List of Papers Browse by Subject Areas Author List
Abstract
While medical Vision-Language models (VLMs) achieve strong performance on tasks such as tumor or organ segmentation and diagnosis prediction, their opaque latent representations limit clinical trust and the ability to explain predictions. Interpretability of these multimodal representations are therefore essential for the trustworthy clinical deployment of pre trained medical VLMs. However, current interpretability methods, such as gradient or attention-based visualizations, are often limited to specific tasks such as classification. Moreover, they do not provide concept-level explanations derived from shared pretrained representations that can be reused across downstream tasks. We introduce MedConcept, a framework that uncovers latent medical concepts in a fully unsupervised manner and grounds them in clinically verifiable textual semantics. MedConcept identifies sparse neuron-level concept activations from pretrained VLM representations and translates them into pseudo-report-style summaries, enabling physician-level inspection of internal model reasoning. To address the lack of quantitative evaluation in concept-based interpretability, we introduce a quantitative semantic verification protocol that leverages an independent pretrained medical LLM as a frozen external evaluator to assess concept alignment with radiology reports. We define three concept scores, Aligned, Unaligned, and Uncertain, to quantify semantic support, contradiction, or ambiguity relative to radiology reports and use them exclusively for post hoc evaluation. These scores provide a quantitative baseline for assessing interpretability in medical VLMs. Code is available at https: //github.com/RakibulHaqueSajal/MedConcept.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5647_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/RakibulHaqueSajal/MedConcept
Link to the Dataset(s)
AbdomenAtlas 3.0: https://huggingface.co/datasets/AbdomenAtlas/AbdomenAtlas3.0Mini
Merlin: https://stanfordaimi.azurewebsites.net/datasets/60b9c7ff-877b-48ce-96c3-0194c8205c40
BibTex
@InProceedings{HaqMd_MedConcept_MICCAI2026,
author = { Haque, Md Rakibul AND Sultan, K. M. Arefeen AND Kataria, Tushar AND Elhabian, Shireen Y.},
title = { { MedConcept: Unsupervised Concept Discovery for Interpretability in Medical VLMs } },
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 presents a fully unsupervised framework that extracts latent medical concepts from pretrained VLMs and grounds them in clinically verifiable text. Sparse neuron-level activations are translated into pseudo-report summaries, enabling interpretable inspection of model reasoning. A quantitative semantic verification protocol employs a frozen medical LLM as an external evaluator to assess concept–report alignment. Three scores—Aligned, Unaligned, and Uncertain—quantify semantic support, contradiction, or ambiguity for post hoc evaluation.
- 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.Fully Unsupervised Concept Discovery. The framework achieves fully unsupervised extraction of clinically salient concepts from pretrained VLM representations via sparse decomposition, identifying interpretable neuron-level latent factors without reliance on manual annotation. 2.Textual Grounding and Interpretability. Discovered latent concepts are explicitly grounded in medical semantics by mapping sparse activations to aligned clinical terminology and generating patient-specific concept summaries, thereby enabling direct inspection of internal model reasoning. 3.Rigorous LLM-Based Semantic Evaluation. The authors introduce a systematic entailment protocol that leverages a frozen medical LLM to assess per-concept alignment with radiology reports, yielding three quantitative metrics—Aligned, Unaligned, and Uncertain—that together capture semantic support, contradiction, and ambiguity for robust post hoc validation.
- 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.Figure 1 contains notational and naming inconsistencies. The inequality “k » p” in the diagram appears mismatched with “k » m” stated in the main text, and the model name “Med-Gamma” should be corrected to “MedGemma”. 2.In the caption of Fig. 1, “textual embeddings generated by a text decoder” should be “. . . text encoder”? 3.In the subsection of “Quantitative Evaluation of Concept-Report Alignment”, the authors state that images lacking concept units above the activation threshold are excluded from evaluation. How should this exclusion be addressed in practical deployment scenarios where model interpretability is still required despite sparse activations?
- 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?
The authors present an unsupervised approach for discovering medical concepts and grounding them in clinically verifiable textual semantics. The core idea is novel and well-motivated. The introduction of three quantitative concept scores enables systematic evaluation, and the experimental results are supportive of the proposed method to a certain extent.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes MedConcept, a unified unsupervised framework for discovering clinically meaningful concepts from pretrained 3D medical vision–language model (VLM) representations using sparse autoencoders (SAEs). It grounds latent neuron activations in medical semantics via cosine similarity matching with a curated vocabulary derived from UMLS and expanded by LLMs.
- 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.This paper represents one of the first attempts to apply concept discovery to 3D medical imaging foundation models, which is a novel step beyond existing work largely focused on 2D images.
- 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 use of sparse autoencoders for concept discovery is not novel, as similar approaches have been extensively explored in both language and vision models. While the paper extends this idea to 3D medical VLMs, the methodological contribution appears largely incremental, focusing more on application and evaluation rather than introducing fundamentally new techniques. E.g., Selective Concept Bottleneck Models Without Predefined Concepts TMLR 2025; Concepts from Neurons: Building Interpretable Medical Image Diagnostic Models by Dissecting Opaque Neural Networks IPMI 2025.2.The paper does not demonstrate how the discovered concepts improve clinical performance, decision-making, or model reliability in practical applications. The connection between the discovered concepts and downstream tasks is weak. 3.The interpretability is derived from SAE latent units rather than the original VLM representations. As such, the method explains a reconstructed proxy of the embedding space rather than the actual internal mechanisms of the Merlin model. It remains unclear whether the discovered concepts faithfully reflect the model’s true representations or are artifacts of the SAE decomposition. 4.The notion of interpretability is primarily based on post-hoc labeling via cosine similarity, without demonstrating any causal relationship between the identified concepts and model behavior, which limits the strength of the interpretability claims. 5.How the dense anatomical annotations generated via R-Super are used?
- 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.
(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?
See the weakness part.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal clarifies the paper’s positioning as an auditing framework rather than a novel SAE method, but my main concern remains unresolved: the work does not demonstrate that the discovered concepts improve clinical reliability, model trustworthiness, or downstream decision-making in any meaningful way. The evaluation is limited to post-hoc semantic alignment analysis, without showing that the identified concepts help detect failures, reduce hallucinations, or provide actionable insights for clinical deployment. Therefore, I remain unconvinced about the practical significance of the proposed framework and maintain my recommendation for rejection.
Review #3
- Please describe the contribution of the paper
This paper proposes MedConcept, an unsupervised framework that uses sparse decomposition for medical concept discovery from pretrained VLM representations. Moreover, authors introduce an LLM-based scoring framework to evaluate SAE-driven concept discovery in medical VLMs.
- 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 overall writing is clear and well-organized.
- The proposed SAE-driven concept discovery method is unsupervised, therefore has relatively high potential for real-world applications.
- The proposed LLM-based scoring framework is easy to conduct and understandable.
- 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.
- As discussed in limitations, the proposed LLM-based scoring framework depends on paired radiology reports for evaluation. Moreover, semantic naming depends on the performance of pre-trained models.
- Lack of experimental comparison with other SAE-based concept discovery methods.
- Results in Fig.3 show that MedConcept tends to perform overconfidently in AbdomenAtlas 3.0, resulting in unaligned concepts.
- 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
Questions:
- Only with the sparsity regularization, how to ensure the w_j lies in the manifold of the joint VLM latent space?
- How \lambda_1 in eq.(2) influence the discovered concepts?
- 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?
Considering the insufficient comparison experiments, the reliance on paired reports for evaluation and other questions regarding methods, this paper is marginally below the acceptance threshold.
- 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.
The rebuttal addressed concerns of reviewers.
Author Feedback
We thank the reviewers for their constructive feedback. In the camera-ready, we will incorporate their suggestions and release code, concept dictionaries, and evaluation protocols (preprocessing, SAE training, and MedLLM prompts) to ensure reproducibility. Positioning and Novelty:MR,R1,R2,R3.MedConcept is not proposed as a new SAE method, but as a semantic auditing framework for 3D medical VLMs. Its core contribution is enabling quantitative, report-grounded evaluation of discovered concepts, which differs from prior SAE-based work that primarily focuses on representation discovery and qualitative analysis in language or 2D vision (including recent concept bottleneck and neuron dissection approaches). Specifically, it: (1) extracts unsupervised concepts using SAEs guided by open-source medical vocabularies, (2) performs report-grounded verification with MedLLMs, introducing Aligned/Unaligned/Uncertain metrics for semantic support, contradiction, and ambiguity, and (3) enables systematic comparison across models, datasets, sparsity levels, and decoding settings. To our knowledge, such quantitative auditing of latent concepts in 3D medical VLMs has not been addressed in any prior work. Scope and Downstream Utility:MR,R2,R3.MedConcept is designed for model auditing and interpretability, not for directly improving downstream predictive accuracy. It extracts clinically relevant concepts from frozen pretrained VLMs and provides report-grounded evidence for expert review, enabling assessment of which concepts are encoded, missing, or unsupported. This supports model reliability analysis and comparison prior to deployment. As medical VLMs move toward clinical use, the ability to systematically audit and verify learned concepts is critical for safe and trustworthy deployment. Evaluating direct impact on clinical decision-making is important future work but beyond this study’s scope. Evaluation Design and Deployment:R1.Samples with no activated SAE units are excluded only from normalized score computation, as the metrics are undefined for empty concept sets. In practice, such cases would be reported as low-confidence / no-salient-concept outcomes, optionally with sub-threshold activations. Importantly, deployment would not rely solely on Aligned concepts; Uncertain and Unaligned concepts are also informative, indicating ambiguity, missing report content, or potentially unsupported model behavior. Interpretability:R2,R3.We clarify that MedConcept provides representation-level interpretability, not mechanistic causality. The SAE reconstructs frozen VLM embeddings, so its latent units represent sparse directions constrained by the original embedding geometry, rather than arbitrary proxies. Concept naming via cosine similarity yields candidate labels, which are validated through report-grounded MedLLM evaluation. Non-perfect alignment across datasets shows current VLMs only partially capture clinical concepts, leaving room for improvement. Report-Grounded Evaluation Effects:R3.Unaligned scores reflect report-grounded disagreement, which may arise from either model error or incomplete reporting. In datasets such as AbdomenAtlas, where reports are sparse, visually present but unreported findings may be labeled Unaligned due to missing text. We will clarify that Unaligned denotes report-unsupported, not necessarily clinically incorrect. Sparsity Ablation:R3.The sparsity coefficient λ₁ controls the trade-off between concept selectivity and coverage. Strong sparsity (λ₁=0.2) suppresses nearly all concepts (Present/Absent/Uncertain ≈ 0). Moderate sparsity (λ₁=0.002) achieves the best balance (Present ≈ 0.29, Absent ≈ 0.46, Uncertain ≈ 0). Weak sparsity (λ₁=0.0002) increases coverage but introduces noise (Present ≈ 0.05, Absent ≈ 0.14, higher variance). This trend is consistent across temperatures (T=0.1, 0.3, 0.5), indicating λ₁ has a stronger effect than T. Thus, λ₁=0.002 provides the best trade-off between concept discovery and reliabil
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 prioritize clarifying the methodological novelty with respect to prior SAE-based concept discovery work and strengthening comparisons to existing approaches. Additionally, address concerns regarding the validity and faithfulness of the learned concepts (e.g., whether they reflect true model representations or artifacts of the SAE) and clarify their impact on downstream utility, including clinical relevance. Finally, provide clarification on evaluation design choices (e.g., exclusion of low-activation samples, reliance on paired reports) and address inconsistencies or presentation issues noted by the reviewers.
- 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’s main strength lies in its fully unsupervised discovery and textual grounding of clinically salient latent concepts in pretrained 3D medical VLMs, together with LLM-based semantic evaluation protocol for quantitative auditing. The primary remaining concern is the extent to which the method demonstrates direct clinical usability. However, the work has clear benefits and represents one of the earliest efforts to quantitatively audit latent concepts in 3D medical VLMs, an important problem that has not been sufficiently addressed in prior work. As the authors note, systematic auditing and verification of learned concepts will be critical as medical VLMs move toward clinical deployment. While evaluating direct impact on clinical decision-making is an important direction for future work, I agree with authors that it is reasonably beyond the scope of this study. Overall, the paper is interesting, timely, and methodologically valuable, and is recommended for acceptance.
Meta-review #2
- 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.
The intent of the paper as clarified in the rebuttal is to make a case that the pre-trained model (Merlin) for CT extracts semantically meaningful features. The way they ‘prove’ is to build a sparse autoencoder which projects the data into a higher dimensional space than the original Merlin embedding and forcing the model to use only a few active neurons, so that the network is forced to learn disentangled, distinct features rather than just compressing the data. Then the features are mapped to textual embeddings of selected concepts and best matches chosen through cosine similarity. The reviewers have raised valid concerns on the downstream utility of the methods and rightly questioned the validity of the method itself. The method explanation has some inconsistencies with the figure Figure 1 mentions the lower dimension as p whereas the text uses m making it confusing. Similarly, the matrix Wd matrix is supposed to be kxm whereas the cosine similarity computation section on page 4 mentions Wj of dimension d. So these errors represent a sloppiness that is unacceptable. One critical experiment to do to prove that it is indeed recognizing these medical concepts and not general language concepts, is to replace the medical concept vocabulary embeddings with textual embeddings derived from other concepts, and see if the cosine similarity will not pick some best matching neurons. If MedGemma cross-check is ruling them out, then a better approach may be to pass the CT report through MedGemma first, select those concepts for the cosine similarity test for matching to the sparse neurons. As you scale the concept vocabulary, the precision and recall of such approach will tank as is well known for the VLM model embeddings.
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.
All the reviewers find the proposed idea of discovering concepts from VLM representations using unsupervised methods novel. Moreover, two of the reviewers recommended acceptance, highlighting the rigorous experiments and potential real-world applications. Uncovering semantic concepts from the dense feature representations of VLMs by employing Sparse Autoencoders can still be considered novel and interesting to a wider medical AI community