List of Papers Browse by Subject Areas Author List
Abstract
Automated polyp segmentation is critical for early colorectal cancer detection, yet performance often degrades across clinical centers due to illumination variability from heterogeneous endoscopic systems and imaging protocols. We propose KAN-AINet, a ConvNeXt-U-Net architecture that integrates Kolmogorov-Arnold Networks (KAN) for adaptive feature modulation through learnable nonlinear activation functions, improving robustness to diverse lighting conditions. KAN-AINet includes two modules: (1) a KAN-based Illumination Modulation Module (KAN-IMM) that adaptively scales encoder features according to brightness patterns, and (2) a KAN-based Boundary Attention Module (KAN-BAM) that refines decoder features for improved structural consistency. Trained on Kvasir-SEG, SUN-SEG, and PolypGen (C1-C5), and evaluated on Kvasir-Sessile, CVC-ColonDB, ETIS-LaribPolypDB, and PolypGen-C6. KAN-AINet outperforms ESPNet and nnU-Net on the combined unseen datasets, improving mDice by 2.20% and 42.27%, and mIoU by 3.05% and 55.51%, respectively. On ETIS-LaribPolypDB and PolypGen-C6, mDice increases by 3.24% and 7.26% over ESPNet, with substantially lower HD95 and ASD, demonstrating strong cross-dataset generalization. Ablation studies show that KAN modules improve mDice by up to 5.81% and mIoU by 6.45% while reducing HD95 and ASD. Stratified analysis on 822 external images shows consistent mDice gains under dark (+4.98%), medium (+3.47%), and bright (+5.06%) conditions. Brown-Forsythe testing further confirms lower prediction variance (F = 11.28, p < 0.001; σ₁²/σ₂² = 0.68), supporting robust and stable performance across imaging variation. The source code and implementation are publicly available at https://github.com/biodatlab/kanainet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5427_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/biodatlab/kanainet
https://huggingface.co/biodatlab/kanainet
Link to the Dataset(s)
N/A
BibTex
@InProceedings{TimWat_KANAINet_MICCAI2026,
author = { Timklaypachara, Watcharapong AND Chiewhawan, Monrada AND Pudhota, Abhiraj AND Aguilera, Stephanie Marie AND Bodduluri, Sandeep AND Achakulvisut, Titipat},
title = { { KAN-AINet: Kolmogorov-Arnold Network with Adaptive Illumination Modulation for Generalizable Polyp Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposed a KAN-AINet for polyp segmentation under diverse lighting conditions. The proposed KAN-AINet is verified on different datasets and performed good cross-dataset generalization.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1)The proposed model is a UNet-like architecture, and the two KAN-based modules are . simple and effective. 2)The proposed method is experimented on different datasets to verify the generalization.
- 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 paper states that KAN-IMM combines local feature descriptors with global illumination statistics (mean, standard deviation, and brightness) to learn per-channel coefficients. However, the manuscript lacks a clear mathematical formulation or architectural detail explaining exactly how these 1D global priors are tensor-fused with the 2D spatial feature maps before entering the KANLinear layer. This ambiguity hinders the reproducibility of the core module. 2)Automated polyp segmentation in clinical colonoscopy strictly requires real-time video stream processing. The proposed KAN-AINet has a massive parameter count of 151.8M and a computational complexity of 71.76 GFLOPs. While the paper thoroughly evaluates illumination robustness, it completely omits any evaluation of inference latency (e. g. , FPS) on standard-resolution endoscopic images. The immense computational overhead casts doubt on the model’s practical viability for real-time clinical intervention. 3)It seems that the paper is not yet complete, with some sections missing.
- 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 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.
(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?
Although the paper provides a detailed description of the experimental part and analyzes the model performance, on the one hand, the ablation experiments are incomplete; on the other hand, the details in the paper are described too simply, such as in sections 2.2, 2.3, and 2.4.Moreover, the structure of the paper is also incomplete.
- 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 reviewing the authors’ rebuttal,I have comprehensively assessed the novelty, experimental rationality, practical application value and overall academic contribution of this manuscript. I finally decide to reject this paper and the detailed reasons are stated as follows: 1,The authors merely restated the basic structural workflow of the proposed KAN-IMM module, but failed to clearly demonstrate its essential innovative advantages compared with existing illumination-aware feature adjustment mechanisms. There is a lack of sufficient verification on the rationality of module design and its targeted adaptability to endoscopic polyp segmentation scenarios. 2,The proposed method shows inferior quantitative results on the authoritative mainstream dataset CVC-ColonDB compared with the state-of-the-art ESPNet. The authors subjectively attributed the unsatisfactory performance to dataset limitations rather than optimizing and analyzing model defects. Performance gains are only reflected on several less commonly used datasets, and the overall performance improvement is marginal, which cannot verify that the proposed method achieves comprehensive superiority over existing polyp segmentation approaches. 3,The photometric augmentation ablation experiment adopts single fixed parameter settings without sufficient multi-group comparative verification. It is inappropriate to draw a hasty conclusion that feature-level illumination modulation is completely superior to input-level data augmentation. The experimental design lacks comprehensiveness, and the deduced conclusions are insufficiently generalizable and scientifically convincing. 4,The selection of comparison baselines is not comprehensive enough, and some classic mainstream polyp segmentation methods are omitted without reasonable grounds. Only vague descriptions are provided for the evaluation of SAM-based models, without complete quantitative comparison results. The insufficient horizontal comparison cannot fully and fairly verify the academic competitiveness of the proposed method. In summary, the authors’ rebuttal fails to effectively resolve the core concerns raised by reviewers.
Review #2
- Please describe the contribution of the paper
This paper proposes KAN-AINet, a ConvNeXt-U-Net architecture that integrates Kolmogorov-Arnold Networks (KAN) to address illumination variability in generalizable polyp segmentation. The framework utilizes a KAN-based Illumination Modulation Module (KAN-IMM) in the encoder to adaptively scale features based on brightness patterns, and a KAN-based Boundary Attention Module (KAN-BAM) in the decoder to refine structural consistency and edge detection.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The study explicitly targets a critical and highly practical challenge in automated colonoscopy, namely the performance degradation of polyp segmentation models caused by heterogeneous endoscope illumination, specular highlights, and shadow transitions. 2.The adaptation of Kolmogorov-Arnold Networks into a 2D spatial formulation (SpatialKAN2D) provides a mathematically sound approach to handling pixel-wise nonlinear photometric corrections while preserving spatial structure, offering a clear structural advantage over standard fixed-activation convolutions. 3.The evaluation protocol is robustly designed across multiple clinical centers, utilizing a diverse set of cohorts to effectively demonstrate cross-dataset generalization capabilities on unseen data such as ETIS-LaribPolypDB and PolypGen-C6.
- 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 experimental methodology could be further validated by evaluating the models with standard illumination augmentations included; deliberately disabling these augmentations to isolate the KAN-IMM effect makes it difficult to ascertain whether the proposed module provides distinct performance advantages over a standard backbone trained with rigorous photometric data augmentation. 2.The baseline evaluations would be more comprehensive if they included comparisons beyond nnU-Net and ESPNet; incorporating recent state-of-the-art domain generalization frameworks or specific illumination-correction networks designed for endoscopic imaging would help to better contextualize the proposed method’s relative advantages. 3.Given the reported computational requirements of 71.76 GFLOPs and 151.8M parameters, the manuscript would greatly benefit from an evaluation of inference latency; reporting real-time processing metrics such as frames-per-second is highly recommended to fully substantiate the method’s practical clinical utility during live colonoscopy.
- 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 proposed KAN-AINet introduces a notable architectural contribution by utilizing SpatialKAN2D for adaptive illumination modulation, demonstrating consistent empirical improvements across multiple unseen clinical datasets. While the decision to omit standard illumination augmentations during training complicates a direct evaluation of the architectural benefits compared to traditional data-driven strategies , the method’s statistically significant gains under extreme lighting conditions and reduced prediction variance provide empirical support for its robustness. Despite the high parameter count (151.8M) and the current absence of real-time inference speed benchmarks , the strong cross-center generalization results and improved structural consistency on datasets such as ETIS-LaribPolypDB represent a sufficient contribution to merit acceptance.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
Acceptable performance is demonstrated in Tables 1 and 2.
- 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.
Positives: 1.The proposed method successfully addresses the problem of illumination variation which is crucial in endoscopy practice. 2.It proposes an innovative use of Kolmogorov-Arnold Networks (KAN) to enable nonlinear adaptation. 3.The proposed work successfully provides experimental proof with cross-dataset evaluation.
- 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.
Technical Issues: 1.KAN novelty is not properly motivated relative to other nonlinearities (GELU, Swish). 2.The model demonstrates marginal performance improvement compared to other well-performing baselines (ESPNet). 3.The model has high complexity (151M parameters, 71.76 GFLOPs) → difficult to implement in real applications. 4.Should include comparisons with transformer architectures used for segmentation (SAM-based, etc). 5.Illumination robustness depends on the distribution of the datasets, not entirely independent from them. 6.Complexity of KAN layers is significant but without efficiency assessment. Inconsistencies with regard to achieved improvements.
- 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?
8/10
- 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 all reviewers for their constructive feedback. We address the key comments raised across reviews below.
KAN-IMM formulation [R1.1]:
KAN-IMM extracts two 1D compact descriptors from the 2D input feature map. The local branch captures spatial patterns, applying Conv-BN-ReLU followed by global average pooling. The illumination branch computes scalar statistics, summarizing the overall lighting condition. These two descriptors are concatenated into a single vector and passed through a KANLinear layer to produce per-channel scaling weights. Each channel of the input is then rescaled with the weights via per-channel multiplication. We will revise the text for clarity upon acceptance. Our source code is also publicly available for full implementation details and reproducibility.
Inference latency [R1.2, R2.3, R3.3]:
KAN-AINet has 151M parameters and inferences at 54.15 FPS for 384x384 images on an RTX 3090 (mean latency 18.47 ms; p95 19.94 ms), exceeding the clinical time threshold (≥25 FPS). While the current state-of-the-art, ESPNet has 297M parameters and inferences at 12.2 FPS for 384x384 images on the same hardware (mean latency 81.95 ms; p95 82.84 ms). This is because KANLinear and SpatialKAN2D are highly parallelizable, making performance more dependent on memory transfer bandwidth than on arithmetic computation. We will include a detailed latency and FPS analysis in the final version.
Marginal gain over ESPNet [R3.2]:
KAN-AINet outperforms ESPNet across all metrics on the two more recent unseen datasets. However, CVC-ColonDB comprises 55% of the unseen set: 345 images drawn from just 15 videos recorded at a single hospital in 2012, limiting its clinical relevance and generalizability. We suspect this skews the combined unseen set results, as reflected in the per-dataset breakdown. Compared to ESPNet, KAN-AINet achieves −1.4% mIoU / +1.7% F^w_β on CVC-ColonDB, +3.5% mIoU / +8.8% F^w_β on ETIS-LaribPolypDB, and +6.9% mIoU / +10.1% F^w_β on PolypGen-C6. Augmentation-based ablation [R2.1]:
We retrained KAN-AINet with synthetic photometric augmentation (BrightnessContrast ±0.3 p=0.7, Gamma [0.7,1.3] p=0.5, HSV ±20/±30/±20 p=0.5). Photometric augmentation consistently results in lower segmentation and boundary metrics across the unseen datasets:
CVC-ColonDB: Dice 0.784→0.752, HD95 78.3→86.7 px, ASD 52.0→57.8 px; ETIS-LaribPolypDB: Dice 0.859→0.857, HD95 48.4→57.1 px, ASD 29.1→33.5 px; PolypGen-C6: Dice 0.872→0.854, HD95 87.6→74.5 px, ASD 26.1→55.3 px; Unseen mean: Dice 0.819→0.799, HD95 70.4→75.8 px, ASD 41.2→49.9 px.
Sequential illumination modulation with boundary refinement from KAN-IMM and KAN-BAM achieves better generalization than photometric augmentation by learning to modulate features adaptively rather than relying on augmentations at the input level. We will include this analysis and discussion in the revised manuscript.
Baseline selection [R2.2, R3.4]:
We selected baselines to cover both canonical references and the current state of the art. nnU-Net serves as the standard U-Net anchor, while ESPNet (MICCAI 2025) reports the strongest cross-center generalization among recent methods (PraNet, SANet, MEGANet, Polyp-PVT, U-KAN). We omitted weaker baselines due to space. We also evaluated SAM-based methods, including fine-tuned variants, but found they did not outperform ESPNet on this protocol.
Existing polyp segmentation methods address illumination variation only indirectly through methods such as preprocessing or photometric augmentation. These methods operate at the image level rather than the segmentation network. KAN-AINet is the architecture embedding illumination correction at the feature level via KAN-IMM. With our experiments in augmentation-based ablation above, photometric augmentation at pre-processing level performance is still below KAN-AINet that modulates illumination in the feature levels.
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 methodological contribution with KAN-AINet for illumination-robust polyp segmentation and demonstrates consistent cross-dataset generalization. Reviewers acknowledge the relevance and empirical validation, but raise concerns regarding unclear module formulation, incomplete experimental design, missing architectural details, and high computational cost without latency evaluation. For rebuttal, authors should strengthen ablations and baselines, and provide efficiency and real-time performance analysis to support practical applicability.
- 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 accepted despite mixed reviews due to overall positive consensus and sufficient methodological contribution.
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 responses addressed the concerns about unclear module formulation, incomplete experimental design, missing architectural details, and high computational cost without latency evaluation.
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.
I recommend rejection. While the paper addresses an important problem—generalizable polyp segmentation under illumination variation—the necessity of introducing KAN is not convincingly established. The proposed KAN-IMM and KAN-BAM largely act as adaptive feature-gating and boundary-attention modules, but the paper does not compare them against simpler and more standard alternatives such as MLP-based gating, SE/CBAM-style attention, dynamic convolution, illumination normalization, or parameter-matched non-KAN modules. As a result, the reported gains cannot be attributed specifically to KAN rather than to increased model capacity, additional attention mechanisms, or architectural tuning.
The empirical evidence is also insufficient to justify KAN as the key contribution. The improvement over the strongest baseline, ESPNet, is modest on average and inconsistent across datasets; for example, the proposed method is worse than ESPNet on CVC-ColonDB in Dice and IoU, and only slightly better on Kvasir-Sessile. The ablation results show that KAN modules help compared with the authors’ own no-KAN variant, but this is not a strong test of KAN’s necessity because no parameter-matched or functionally equivalent non-KAN variant is included. Moreover, some single-module variants are competitive with or better than the full model on certain metrics, weakening the claim that the complete KAN design is consistently necessary.
The illumination robustness analysis is relevant but not fully convincing. Brightness-stratified results only compare KAN-AINet with its no-KAN counterpart and do not show that KAN is superior to conventional illumination-aware designs or augmentation-based solutions. The claimed boundary benefits are also only partially supported, since HD95/ASD are not reported for the main SOTA comparator ESPNet. Overall, the paper demonstrates that adding the proposed modules may improve performance, but it does not demonstrate that KAN is a necessary or uniquely effective mechanism for polyp segmentation. The methodological novelty and experimental justification are therefore not strong enough for acceptance.
