT under the 4-shot setting. While informative, broader ablations across additional datasets or shot settings would make the conclusions about component effectiveness more convincing.
  • 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?

    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?

    At this stage, I lean positive on this paper. The problem is reasonably well motivated, and the method presents a coherent approach that combines structural priors, high-resolution feature aggregation, and support-set memory to improve spatial localization in few-shot anomaly detection. The empirical results are generally strong in terms of Dice and support the claimed improvements in segmentation quality My main reservations are about contribution attribution and computational cost. The largest gain appears to come from the sliding-window aggregation, while the role of the VFM prior is less clearly isolated. In addition, the paper does not sufficiently discuss the computational overhead of the sliding-window design. There is also some room for improvement in the completeness of the empirical analysis, including broader ablation validation and clearer positioning of the observed gains across different metrics. Overall, I find this to be a well-executed piece of work with meaningful empirical improvements. My final decision will depend on how these points are addressed in the rebuttal.

  • 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 thank the authors for the rebuttal. The response reasonably clarifies my main concerns regarding the role of SWA, the contribution of the DINO-guided module, and the computational cost of the proposed memory scheme. While broader ablations would still strengthen the paper, I find the main concerns sufficiently addressed. I maintain my positive recommendation.



Author Feedback

We are glad and encouraged that reviewers find our work “well-motivated” (All), “addresses an important limitation” (All), “strong empirical gains” (All), “solid” (R1), “comprehensive evaluation” (R2), and “supportive qualitative results” (R3). All the reviewers expressed positive ratings. Thanks for the constructive comments. Responses to the remaining concerns are as follows.

Q1 (R1): Difference from SSR. A1: Our method differs from SSR in two main aspects: (i) the role of DINO and (ii) whether CLIP features are directly enhanced. SSR uses DINO only as a post-hoc auxiliary cue by combining DINO affinity with CLIP affinity to suppress CAM noise, without improving CLIP features themselves. In contrast, we use DINO as a structural prior during feature learning to directly enhance CLIP representations, leading to stronger spatial localization for anomaly detection.

Q2 (All): Concerns on sliding-window aggregation (SWA). A2: (1) Despite SWA’s gains, the other two modules (DINO-guided fusion & PSM) constitute the core scientific contribution in this work, aiming at addressing two fundamental problems of few-shot anomaly detection: the lack of fine-grained local structural information for precise spatial localization and inefficient use of limited support data. These modules and SWA are highly complementary, as our core modules build robust and discriminative features that provide the foundation for SWA to further enhance spatial resolution and improve the final localization quality. (2) Notably, even without SWA, adding only the DINO-guided module to the baseline still brings a +15.85% Dice gain, showing that the improvement is not mainly driven by SWA. (3) For fair comparison, we also apply SWA to competing methods. With SWA, MVFA and MadCLIP achieve 48.19 and 50.61 Dice, respectively, still clearly below our 56.33, demonstrating the effectiveness and complementarity of our core modules.

Q3 (R1, R3): Efficiency and scalability of our method. A3: (1) For SWA, as stated in our manuscript, we concatenate all cropped sub-images along the batch dimension, obtaining whole-image embeddings in a single forward pass. This utilizes GPU parallelism and incurs only a marginal increase in inference time: < 0.025s on an NVIDIA A40 GPU. (2) For our memory scheme, the main overhead comes from storing the memory bank and computing patch-prototype similarities. Let d be the feature dimension, m^2/n^2 the patch number per image for MVFA/ours, and n_n/n_a the normal/abnormal prototype number in our memory scheme. Under the k-shot setting, memory size comparison: MVFA:kdm^2 vs Ours:(n_n+n_a)d. Similarity FLOPs per image: MVFA:2kdm^4 vs Ours:2d(n_n+n_a)*n^2.With k=4, n_n=50, n_a=4, m=17, n=37 (typical for MVFA and ours), our memory and FLOPs are < 5% and <23% of MVFA, respectively. Also, our design is independent of the shot number k, yielding better scalability.

Q4 (R2): More ablation studies on PSM. A4: Although adding abnormal images to a dense memory bank improves Dice from 53.65 to 55.38, it remains below our PSM (56.33) and incurs higher computation (See Q3). This confirms that the gain also benefits from the prototype design rather than merely adding abnormal data. For fusion weights, the fixed average already yields consistent gains, indicating our robustness without complex coefficient tuning. Additional optimization may further improve performance.

Q5 (R2): Evaluation of DINOv3 as a standalone anomaly detector. A5: When frozen DINOv3 is used alone for anomaly detection via support–query patch matching, it yields only 12.43% Dice on LiverCT (4-shot). This suggests that, in the medical domain, DINOv3 may not reliably establish cross-image patch correspondences without task-specific adaptation. In contrast, within-image affinities are more reliable and can be effectively exploited by our method to improve spatial localization.

Minor concerns like the Fig. 3 caption and careful claim of classification results will be carefully revised.




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 paper proposes a coherent framework for improving spatial localization in CLIP-based few-shot medical anomaly detection and demonstrates consistent empirical gains across multiple datasets and settings. Reviewers agree that the problem is important and the overall approach is well motivated, with strengths in experimental performance and clarity of presentation. At the same time, several substantive concerns are consistently raised across reviews, including the need to better clarify novelty relative to prior CLIP+DINO and sliding-window approaches, more clearly isolate the contribution of individual components (particularly the dominant effect of sliding-window aggregation), provide deeper analysis of key design choices (e.g., DINO priors and prototype design), and discuss computational cost and fairness of comparison .

    Given these points, the paper is forwarded to rebuttal. Authors are encouraged to address these concerns through careful clarification, additional analysis based on existing results, and improved positioning. In accordance with conference policy, no new major experiments should be introduced; instead, rebuttal should focus on explanation, analysis, and justification. Reviewers are requested to re-evaluate the paper in light of the rebuttal.

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

    Accept

  • Please justify your recommendation.

    The paper proposes a coherent framework for improving spatial localization in CLIP-based few-shot medical anomaly detection by combining vision foundation model priors with CLIP-based anomaly scoring. Reviewers recognized the importance of the problem, the practical relevance of spatial-aware anomaly detection, the generally clear presentation, and the consistent empirical performance across multiple datasets and shot settings. The initial concerns mainly related to attribution of gains, novelty relative to prior CLIP+DINO and sliding-window approaches, the role of the DINO-guided module and prototype memory design, and computational cost. The rebuttal clarified these points in a generally satisfactory manner and helped position the contribution more clearly. While the novelty remains moderate and some analyses could be strengthened in the final version, the paper presents a well-motivated and practically useful contribution. I therefore recommend acceptance. The authors should incorporate the rebuttal clarifications, more carefully discuss the contribution of each component, and explicitly acknowledge remaining limitations in the camera-ready version.



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.

    All three reviewers recommend acceptance post-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.

    The authors have provided solid and comprehensive responses in the rebuttal, satisfactorily clarifying the key technical and methodological queries raised during the review phase. Specifically, the authors have successfully articulated the core innovations of their framework over existing CLIP+DINO baseline. Crucially, they have provided compelling empirical evidence to quantify the independent contributions of individual modules; specifically, the authors demonstrated that incorporating the DINO-guided module into the baseline alone yields a 15.85% improvement in Dice similarity without relying on sliding-window aggregation, effectively proving that the performance gain is inherently driven by their core methodological design rather than the sliding-window mechanism. Furthermore, the deeper analyses regarding the DINO prior, prototype design, computational overhead, and baseline fairness have successfully resolved the remaining technical doubts.

    By rigorously untangling the driving forces behind the performance gains and justifying the theoretical depth of their foundation model integration, the paper delivers validated, high-impact insights to the community.

    In light of MICCAI policy that no new results can be added in the rebuttal phase, I have re-evaluated this paper based on its original content. The design works effectively (bringing large performance gains, as evidenced in the original manuscript), is technically sound, and the paper is easy to follow. Therefore, I recommend acceptance.



back to top