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

Abstract

Despite providing concept-based explanations for decision-making in medical AI, concept bottleneck models (CBMs) suffer from prediction distortions under insufficient concept annotations. Inspired by real-world scenarios where experts continually learn diagnostic signs, we introduce the framework of concept incremental learning (ConceptIL). Such learning framework acquires knowledge of new concepts from sequential data with limited annotation of concepts, and further promotes interpretability of AI decision-making. However, achieving this ConceptIL can be challenging due to the limited access of historical data and imbalanced distribution of concepts. To address these, we propose an analytical concept learning scheme, which optimizes parameters through a recursive formulation. This ConceptIL method achieves concept prediction accuracy theoretically proven to be equivalent to that of joint training. Additionally, we propose a dynamic weighting strategy to mitigate the concept imbalance problem. Results on a skin disease dataset show that our method matches the diagnostic performance of models trained on the full concept set, demonstrating the effectiveness of our incremental concept learning. Moreover, our work obtains the best performance compared to competing incremental learning algorithms, establishing a new state-of-the-art for ConceptIL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1480_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhoHan_ConceptIL_MICCAI2026,
        author = { Zhou, Hangqi AND Wang, Bomin AND Zhuang, Xiahai},
        title = { { ConceptIL: Concept Incremental Learning for Interpretable Medical AI } },
        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 Concept Incremental Learning (ConceptIL) for medical AI, where a model gradually acquires new expert-defined concepts from sequentially arriving datasets with partial concept annotations, and then uses the accumulated concepts for downstream diagnosis. Methodologically, the paper proposes an Analytical Concept Predictor (ACP) with a recursive least-squares style update, aiming to learn concept predictors incrementally without storing past samples, while claiming equivalence to joint training under the stated formulation. The paper further adds a dynamic weighting strategy to mitigate concept imbalance, and evaluates the framework on Fitzpatrick17k/SkinCon-derived concepts for benign-vs-malignant skin lesion classification.

  • 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 relevant and practically motivated problem setting: in real clinical annotation workflows, concept vocabularies do evolve over time and concept labels are often incomplete, so incremental concept acquisition is a meaningful extension of concept-based medical AI. The authors compare against several continual learning baselines, report both concept-level and task-level metrics, and include a simple qualitative example showing how newly acquired concepts can improve downstream prediction.

  • 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 methodological novelty appears limited. The analytical recursive update is very close in spirit to prior analytic continual learning work such as ACIL and GACL, which already study exemplar-free recursive analytic updates with privacy/memorization claims; here the main change is to apply that machinery to concept prediction rather than standard class-incremental classification. The paper does not sufficiently clarify what is fundamentally new beyond this adaptation. Relevant prior work includes ACIL [21] and GACL [19]. 2.The conceptual novelty relative to prior concept-based literature is also not fully convincing. The paper positions ConceptIL as a new framework, but the space already includes CBMs [6], post-hoc CBMs [15], label-free CBMs [7], and even incremental residual concept bottleneck models [12]. The submission does not sharply distinguish whether the key novelty is a new problem definition, a new optimization method, or simply a new combination of existing ideas from CBMs and continual learning.

  • 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?

    The paper studies an interesting and clinically relevant problem, and the privacy-aware analytic update is technically clean. However, for a top conference submission, the core methodological contribution is not sufficiently novel relative to ACIL/GACL-style analytic continual learning and existing concept-bottleneck variants, and the experimental evidence is too limited to establish broad significance.

  • 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 well answered my comment.



Review #2

  • Please describe the contribution of the paper
    • This paper introduces ConceptIL, a framework for incrementally learning clinical concepts within concept bottleneck models (CBMs)
    • Addresses the practical scarcity of concept annotations and the limitations in data sharing due to privacy issues
    • The core contribution is an Analytical Concept Predictor (ACP) with a recursive least-squares update proven equivalent to joint training, alongside a dynamic weighting strategy for concept imbalance.
    • Experiments on Fitzpatrick17k show competitive performance against models trained on the full concept set.
  • 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.
    • Novel and Well-motivated problem formulation
    • Although incremental concept learning is not new, its formulation and adaptation for medical AI seems to be novel.
    • The motivation (gradual accumulation of clinical knowledge, costly concept annotation, problematic data sharing) is realistic and well-argued.
    • Theoretical Guarantees
    • The paper shows that the ACP’s recursive update is equivalent to joint training over all accumulated data.
    • This is shown both theoretically and empirically
    • Privacy-Preserving by Design
    • The inverse correlation matrix is stored instead of raw patient data
    • Handling of Concept Imbalance
    • It is proposed the dynamic weighting strategy and a weighted least-square formulation that connects naturally to the ACP framework
  • 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.
    • Evaluation on a Single Dataset
    • All experiments use a single dataset (Fitzpatrick17k / SkinCon) with one binary task (Benign vs. Malignant).
    • This is insufficient to claim state-of-the-art or demonstrate generalisability
    • Concept Prediction Accuracy Trade-off is Not Addressed
    • In the T = 5 setting, the method achieves lower C-ACC than the Finetune baseline.
    • Given that interpretability is a central claim, this trade-off between concept accuracy and downstream performance deserves explicit discussion.
    • Missing Concept-Specific Baseline
    • All IL competitors are general-purpose methods. Res-CBM [12], discussed in the introduction as closely related work on concept discovery within CBMs, is absent from the comparison.
    • Scalability & Privacy
    • ACP stores P_j matrix per concept scaling as O(Kd_h^2), which may be limiting.
    • No memory or runtime analysis is provided
    • Privacy is included in the motivation, but never discussed in the design choice or evaluation. It is not clear whether the proposed design is actually safe.
  • 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

    Comments

    • The BFC metric is introduced without precedent or citation. While intuitive, its relationship to established incremental learning evaluation protocols should be better justified or grounded in the existing literature.
    • The ablation study covers only the dynamic weighting strategy. Ablations on \lambda, the number of concepts per task, and backbone freezing would strengthen the empirical analysis.
    • The concept of task overlap is described as a realistic clinical property, but its effect on performance is never studied. Even a brief analysis of how the overlap rate affects C-F1 or T-F1 would strengthen the paper.
    • The T = 15 setting shows more volatile per-task performance (Fig. 2, right) than T = 5.This instability is not discussed. Understanding when and why performance fluctuates would be valuable.
    • Figure 3 presents only two images as a case study. While illustrative, this is anecdotal and does not constitute systematic evidence for the interpretability claims. A brief quantitative analysis of concept-level decision changes across the test set would be considerably more convincing.
    • The claim from the introduction “This reflects real-world scenarios where experts gradually recognize evolving diagnostic signs based on the stepwise acquisition of patient data, updates of clinical guidelines, and newly introduced devices.” would deserve a citation.

    Typos

    • Abstract: “our work obtains best the performance” -> “our work obtains the best performance”
    • Sec 2.1: “with enhance interpretability” -> “with enhanced interpretability”
    • Sec 2.3: “To address the this challenge” -> “To address this challenge”
    • Sec 3.2: “As results in Table 1 shows” -> “As the results in Table 1 show” Other: Introduction: “pleasant performance” is informal; consider “strong performance” or “competitive performance”
  • 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?

    Interpretable models and the idea of incremental concept learning are genuinely important and interesting in medical AI, and the community would benefit from knowing about it. The experimental limitations are real concerns, but do not undermine the core contribution.

  • 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.

    The rebuttal clarifies some technical details, but fails to convincingly address the core novelty concern raised by R1.The authors frame their main contribution as problem formulation rather than methodology, but the experimental evidence (a single dataset with a single task) is too narrow. I am inclined to the view of R1.The paper is quite good, but not enough for this top venue.



Review #3

  • Please describe the contribution of the paper

    The paper derived the fundamental algorithm for concept incremental learning to address the problem of progressively learning new concepts and improving interpretability in medical artificial intelligence. It outperforms the other existing algorithms.

  • 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 claimed that it is the first attempt to address the problem in learning new concepts from sequential data with limited annotation and handling the concept imbalance through a dynamic weighting strategy. The paper also presented an analytical concept predictor and achieves incremental learning without sample storage through recursive updates.

  • 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 pictorial illustration is not enough to demonstrate the implementation of mathematical equations, as the pure mathematical explanation is somewhat obscure and difficult to understand.

  • 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 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 fundamental algorithm for concept incremental learning provided by the paper would contribute to the development of the field.

  • 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



Author Feedback

We thank the meta reviewer and all the reviewers for providing valuable comments and suggestions. Our response is as follows.

Q1: Regarding contribution and novelty of the paper (R1).

  • We clarify that our main contribution is the introduction and formulation of the new problem, i.e., concept incremental learning (ConceptIL) for medical AI. Obtaining sufficient expert-labeled clinical concepts is prohibitively costly. ConceptIL enables models to acquire new concepts from sequentially arriving datasets with limited concept annotations. To the best of our knowledge, this is the first attempt to investigate such a problem for medical AI.

  • Analytic continual learning: We agree that the analytical recursive update is inspired by ACIL/GACL through recursive analytical optimization, but our work addresses a different learning problem. ACIL/GACL assume fully observed task labels for classification, whereas ConceptIL studies an expanding multi-label concept space with partial annotations. In this case, missing concept labels should not be treated as negatives, which leads to the concept-specific LS formulation in Theorem 1.Since medical concepts are often sparse and imbalanced, we further introduce dynamic weighting to improve learning of rare clinical concepts.

  • Concept learning: We would like to clarify that ConceptIL targets a realistic scenario for medical AI, i.e., learning the continually growing concept set, which is not addressed by existing CBM variants. Conventional CBMs [6], post-hoc CBMs [15], and label-free CBMs [7] mainly study how to build interpretable models under a fixed concept set. Res-CBM [12] discovers new residual concepts from a candidate concept bank within a single task, but the number of residual concepts is pre-defined and fixed.

Q2: Concept prediction accuracy trade-off is not addressed (R2).

  • Thank you. We would like to emphasize that ConceptIL w/o weighting achieved comparable C‑ACC to the baseline (T=5: 92.61 vs 92.62; T=15: 89.38 vs 89.36) while obtaining better C-F1, T-ACC and T-F1, showing that both concept accuracy and downstream performance get improved. When the dynamic weighting strategy is applied, C-ACC of ConceptIL decreased, but C‑F1 further increased. It aligns with the objective of such strategy, i.e., detecting imbalanced concepts.

Q3: Regarding missing concept-specific baseline (R2).

  • Res-CBM [12] aims to discover new concepts from a pre-defined concept candidate bank within a single task. Therefore, it is not applicable for the experimental setting, where models incrementally learn new concepts across sequentially arriving tasks.

Q4: Regarding scalability & privacy (R2).

  • ACP only needs 1 epoch for training. During inference, it’s runtime remains the same as the backbone.
  • As mentioned at the end of Section 2.2, ACP stores the inverse correlation matrix instead of original samples from previous tasks, thereby preserving data privacy. Given d_h = 1024, ACP needs around 2MB memory for each concept.

Q5: Regarding the details on the experimental settings (R2).

  • BFC: Note that BFC = \bar{A} - \bar{A}^(0), where \bar{A}^(0) represents the average incremental performance of the same model trained with fixed concept set. Therefore, BFC is also related to the established protocol in [9].
  • Concept overlap: High overlapping rate results in a small concept set, which may improve concept accuracy and restrict the downstream performance.
  • Volatile per-task performance: We randomly divide the whole training and validation sets into T subsets. Therefore, the training subset is small when T=15, which may cause the instability.

Q6: Regarding suggestions on writing, e.g., fixing typos, add citations and polishing figures (R2, R3).

  • (R2) Thank you. We will correct the typos, add citations and improve Figure 3 in revision.
  • (R3) Thank you. We will make Figure 1 clearer in revision.



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.

    Reviewer comments are mixed. The authors need to address the concerns raised by the reviewers, especially regarding reviewer #1’s concern on distinguishing the contribution of the proposed method from the existing ones, and Reviewer #2’s suggestions about more discussion and how the proposed method could help the real-world practice.

  • 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.

    Reviewer #1 was convinced by the rebuttal that the authors have addressed all the concerns, and changed from rejection to acceptance. Reviewer #3 has not made changes. Reviewer #2 was inclined to the Reviewer #1’s initial view regarding limited novelty and changed to rejection. However, in light of the paper’s claims and the rebuttal’s arguments, the AC was convinced that the proposed framework is a novel contribution and that the authors have conducted an interesting study exploring this problem, which resonates with real-world challenges.



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.

    Started with diverse comments and concerns and diverging ratings, the rebuttal is not effective in addressing the major concerns, including but not limited to concerns over the technical contributions. Overall it requires significant efforts that goes beyond minor nature to get into MICCAI in this round.



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.

    In my opinion, the authors have addressed most of the concerns raised by reviewers. I recommend Accept.



back to top