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

Abstract

Vision–language models trained on large-scale medical image–text data achieve strong zero-shot performance but typically rely on global, dense representations that are difficult to interpret and provide little to no visual grounding. We introduce a framework to extract and exploit sparse, local features from a medical vision–language model (MedSigLIP) while preserving text-aligned capabilities. First, we obtain local, text-aligned image embeddings using (1) distillation and (2) token alignment. We analyze these local embeddings using a sparse autoencoder, yielding sparse latents and show that they correspond to recurring, clinically meaningful image patterns. By ranking these latents using MedSigLIP’s text embeddings and selectively recombining latents back into the teacher embedding space, we construct new classification probes that focus on interpretable local evidence. We do not fine-tune MedSigLIP and thus retain text-aligned capabilities. Evaluated on VinDR-CXR with external testing on NIH ChestXray, our approach improves mean AUROC on VinDR by up to 0.05 and significantly increases pointing game localization recall by up to 9% on both datasets, across multiple attribution methods. These results demonstrate that sparse latent structure in distilled local features provides an effective mechanism for interpretable and clinically grounded reasoning in medical vision–language models.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1037_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/martingoe/sparse_local_latents

Link to the Dataset(s)

VinDR CXR: https://vindr.ai/cxr NIH: https://www.kaggle.com/datasets/nih-chest-xrays/data

BibTex

@InProceedings{GoeMar_Sparse_MICCAI2026,
        author = { Goetze, Martin AND Wienholt, Patrick AND Eschweiler, Dennis AND Gazibaric, Marvin AND Kuhl, Christiane AND Nebelung, Sven AND Truhn, Daniel},
        title = { { Sparse Local Latents for Explainable Zero-Shot Reasoning in Medical Vision-Language Models } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a framework that decomposes the internal model representations of a medical vision-language model while preserving its zero-shot capabilities. They achieve this by extracting sparse local features of the model and exploring features that are fired for a given diagnosis. This enables localisation of visual evidence for a pretrained vision-language model directly without further fine-tuning.

  • 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 classification pipeline shows that a vision-language model’s reasoning can be localised through sparse local evidence.
    2. The paper demonstrates that sparse local embedding corresponds to clinically meaningful image patterns.
    3. The paper show that sparse local features in vision-language models can enhance post-hoc interpretability while simultaneously supporting clinically zero-shot reasoning.
    4. The paper proposes an evaluation of whether high- and low-similarity latent for a given diagnosis tend to co-activate is novel to assess the coherence of the learned sparse representations.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. Section 2.2 introduces two approaches to constructing local embeddings, but the results section does not clearly compare them in terms of performance, interpretability quality or practical trade-offs. Since the paper introduces both variants, a direct comparison would strengthen the results.
    2. The paper states that post-hoc interpretability methods operate on dense features. However, there are post-hoc interpretability methods that do not necessarily need dense features to work, such as GradECLIP [1].
    3. The choice of SigLIP as the backbone is not sufficiently discussed in the paper.
    4. The figures in the paper lacks in-figure labels making readability difficult. Relying solely on captions makes the figures difficult to interpret.
    5. A statement of clarification in the second paragraph of the Introduction section: Foundations models are not always multi-model architectures. Most large language models are not multi-modal, eg. LLaMA [2]. References
    6. Zhao, C. , Wang, K. , Hsiao, J. H. , & Chan, A. B. (2025). Grad-eclip: Gradient-based visual and textual explanations for clip. arXiv preprint arXiv:2502.18816.
    7. Touvron, H. , Lavril, T. , Izacard, G. , Martinet, X. , Lachaux, M. A. , Lacroix, T. , . . . & Lample, G. (2023). Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

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

    N/A

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

    N/A

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

    (5) Accept — should be accepted, independent of rebuttal

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

    The approach introduced in this paper presents a practical and meaningful approach for improving the interpretability of medical vision-language models without requiring fine-tuning. This is an important contribution as pretrained VLMs are increasingly used in the medical domain, yet their predictions often remain difficult to interpret in a clinically meaningful way. The proposed method addresses this limitation by extracting sparse local features that are relevant to the prediction, thereby providing localised explanations.

  • 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.

    While the various methods used in the study are not novel, the idea of making VLM models interpretable is appealing. Authors have indicated that they will resolve concerns raised.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a framework for improving the interpretability of zero-shot medical vision-language models by introducing sparse local representations. Specifically, the author derives local text-aligned features from MedSigLIP, discovers clinically meaningful sparse latents using a TopK sparse autoencoder, and uses these latents to build modified classification probes. The paper shows that this approach can improve both zero-shot classification and localization while preserving the original model’s zero-shot capability.

  • 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.

    I think the main strength of this paper is that it addresses an important problem in a meaningful way. Medical vision-language models can perform well in zero-shot settings, but they are often difficult to interpret, and this paper tries to improve that by introducing sparse local representations.

    I also found the overall pipeline well designed. The paper combines local text-aligned features, sparse autoencoding, and sparse probe construction in a way that feels consistent and well motivated. What I found particularly interesting is that the sparse features are not only used for interpretation, but also for improving the actual zero-shot probes.

  • 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.

    While the paper puts the components together in a meaningful way, the main ingredients are individually not very new. The paper itself builds on existing directions including BoF-style local modeling, token-level alignment similar in spirit to PACL, and sparse autoencoders for interpretable feature discovery. Therefore, the contribution seems strongest as a careful integration of existing ideas into the medical zero-shot VLM setting, rather than as a fundamentally new technical formulation.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

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

    N/A

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

    N/A

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

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

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

    My overall recommendation is “weak Accept”. I think the method is reasonable and the reported results show modest but meaningful gains. At the same time, my enthusiasm is somewhat limited because the technical contribution seems to come more from combining existing ideas in a careful way than from introducing a clearly new methodological component.

  • 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.

    The authors have sufficiently addressed my concern.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a framework for improving interpretability in medical vision-language models (VLMs) while preserving zero-shot capabilities. The method extracts local image features using two approaches (bag-of-features distillation and token alignment), and applies a sparse autoencoder (SAE) to obtain sparse latent representations. These latents are then used to construct modified classification probes by selecting the most relevant components based on similarity to text embeddings.

    The key idea is to decompose dense global representations into sparse, local, and potentially interpretable components, and to leverage these components both for analysis and for constructing more focused classification signals. The paper demonstrates improvements in zero-shot classification (AUROC) and localization (pointing game recall) on chest X-ray datasets.

  • 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.Complete pipeline The work presents a clear and structured pipeline that integrates multiple components: extraction of local image features, sparse representation learning via SAE, and probe construction for classification.

    2.Non-invasive approach The method does not require fine-tuning the underlying VLM (MedSigLIP), which is a practical advantage. This allows the approach to be applied post hoc without retraining large models, maintaining zero-shot behavior.

    3.Empirical improvements The method shows consistent improvements over the zero-shot baseline: improved AUROC on the VinDR dataset, and improved localization performance (pointing game recall) across multiple attribution methods. These results suggest that restricting the representation to a subset of sparse features can improve signal quality.

    4.Connection to emerging interpretability methods The use of sparse autoencoders connects this work to recent advances in mechanistic interpretability, particularly the idea of decomposing representations into sparse, disentangled features.

  • 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 methodological novelty The proposed approach is primarily a combination of existing techniques: bag-of-features models (classical computer vision), sparse autoencoders (recent interpretability work), linear probing / feature selection. The paper does not introduce fundamentally new algorithms or formulations. The contribution is largely at the level of integration rather than innovation, which limits its novelty.

    2.Interpretability claims are not sufficiently validated The paper claims interpretability via sparse latents, but does not clearly demonstrate what these latents actually represent. For example, it does not identify a latent as corresponding to a specific clinical pattern (e.g., “pleural effusion region” or “cardiac enlargement boundary”). Instead, it only shows activation patches and similarity rankings, which do not establish clear meaning. In original SAE-based interpretability work (like in NLP), it is common to assign semantic meaning to specific activations (e.g., a feature corresponding to punctuation, or a specific entity type). In this work, similar results are expected, such as identifying which latent corresponds to which medical pattern, but this is not provided. More importantly, there is no validation of these interpretations. A standard approach would be: hypothesize that a latent corresponds to a specific concept, then test on unseen images whether that latent consistently activates when the concept is present. However, this type of validation is missing. As a result, the paper does not convincingly show that the learned sparse features correspond to meaningful or reliable clinical concepts.

    3.Lack of comparison with supervised baselines The evaluation is limited to comparisons within the zero-shot setting, without including standard supervised baselines (e.g., CNNs or fine-tuned models). While the goal is not necessarily to outperform supervised methods, one would expect at least comparable performance. Without such comparisons, it is difficult to assess how competitive the proposed approach is in realistic scenarios where labeled data is available, which is common in medical imaging settings such as chest X-rays.

    4.Open-vocabulary and zero-shot claims are not substantiated While the paper is motivated by zero-shot and open-vocabulary capabilities, which is a big advantage compared to supervised methods trained on closed-vocabulary, all experiments are conducted on fixed label sets. There is no evaluation on: unseen disease categories or true open-vocabulary scenarios. Therefore, the claimed advantages of zero-shot learning are not directly demonstrated.

    5.Method resembles feature selection without deeper analysis The improvement in classification performance (Table 3) appears to stem from selecting a subset of latent features, which is conceptually similar to feature selection. The paper does not provide deeper analysis of why this improves performance beyond intuitive arguments, nor does it compare to simpler baselines for feature selection.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

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

    N/A

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

    N/A

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

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    The main factors leading to my recommendation are the limited methodological novelty and insufficient validation of the core claims.

    While the paper presents a coherent pipeline combining local feature extraction, sparse autoencoders, and probe construction, the approach largely relies on existing techniques (e.g., BoF, SAE, linear probing) without introducing fundamentally new ideas. The contribution is therefore primarily at the level of integration.

    More importantly, the interpretability claims are not convincingly supported. The paper does not clearly demonstrate what individual sparse latents represent, nor does it validate whether these features correspond to stable and generalizable clinical concepts. In particular, there is no attempt to verify whether a hypothesized latent consistently activates on unseen data, which is a standard practice in prior SAE-based interpretability work.

    The experimental evaluation is also limited. Results are only compared within the zero-shot setting, without including supervised baselines, making it difficult to assess practical competitiveness. Additionally, although the work is motivated by zero-shot and open-vocabulary capabilities, all experiments are conducted on fixed label sets, and no true open-vocabulary evaluation is provided.

    Overall, while the paper is well-motivated and technically sound, the lack of novelty and insufficiently validated interpretability claims make the contribution relatively weak.

  • 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.

    While the rebuttal clarifies the intended framing and acknowledges several limitations, it does not sufficiently address the core concern that the SAE latents are not rigorously validated as meaningful or stable clinical concepts. Also, the evidence provided remains largely qualitative and anatomical rather than demonstrating true semantic interpretability in the sense commonly expected in SAE-based mechanistic interpretability work. As a result, the paper remains primarily an integration of existing techniques with insufficient validation of its interpretability claims.



Author Feedback

We thank the reviewers for their constructive feedback. We are encouraged by the consensus that the paper addresses an important problem and presents a coherent, technically sound pipeline with improvements in both classification and localization. We address the main concerns below and will revise the manuscript accordingly.

W1: Limited Methodological Novelty (R2, R3) While individual components are known, their combination requires design choices that are themselves contributions. Specifically: (1) preserving the shared image-text embedding space of a contrastive medical VLM locally, without retraining, enabling zero-shot localization out of the box; (2) using a TopK archetypal SAE on local patch embeddings in the medical domain, most SAE work in vision operates on CLS tokens; (3) selecting sparse latents by text alignment and projecting back into the teacher embedding space, which is a new formulation. We will clarify this framing in the final version.

W2: Interpretability Claims Insufficiently Validated (R3) R3 argues that the paper does not demonstrate what individual sparse latents represent. We respectfully point to existing evidence in the paper. While qualitative, figure 3 shows that high-similarity latents fire in diagnosis-specific regions like heart for cardiomegaly and lateral lung fields for pleural effusion across test data. Laterality-specific latents (separate for left vs. right effusion) emerge without any positional supervision, constituting unsupervised semantic validation. Table 2 quantitatively demonstrates that high- and low-similarity latents are nearly mutually exclusive (lift approx 0), showing that the sparse latents represent distinct, non-overlapping concepts. We acknowledge that qualitative assignment of individual latents goes beyond what is shown, and we list this as a limitation. However, we believe the anatomical consistency and mutual exclusivity of the latents constitute meaningful evidence that they correspond to stable, clinically grounded concepts.

W3: Missing Comparison Between the Two Variants (R1) Tables 3 and 4 report both variants side by side. The key trade-offs are: BoF probes achieve higher AUROC on VinDR (0.88 vs. 0.81); MedSigLIP token probes generalize better in localization on NIH (0.45 vs. 0.38); BoF latents occasionally fire in anatomically implausible positions due to translation invariance, whereas token latents retain positional context. We will add a dedicated paragraph making these trade-offs explicit.

W4: Zero-Shot Claims Not Substantiated (R3) “Zero-shot” refers to the absence of labeled supervision for probe construction: classification uses only cosine similarity between image and text embeddings, consistent with standard VLM evaluation. We agree “open-vocabulary” in the strict sense (unseen disease categories) is not tested and will revise the framing accordingly, listing this as explicit future work.

W5: Performance Gains Resemble Feature Selection (R3) The key distinction from standard feature selection is that latents are scored by semantic alignment with a text embedding, requiring no labeled data. A data-driven feature selector would need labeled examples per diagnosis and would forfeit the zero-shot property. We will make this distinction explicit in the discussion and note potential future work (like PCA).

W6: Grad-ECLIP; Backbone Choice; Figures; Foundation Model Claim (R1) Grad-ECLIP: We thank R1 for this reference and will discuss it as related work. We note our approach modifies the probe itself, and localization improvements hold across multiple attribution methods, suggesting the effect is not attribution-specific. Backbone: MedSigLIP was chosen as a representative, publicly available contrastive medical VLM; we will add a brief justification and note other backbones as future work. Figures: We will add in-figure labels where appropriate. Foundation model claim: We will restate related work to include both single and multi-modal foundation models.



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.

    Please provide a rebuttal for the following weakness:

    Limited methodological novelty — the framework integrates existing components (BoF, PACL-style token alignment, SAE, linear probing) without a fundamentally new formulation (R2, R3). Interpretability claims insufficiently validated — no semantic assignment of individual latents to clinical concepts, and no test of whether a hypothesized latent consistently activates on unseen images with that concept (R3, standard SAE practice). Missing comparison between the two local-embedding variants in Section 2.2 — performance and interpretability trade-offs not reported (R1). Zero-shot / open-vocabulary claims not substantiated — all experiments use fixed label sets; no unseen-category or true open-vocabulary evaluation (R3). Performance gains resemble feature selection — no comparison to simpler feature-selection baselines, and no deeper analysis of why sparse-subset probing helps (R3). Missing comparison with non-dense post-hoc methods (e.g., Grad-ECLIP); SigLIP backbone choice not justified; figure readability and an inaccurate claim about foundation models being multi-modal (R1).

  • 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.

    I agree with R3 that the clinical interpretation ability is not fully evaluated, which is the main weakness in this paper.

    Also, the combination of different papers has limited novelty.

    Despite the importance of research on model interpretation, I do not agree that this paper meets the MICCAI standard.



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 have mainly pointed out the insufficient novelty in this work, while one reviewer also emphasized the lack of proper evaluation of the interpretability. While it’s not ideal that the work hasn’t conducted sufficient evaluation of the interpretability, this seems to be a common limitation in most of such studies. Nevertheless, the proposed method could be of wide interest to the community. The revised version also needs to remove unnecessary claims such as open-vocabulary and explain the zero-shot setting clearly.



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 method is more an integration of existing sparse-representation and probing techniques than a fundamentally new formulation, and the interpretability claims should be moderated. However, the paper addresses an important problem in medical VLM explainability and provides meaningful improvements in zero-shot classification and localization.



back to top