Abstract

Faithfulness and interpretability are essential for deploying deep neural networks (DNNs) in safety-critical domains such as medical image analysis. B-cos networks modify the parameterization of convolutional and classification layers to measure class evidence via feature-weight alignment, enabling built-in, class-specific contribution maps without post-hoc explanations. While maintaining diagnostic performance competitive with state-of-the-art DNNs, standard B-cos networks exhibit severe aliasing artifacts in their explanation maps, rendering them unsuitable for clinical use, where clarity is essential. In this work, we address this limitation by introducing anti-aliasing strategies using ASAP and BlurPool (BP) to significantly improve explanation quality. Our experiments on chest X-ray datasets demonstrate that the modified B-cos_ASAP and B-cos_BP preserve strong predictive performance while providing faithful and artifact-free explanations suitable for clinical application in multi-class and multi-label settings. Code is available at: https://github.com/shrebox/Artifact-free-B-cos-Networks

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3915_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/shrebox/Artifact-free-B-cos-Networks

Link to the Dataset(s)

RSNA Pneumonia Detection Challenge dataset: https://www.kaggle.com/competitions/rsna-pneumonia-detection-challenge/overview VinBigData Chest X-ray Abnormalities Detection dataset: https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection

BibTex

@InProceedings{AryShr_Faithful_MICCAI2026,
        author = { Arya, Shreyash AND Agnihotri, Shashank AND Kleinmann, Marcel AND Schiele, Bernt AND Keuper, Margret},
        title = { { Faithful, Interpretable Chest X-ray Diagnosis with Artifact-Free B-cos Networks } },
        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

    The primary technical contribution of the paper lies in mitigating grid artifacts in B-cos networks through anti-aliasing techniques such as ASAP and BlurPool, while simultaneously improving the interpretability of chest X-ray classification, a widely used and clinically relevant radiological task.

  • 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 strength of the paper lies in its clear identification of the need for inherently interpretable architectures, rather than relying on post hoc attention-based methods. It introduces a principled mathematical approach to mitigate grid artifacts, which is important for practical deployment. The authors also provide relevant empirical evidence demonstrating that B-cos networks offer improved interpretability over post hoc methods such as Grad-CAM, while preserving classification performance.

  • 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 primary weaknesses lie in the experimental design and presentation. While the authors select appropriate baselines, include B-cos variants, and conduct ablation studies, the overall presentation of results could be improved. The current figures suggest that the gains are relatively marginal; presenting results in a clearer tabular format may better highlight these differences. Additionally, although improvements over baselines are documented, cross-modal performance comparisons are difficult to interpret. For instance, it is unclear which model would be most suitable for deployment, as a clear “best overall” model is not explicitly identified.

    The choice of faithfulness metric is also somewhat limited. Including IoU alongside EPG would provide a more comprehensive evaluation. Finally, while extending the approach from multi-class to multi-label classification is necessary for the domain, it does not constitute a significant methodological contribution.

  • Please rate the clarity and organization of this paper

    Good

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

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

    N/A

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

    N/A

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

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

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

    The work is novel and timely for the field and the contribution is also interesting to the community. With better presentation of the experiments, the work may be considered for the conference.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I continue with my score.



Review #2

  • Please describe the contribution of the paper

    A solution combining ASAP and BlurPool is proposed to address the issue of interpretation image artifacts caused by the inherent interpretable model B-Cos network in medical images. The core contribution lies in improving the visual quality of the explanatory image while ensuring that the core diagnostic performance does not decrease.

  • 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.

    Using a relatively simple method, the author achieved the desired effect of removing artifacts.

  • 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.

    Although it achieved the effect of removing artifacts, it reduced the quantitative indicators of the model.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

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

    N/A

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

    Although the effect of removing artifacts has been achieved, there has been a significant decrease in the quantitative evaluation indicators. Especially compared to B-cos, the indicator has significantly decreased, which is not an ideal effect.

  • 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 proposed method did not achieve a balance between interpretability and accuracy.

  • 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 article was recommended and accepted, and the author gave a clear response.



Review #3

  • Please describe the contribution of the paper

    Overall, the manuscript is well-structured and addresses the critical issues of faithfulness and interpretability in medical image analysis. The study introduces B-CosASAP and B-CosBP models, aiming to enhance the reliability of medical AI systems.

  • 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 technical contributions are noteworthy: B-CosASAP effectively provides artifact-free explanations and improved localization without compromising predictive performance. Furthermore, B-CosBP addresses the common issue of aliasing-prone downsampling by incorporating spatial blurring, resulting in superior explanation quality and localization accuracy. The authors have validated the efficiency and robustness of their proposed approaches through empirical experiments using two widely recognized public datasets, RSNA and VinBigData (VinBigXray).

  • 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.Insufficient Presentation of Empirical Results While the authors claim that the proposed approach outperforms existing attribution methods, it is difficult to be fully convinced of its efficiency based on the current manuscript. In the experiments section, various metrics (Accuracy, AUROC, Recall, EPG micro Precision/Recall) are presented in Figure 3.However, relying solely on a performance comparison chart is insufficient for a rigorous technical evaluation. Notably, several lines in the plots overlap, making it impossible to distinguish the specific performance of each method. To improve clarity and facilitate a detailed comparison, the authors should provide a comprehensive table including the exact numerical values and statistical summaries rather than relying on cluttered figures.

    2.Lack of Statistical Rigor The manuscript currently lacks the necessary statistical analysis to support its claims. Although I acknowledge the page constraints of a conference paper, reporting minimal statistical significance—such as p-values, Cohen’s d, or E-values—is essential to demonstrate that the observed improvements are statistically meaningful. Without these measures, the “efficiency” of the proposed framework remains speculative.

    3.Inadequate Localization Validation In Figure 1, the authors present only a single IoU result as a qualitative example. Providing a single case is not acceptable as representative evidence for the performance across the entire dataset. A more extensive quantitative evaluation of IoU (or similar localization metrics) across the full test set is required to validate the model’s interpretability and localization capabilities.

  • 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?

    The lack of statistical rigor, insufficient quantitative data representation, and limited validation across the entire dataset fail to provide a convincing demonstration of the proposed method’s efficiency.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I think that the author’s reply is reasonable to address my concerns.



Author Feedback

We thank the AC & reviewers for the feedback. The overlapping radar curves in Fig 3 were intentional: for a fixed architecture, metrics differ little, while the main gain is artifact-free explanation. We agree tables are clearer so exact values are below. Values are %, mean ± std over 3 seeds 5 folds. ConvNeXt-T on VinBigXray & some metrics are omitted due to space but follow the same trend. RN50=ResNet-50, CNXt-B=ConvNeXt-B, EPG_P=EPG Precision. RSNA Pneumonia | Arch | Method | Recall | AUROC | EPG_P | mIoU | |—|—|—:|—:|—:|—:| | RN50 | Baseline | 50.6±1.7 | 83.3±0.6 | NA | NA | | RN50 | B-cos | 74.8±2.2 | 86.7±0.7 | 24.1±0.8 | 4.7±0.5 | | RN50 | B-cos_ASAP | 77.5±1.1 | 87.1±0.5 | 21.1±1.3 | 9.4±1.8 | | RN50 | B-cos_BP | 79.2±1.7 | 87.1±0.7 | 20.1±1.9 | 10.7±2.0 | | CNXt-B | Baseline | 44.8±2.3 | 81.7±0.6 | NA | NA | | CNXt-B | B-cos | 79.9±1.6 | 88.2±0.5 | 18.6±1.1 | 1.4±0.5 | | CNXt-B | B-cos_ASAP | 80.2±2.3 | 88.0±0.4 | 21.9±2.8 | 2.0±0.3 | | CNXt-B | B-cos_BP | 80.7±1.4 | 88.1±0.4 | 20.1±1.6 | 2.1±0.6 | VinBigXray | Arch | Method | Recall | AUROC | EPG_P | mIoU | |—|—|—:|—:|—:|—:| | RN50 | Baseline | 38.6±1.8 | 80.8±1.1 | NA | NA | | RN50 | B-cos | 38.6±1.5 | 91.9±0.6 | 9.4±0.2 | 2.3±0.2 | | RN50 | B-cos_ASAP | 37.0±1.1 | 87.3±0.8 | 9.9±0.4 | 5.2±0.9 | | RN50 | B-cos_BP | 37.3±1.6 | 87.7±0.9 | 9.0±0.2 | 5.4±0.7 | | CNXt-B | Baseline | 39.0±0.9 | 89.7±0.6 | NA | NA | | CNXt-B | B-cos | 44.8±1.4 | 92.2±0.5 | 13.2±0.4 | 1.6±0.1 | | CNXt-B | B-cos_ASAP | 31.5±1.7 | 91.7±0.5 | 13.4±1.1 | 2.8±0.4 | | CNXt-B | B-cos_BP | 40.7±1.6 | 90.3±1.1 | 16.2±0.7 | 4.6±0.4 | [AC, R1, R2, R3] Trade-off & best model choice. The goal is obtaining faithful B-cos explanations that are aliasing-free while preserving classifier performance, rather than improving classifier performance. This is what we observe. For RSNA CNXt-B, vs baseline, B-cos_ASAP/BP have only about 2% lower accuracy but improve recall from 45% to ~80% & AUROC from 82% to 88%, while adding inherent explanations. For VinBigXray CNXt-B, B-cos_BP is at par or better than baseline in accuracy, recall, precision, & AUROC; B-cos_ASAP has a recall trade-off but keeps high AUROC & improves mIoU. A stronger encoder also improves metrics, as CNXt-B vs RN50 shows. Our aim is not one universally best model but showing that anti-aliased downsampling can be combined with B-cos backbones for artifact-free explanations. ASAP & BP are both effective; there is no clear numerical winner across settings & we want to be transparent about that. Visually ASAP appears slightly cleaner, so B-cos_ASAP is preferred when explanation quality is primary.

[AC, R1, R3] Localization beyond examples. EPG_P & mIoU above are computed over the full validation set where EPG_P is localization precision: the fraction of positive attribution energy inside boxes. This is an established measure for attributions since bounding boxes usually enclose the relevant areas but not every region inside the box is relevant. We additionally report mIoU as requested, & it supports the same conclusion: B-cos_ASAP/BP improve mIoU over B-cos, e.g., RN50 RSNA 5% to ~10% & RN50 VinBigXray 2% to 5%. Baseline EPG_P/mIoU are NA because baseline models do not produce inherent B-cos contribution maps; mixing GradCAM/LayerCAM in the same columns would conflate post-hoc & built-in faithful explanations.

[R3] Statistical rigor. All metrics are mean ± std over 3 seeds with small stds showing stable trends. We also computed paired t-tests over matched seed/fold runs. For mIoU, B-cos_ASAP/BP significantly improves over B-cos in all tested dataset-backbone comparisons with max p=7.0e-3.On RSNA, B-cos_ASAP/BP are not significantly worse than B-cos in Acc/Recall/Precision/AUROC for both methods. Thus localization improves significantly while predictive performance remains competitive.

In the final version we will use the extra half page to discuss this trade-off, replace plots with full tables & add visualizations.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper addresses an important topic: improving the faithfulness and visual quality of explanations in inherently interpretable B-cos networks for chest X-ray diagnosis. Reviewers appreciated the motivation, the use of anti-aliasing strategies to reduce grid artifacts, and the potential relevance for interpretable medical AI.

    However, the reviews raise important concerns. The main issues are the unclear trade-off between artifact reduction and diagnostic performance, insufficient quantitative reporting, lack of statistical analysis, and limited localization validation. Reviewers also noted that the current figures make it difficult to assess the true magnitude of the gains. I recommend inviting the paper for rebuttal. The authors should address: 1- Clarify the performance–interpretability trade-off. 2- Strengthen localization validation beyond single examples. 3- Better justify which model is preferable for deployment.

  • 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 the rebuttal, the authors substantially clarified the main concerns raised in the initial reviews. In particular, they provided full quantitative tables, additional localization metrics beyond isolated examples, and statistical analysis over repeated runs. These clarifications make the performance-interpretability trade-off much clearer. The rebuttal shows that the proposed variants improve localization quality, including mIoU, while maintaining competitive predictive performance, even if no single variant dominates across every metric and dataset. The explanation of when B-cos_ASAP is preferable versus when B-cos_BP is preferable is also reasonable and improves the practical interpretation of the results. The paper still has some limitations. The gains are not uniformly large across all settings, and the final paper should present results in tables rather than crowded plots, with a clearer discussion of deployment preferences and trade-offs. However, the main concerns about insufficient quantitative reporting, limited localization validation, and lack of statistical support were addressed convincingly in the rebuttal, and all reviewers moved to 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.

    Accept

  • Please justify your recommendation.

    The paper is one of the few works on visualization anti-aliasing offering better explainability on anomaly detection in chest X-ray imaging. All reviewers appear convinced with the rebuttal and have recommended acceptance.



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 three reviewers now propose to accept the paper, as they indicate that all concerns have been properly addressed. Clear-cut case.



back to top