Abstract

The integration of deep learning into medical vision applications has led to a growing demand for interpretable predictions. Typically, classification and explainability are treated as separate processes, with explainability methods applied post hoc to pre-trained classifiers. However, this decoupling introduces additional computational costs and may lead to explanations misaligned with the underlying model. In this paper, we propose One For All (OFA), an efficient, single-stage approach that jointly optimizes classification accuracy and self-explanation during training. OFA achieves this through a multi-objective framework, eliminating the need for separate explainability models while ensuring faithful and robust explanations. Extensive experiments on medical datasets confirm that OFA delivers competitive classification performance while providing high-quality, inherently interpretable explanations, making it a scalable and versatile solution for fully explainable classification.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4655_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)

https://medmnist.com/

BibTex

@InProceedings{NaoMeh_One_MICCAI2025,
        author = { Naouar, Mehdi and Vogt, Yannick and Boedecker, Joschka and Kalweit, Gabriel and Kalweit, Maria},
        title = { { One For All: A Unified Approach to Classification and Self-Explanation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15973},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper aims to learn an attribution map for network explanation simultaneously with classification. For that, the authors rely on Salvage [1] which learns a network to approximate Shapley values by optimizing the Jenson-Shannon divergence between the estimated distribution of Shapley values and the distribution of the prediction of the classifier for which the explanation is desired. Once learned, this explainer can be used to solve the original classification task by summing the estimated Shapley values and applying an activation function. Here, the authors propose to directly learn the attribution maps to obtain the desired classification (as in Salvage with of the sum the estimated Shapley values and an activation function). In addition, they propose two other losses to ensure robustness to masking. The first one constrains the network to predict the right classification even when the image is masked and the second one to have similar predictions for the original image and the masked one.

    They evaluate their method on 8 public datasets from MedMNIST. They first compare their model to state-of-the-art classification models on the classification task. The results show that they are equivalent or better on almost all datasets. Then, they compare themselves to an attention method and a gradient method on the explainabilty using (relatively) symmetric relevance metric. They outperform both state-of-the-art methods.

    [1] Naouar, M., Raum, H., Rahnfeld, J., Vogt, Y., Boedecker, J., Kalweit, G., Kalweit, M.: Salvage: Shapley-distribution approximation learning via attribution guided exploration for explainable image classification. In: The Thirteenth International Conference on Learning Representations (2025)

  • 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.
    • Integrating the attribution map estimation to the training is more efficient than Salvage, which trains a separate explainer after the classifier training, if we already know that we want an explanation.
    • The method outperforms the state-of-the-art methods compared.
  • 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 novelty is limited as Salvage already proposes to use the learned explanations for the classification task.
    • In addition, the contributions of the two other proposed losses are not evaluated since there is no ablation study.
    • The comparison with state-of-the-art methods for the explainability is very limited with only two very basic methods. At least, a comparison to Salvage and to Shapley value (estimated by the proposed method) is necessary to correctly quantify the contribution of the proposed method.
    • The loss proposed in the section 4.3 seems very redundant with the other two losses: if the predictions for both original and masked images are optimized to be the ground truth, the distributions should be similar. Again, an ablation study is necessary to quantify the significance of this loss.
    • The number of patches used to partition the images should be indicated.
    • Is Segm-B the baseline for classification? It is not clear.
    • Please add references for the state-of-the-art classifiers.
    • In Table 1, for DermaMNIST ACC, the last line (OFA-B) should be in bold.
  • 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.

  • 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/2025/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 — could be rejected, dependent on rebuttal

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

    The novelty is limited, in particular because no ablation study is performed to evaluate the contribution of all the losses. The comparison to state-of-the-art is weak for explainability. Therefore, it is hard to evaluate the proposed contributions.

  • 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 evaluation (without abalation study) of the method is not sufficient to quantify the impact of adding the losses l1 and l2. In the reply, the authors do not develop this point sufficiently. In addition, they have not answered the question about the redundancy of the losses. The comparison with the state of the art on the explainability part is not sufficient (at least Shapley values, Salvage and the Seg-m baseline should be compared) and the authors have once again not replied on this subject.



Review #2

  • Please describe the contribution of the paper

    This paper proposes One For All (OFA), a novel, single-stage framework for medical image classification with built-in explainability. Instead of treating explainability as a post-hoc process, OFA jointly optimizes both classification and self-explanation during training. The method integrates Shapley-value inspired attributions within a multi-objective loss framework, achieving competitive classification accuracy with faithful, robust, and inherently interpretable explanations. Classification accuracy of OFA is evaluated across nine medical imaging datasets from Medical MNIST+ and explanations compared against gradient and attention-based explainability methods.

  • 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.
    • OFA’s end-to-end training of classification and explainability eliminates the need for separate explainer models, improving computational efficiency and alignment between predictions and explanations.
    • Strong experimental evaluation were conducted on diverse medical datasets which shows effectiveness of OFA.
    • Quantitative metrics (AUC, accuracy, SRG, R-SRG) demonstrate OFA’s superior performance both in classification and attribution quality.
  • 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 OFA introduces multiple auxiliary loss terms, the individual impact of each term (classification, robustness, explanation) is not dissected via ablation study. It would strengthen the claims to show how much each term contributes to performance and explanation quality.
    • OFA having same architecture with Segm-B which already shows superior performance in the table. Since the training objective can be applied into other architectures, comparitive results in other architectures like CNNs (ResNet-18,50 or DenseNet) and ViTs would be helpful.
    • The paper focuses on attribution maps for explainability. Other forms of explanation (counterfactuals, concept-based) are unexplored, though potentially complementary in clinical settings.
    • The paper claims improved efficiency over post-hoc methods in introduction section, but quantitative inference-time comparisons are missing. Reporting actual inference speedups would solidify this claim.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

  • 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/2025/en/REVIEWER-GUIDELINES.html
    • Equation label of Shapley distribution u(S) is missing in page 4.
    • There are missing word in page 6 line 4, “compute the positive and shapley values (ϕ+ and ϕ−)” to “ compute the positive and negative shapley values (ϕ+ and ϕ−)”
    • The highlighted(bald) result in Table 1 Derma ACC should change. In table 1, OFA-B shows better performance(91.6) than ViT-B(90.9) but highlight is on the ViT.
  • 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 — could be 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 promising contribution in the domain of interpretable medical AI by proposing OFA, a unified framework that integrates classification and explanation through end-to-end training. The strengths of the paper are notable, including a well-designed framework that eliminates the need for separate explainer models, which potentially improves prediction-explanation alignment. Additionally, the experimental evaluation is thorough, utilizing multiple medical datasets and reporting strong performance across both classification and explanation metrics.

    However, the paper has several limitations. A primary concern is the lack of an ablation study to assess the individual contributions of the multiple loss terms introduced in OFA. Without this analysis, it is difficult to discern how each component drives the improvements. Furthermore, while OFA uses the Segm-B architecture, which already performs well, the authors do not explore the generalizability of their training objective across other widely used architectures such as CNNs or ViTs. This limits the broader applicability of the proposed method. Lastly, the claim of improved computational efficiency over post-hoc methods lacks quantitative support, as no inference-time comparisons are provided. Therefore, I would recommend weak reject but encourage the authors to address the above issues.

  • 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.

    After carefully reading other reviewer’s comments and the authors’ rebuttal, I acknowledge the authors’ efforts to clarify several technical points. However, I believe several key concerns remain insufficiently addressed in this paper.

    1. While the paper proposes a unified framework combining classification and explanation, this paradigm has been explored in previous works such as Salvage. Although OFA simplifies the training process into a single stage, the methodological novelty is still limited given the lack of in-depth analysis.
    2. The comparisons remain limited. While the authors emphasize computational efficiency and joint training, the lack of comparison with stronger post-hoc attribution methods. Therefore, I maintain a reject recommendation.



Review #3

  • Please describe the contribution of the paper

    This paper focuses on XAI. The authors put forth a proposal to modify the loss and training process of a visual transformer to explicitly introduce the calculation of Shapley values of each input patch. These values are then utilized for three distinct sub-tasks: classification, assessing the robustness to masking of the input image (i.e., the ability to recognize a masked image), and self-explanation (i.e., ensuring that the explanation is equivalent between masked and original images). The efficacy of the model is evaluated using several datasets from MedMNIST, yielding comparable results in terms of accuracy but noteworthy advancements in explicability.

  • 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 topic is of significant importance to the community. The paper is well-written, offering a clear explanation of the usage of Shapley values. The concept appears to be straightforwardly implementable. The methodology is appraised through its application to an assortment of datasets.

  • 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. Some points need clarification:
      • 1.a. The equation 5 in section 4.2 should be better explained. $S$ is the masked image, but does it mean that $i \in S$ enumerates the non-masked patches? If yes, please explicitly state it. In the robustification to masking, I understand that the $l_2$ norm is using the set of visible patches from the image for the positive values, but the usage of negative values is not clear. Perhaps the link with section 3.2 should be emphasized (definition of $u(S)$).
      • 1.b. In the overall pipeline, how are the masked images generated? This could be important for reproducibility. It seems that only one masked image for each real image per batch is generated. Is this sufficient with regards to the discussion in section 3.2 about the estimation of $u(S)$?
    2. Evaluation should be developed:
      • 2.a. Only 2 approaches are used for comparison. The attention-based one is dedicated to Transformer models, but convolutional-based approaches using attention mechanisms have also shown good results on XAI. Please specify which “Grad” method is used, as several exist.
      • 2.b. Only additive metrics are used, but there exist several other metrics for saliency map evaluation showing different aspects of the explanation. Here is a suggestion of reference to complementary metrics : Gomez, T. et al. “Metrics for Saliency Map Evaluation of Deep Learning Explanation Methods.” ICPRAI 2022. https://doi.org/10.1007/978-3-031-09037-0_8
  • 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.

  • 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/2025/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 proposal is very interesting and addresses a hot topic for the community, even if the evaluation is limited.

  • 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.

    Authors have well answered my comments, either on clarifications and evaluations. I agree with other reviewers that a deeper analysis of the contributions of each loss would be interesting, but the experiments on Segm-B and OFA-B is already convincing from my point of view.




Author Feedback

We first thank the reviewer for their thoughtful and constructive feedback

Reviewer#1: 1.a. We denote by S the set of visible patches, where i ∈ S indexes the non-masked patches. Following (Naouar et al. 2025), the classifier’s prediction on a masked image is decomposed into a sum of the positive values from the visible patches and the negative values from the masked patches. In Eq 5 (loss l2), we enhance the classifier’s robustness to masking by optimizing this decomposed prediction w.r.t ground truth label 1.b. Please see G3 2.a. In our paper, “grad” refers to saliency maps (Simonyan et al., 2014) 2.b. Thank you for highlighting this paper. It addresses two limitations of IAUC and DAUC: 1) Out-of-distribution masking: This issue is mitigated in our case since our classifier is explicitly trained to be robust to masking 2) Pixel score ranking: While SRG (SRG = IAUC - DAUC) shares this limitation, we additionally report R-SRG, which addresses this limitation by evaluating the consistency of relative attribution scores

Reviewer#3

  • You’re absolutely right that Salvage functions as an explainable classifier. However, it involves three distinct training stages: training a classifier model, a surrogate, and then the explainer—leading to high computational cost. Our approach addresses this by combining classification and explanation in a single unified training process
  • Please see G1
  • We divide each image into 196 patches

Reviewer#4

  • Please see G1 and G2
  • Our method is compatible with any segmentation architecture, as it requires computing shapley values for each image patch
  • We fully agree that inference-time comparisons are valuable. While we were unable to include them due to space limitations, we would like to emphasize that our model outputs predictions and attributions jointly in a single forward pass, unlike post-hoc methods which require additional computations

Meta-Reviewer#2: Thank you for raising this important distinction. Our method is fundamentally different from SEEN, ConceptSHAP, and RISE: 1) SEEN and ConceptSHAP rely on concept learning and aggregation into predictions or concept importance estimation, OFA does not learn or use concept-based explanations 2) Our approach learns attribution scores directly during training via masked inputs, rather than via post-hoc concept relevance or perturbation strategies 3) Unlike RISE, which requires over 2,000 test-time inferences per image (to sample mask combinations), our model produces both predictions and attribution maps in one forward pass of the unmasked image—greatly improving efficiency

  • Please see G2 and G3

G1: Our ablation study comparing two settings, l1 (Segm-B) and l1+l2+l3 (OFA-B), showed that l2 and l3 don’t affect the classification performance of our model. However, removing either l2 or l3 leads to a collapse in explanation performance. This is expected: l2 ensures robustness so irrelevant features do not influence predictions, and l3 is the only loss optimizing faithfulness of attributions on the unmasked image

G2: Attribution, counterfactual, and concept-based explanations each address distinct, complementary questions: Which parts of the input contributed most to the model’s decision? (attribution), How important is a high-level concept (e.g. “stripes”, “fur”) for a particular class? (concept-based), and What minimal change to the input would flip the model’s prediction? (counterfactual). These approaches operate at different representation levels—pixel space, abstract concept space, or through input perturbations—and are thus inherently evaluated using different metrics and protocols. As a result, a direct comparison between them is non-trivial, and there is currently no unified evaluation framework across these paradigms

G3: We use mask embeddings (from DINOv2) as a masking strategy. For l2, we sample a single masked image. For l3 (self-explanation), we use 32 masks per image from the informed sampling distribution (Naouar et al. 2025)




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    While methodologically sound, the work conceptually overlaps with earlier integrated explainability frameworks such as SENN, ConceptSHAP, and RISE. The paper lacks a clear differentiation from these existing approaches. Additionally, the masking strategy is not sufficiently clarified, and further elaboration on the mechanics of this process is needed. The paper would also benefit from an evaluation of human interpretability or an experiment focused on concept-level aggregation

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Reviewers raised concerns primarily centered around novelty (c.f. Salvage) and missing ablation studies but agree on the general relevance of the work. Overall the authors reasonably mitigated some of those concerns in the rebuttal, such that this paper, while of limited novelty, is overall suitable to discuss at MICCAI.



Meta-review #3

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    This paper presents OFA, a unified framework for medical image classification and self-explanation using Shapley-inspired attributions integrated into the training objective. While the paper tackles a timely and important challenge in explainable AI and is generally well-written, multiple reviewers raised valid concerns about insufficient methodological analysis. In particular, the lack of ablation studies prevents clear attribution of performance gains to the proposed loss components. Additionally, the evaluation is limited—comparisons to state-of-the-art XAI methods (e.g., Salvage, Shapley baselines) are missing, and claims of computational efficiency remain unquantified. Despite promising results and reviewer interest in the concept, the paper lacks the experimental depth and novelty assessment required for MICCAI acceptance.



back to top