List of Papers Browse by Subject Areas Author List
Abstract
Deep learning models in clinical workflows increasingly pair chest X-ray classifiers with Vision-Language Models (VLMs) to generate natural-language justifications, but no prior work explores whether Kolmogorov-Arnold Networks (KANs), whose spline functions yield interpretable units, translate architectural interpretability to trustworthy text explainability. We introduce KANEx, a framework that leverages the symbolic transparency of Kolmogorov-Arnold Networks (KANs) to ground VLM reasoning. Unlike traditional post-hoc explanations (e.g., Grad-CAM) which often approximate model behavior, our approach derives Interpretability-Grounded Heatmaps and Context Aided Explanations bridging the gap between the mathematical interpretability and medical explainability. Benchmarked on the MIMIC-CXR dataset, we demonstrate that KAN-based architectures achieve AUC parity with SOTA ResNet/ViT baselines while providing significantly more faithful saliency maps. By feeding these grounded contexts into downstream VLMs, we show a measurable boost in semantic alignment and clinical reasoning accuracy compared to baselines. Our results suggest that grounding linguistic explanations in mathematically interpretable units is a necessary step toward verifiable and trustworthy medical AI.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4330_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/cerai-iitm/KANEx
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ShaKri_KANEx_MICCAI2026,
author = { Shailya, Krithi AND Ravi, Ananya Lakshmi AND K. V., Venkatanathan AND S. Sundaram, Sowmya AND S. Krishnan, Gokul AND Anand, Aditi AND Ravindran, Balaraman},
title = { { KANEx: Translating Kolmogorov-Arnold Networks’ Interpretability to Medical Explainability } },
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 KANEx, a novel explainability framework for chest X-ray analysis which combines Kolmogorov–Arnold Networks (KANs) with Vision-Language Models (VLMs) to produce explanations about model predictions. They also introduce KAN-Map, a novel heatmap generation method leveraging KANs for image-based explainability. The main contributions are:
- A unified pipeline combining KAN models and VLMs for multi-label diagnosis and explanation generation.
- A new heatmap method, KAN-Map, which derives spatial importance directly from KAN spline activations instead of gradients, avoiding approximation errors.
- A systematic evaluation of KAN variants integrated into ResNet and ViT backbones.
- 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 goes beyond standard saliency methods by leveraging KANs instead of post-hoc explanations. This is a meaningful conceptual shift compared to standard Grad-CAM-based pipelines.
- KAN-Map is a novel and interesting contribution which directly evaluates spline functions over spatial features, eliminating the need for gradients. This approach is both computationally efficient and theoretically grounded.
- Strong motivation in the field of medical trustworthy AI
- Solid evaluation setup and empirical analysis
- Analysis of KAN variants is extremely valuable for future research opportunities
- 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 KAN-Map is novel, the broader pipeline is conceptually similar to prior works combining saliency maps with report generation (e. g. , Grad-CAM + VLM pipelines). The novelty lies mainly in replacing components rather than introducing something fundamentally new.
- While the paper discusses recent radiology VLM approaches like RadAlign, it does not compare the performance of these approaches to KANEx. Comparisons only compare heatmap quality to GRADCAM/Attn-R, and text explanations under various inputs. No full end-to-end VLM system is compared (e. g. RadAlign which directly generates textual explanations. Further evaluation is needed to determine if KANEx is better than existing SOTA methods which produce grounded explanations.
- The use of computational approaches for evaluating model performance and explanation quality may not actually be reflective of actual clinical need. Having a subset of explanations evaluated by human radiologists would be beneficial to support the claims.
- Some parts (e. g. , KAN-Map formulation) are mathematically dense and could benefit from clearer language or visual aid.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a promising and well-motivated approach to improving explainability in medical imaging by combining KANs with VLM-based reporting. The proposed KAN-Map is a technically sound and novel contribution that shows improvements in explanation quality over gradient-based methods. However, there is limited comparison to recent methods, a lack of clinical validation, and the overall pipeline itself is not a novel contribution.
- 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
Review #2
- Please describe the contribution of the paper
A nice way to get explainability using KANs in 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.
Novel way to get grad-cam like heatmaps using KAN. Increase in faithfullness in certain settings.
- 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.Confusing writing. A. Table 2.Vit normal has the best energy but it is not in bold? Why only bold based on that category? B. What are we even supposed to see in table 2? C. Lext-p is only mentioned once and never used again?
2.Why compare with Grad-CAM and not Score-CAM? That’s gradient free too.
3.I also feel the title is misleading. There is no translation of interpretability happening. Its the same as CAM in terms of visual output or feedback.
4.Eval size is not mentioned anywhere.
[Minor] LeXT and LExt is used interchangebly, should solve for this.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 writing is just too confusing to make an informed decision. Also the metrics do not exactly shout an improvement. Agreed, there is a good methodological progress, but the current evaluation harness and explanations do not really motivate that.
- Reviewer confidence
Not confident (1)
- [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 #3
- Please describe the contribution of the paper
The authors propose a KANEx, a new method for (1) creating attribution maps highlighting key regions from the model’s decision-making process (KAN-Map) and (2) generating textual explanations in a natural language. The authors propose to replace MLPs with KANs on top of visual backbones. KAN-Map does not require gradients, as the needed values are aggregated during inference.
- 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 method for generating heatmaps and textual explanations that outperforms other methods and has many good properties as there is no need for gradients which is computationally expensive in other methods
-
originality thanks to the use of KAN in the field of explainable AI
-
many experiments and nicely presented results
-
well written paper
-
- 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.
-
it is not clear why the particular criteria when filtering MIMIC-CXR to create custom dataset
-
in various references, it is claimed that the work is a preprint whereas it was published (Ref 1, 6, 15, 16, 27)
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
In my opinion the use of KANs in this domain is interesting and novel. The results with comparisons of the proposed method to other solutions in literature are convincing. KAN-Map have also significant advantage that unlike other methods, it does not require the computation of gradients.
- 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.
I agree with reviewer #1 that on the high-level the proposed pipeline is conceptually analogous to the ones used already in literature. However, I find the incorporation of KAN into the pipeline within the explainable AI field as novel approach that potentially can open up new directions.
I accept the explanation of the authors that due to the limited number of pages, they coudn’t provide more comparisons with other methods from the literature. I think that the provided results are satisfying enough to decide that it would be valuable for the community to learn about the proposed method.
Author Feedback
We thank all reviewers for their constructive and detailed feedback. We address major concerns below:
Comparison with Textual Explanation systems [R1,MR] KANEx is proposed as an explainability module designed to explain classification outputs through faithful, visually grounded heatmaps via KAN-Map. Textual explanations are a downstream effect of providing better heatmap-informed visual context to a VLM. Systems like RadAlign optimize report generation quality without evaluating attribution faithfulness or explanation alignment. KANEx serves as a complementary grounding framework for such downstream VLMs; we identify this integration as a valuable evaluation for future work.
Clinical Relevance [R1, MR] We acknowledge and identify clinician evaluation as future work (Section 3.1), warranting a separate study across KAN variants. Our evaluation setup incorporates clinically grounded signals by using radiologist-annotated bounding boxes for heatmap evaluation, and real MIMIC-CXR reports used by LExT as a proxy for clinical correctness.
Novelty [R1] Current heatmap methods rely on approximations-linear importance for Grad-CAM, propagated attention weights for Attention Rollout. KAN-Map derives spatial importance directly from learned intrinsic spline activations in a forward-only pass, This is a fundamentally different formulation (Section 2.2) resulting in different interpretability claims-transparency through inspectable functions with improved faithfulness, localization and is deterministic without post hoc estimates.
Misleading Title [R2] We acknowledge that the visual output of KAN-Map resembles CAM-style heatmaps; however, the title emphasizes how KANs provide mathematical interpretability through inspectable polynomial spline functions at the model level, which KANEx leverages for user-facing explainability via heatmaps and downstream textual explanations. As KANs gain traction in medical imaging models, this translation becomes a key distinction from approximation-based methods.
Clarity and Presentation [R2] Table 2 consolidates three simultaneous comparisons: KAN-Map vs. Grad-CAM/Attn-R within the same backbone, KAN-enhanced vs. vanilla baselines, and inter-KAN variant analysis. Bolding is applied per-method-group highlighting the best relative gains among KAN variants. This goes beyond showing that KAN-Map outperforms existing methods, and characterizes which KAN variant benefits most from this attribution approach. We will revise the caption and the table with clearer bolding convention and reading directions.
Baseline Selection [R2] Grad-CAM and Attention Rollout are mechanistically comparable to KAN-Map: all three derive spatial importance from internal model representations, via gradient-weighted feature maps, propagated attention weights, and spline activations respectively. Grad-CAM is also a practical baseline as a standard in medical imaging explainability. Score-CAM derives importance by input sensitivity, which is conceptually different. As per rebuttal guidelines we cannot present additional results, but we emphasize that our baseline selection deliberately favours theoretical comparability and depth over breadth, given page constraints.
Dataset Filtering [R3] The filtering criteria was designed to enable end-to-end evaluation comparison, requiring cases with both radiologist-annotated bounding boxes and ground-truth explanations covering clinical reasoning.
Minor Edits: LExT-C, a subset of LExT-P, evaluates explanation correctness which is most relevant to our task, we will correct this and standardise the notation throughout. The full dataset comprises 20,000 cases with a 30% test split; explicit split sizes will be added in Section 3.Preprint references will be updated to published venues. KAN-Map formulation will include an intuitive walkthrough.
We hope this rebuttal clarifies the scope and novelty of KANEx as a meaningful step toward trustworthy, interpretability-grounded medical AI.
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 presents a promising explainability framework (KANEx) with a novel KAN-Map. While technically sound, concerns remain regarding limited comparison with end-to-end VLM systems, which integrate diagnosis and explanation in a unified manner. The rebuttal should therefore strengthen benchmarking against such methods and provide clearer evidence of clinical relevance, ideally through expert evaluation or more clinically grounded metrics.
- 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.
Despite concerns regarding clarity and evaluation scope, the paper is recommended for acceptance due to its promising methodological contribution and strong relevance to trustworthy medical AI.
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 manuscript presents an interesting explainability module designed to explain classification outputs. The rebuttal clarifies several concerns raised by the reviewers.
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.
While the paper would benefit from stronger clinical validation and comparison with full end-to-end VLM explanation systems, the proposed KAN-Map is a technically interesting and reasonably novel contribution for gradient-free visual explainability.