acceptodds
Withdrawn from ICLR 2027

Deformable Kolmogorov-Arnold Networks for Medical Image Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Medical image segmentation remains difficult when boundaries are faint and anatomy varies across modalities. U-Net variants replace the bottleneck with a transformer, an MLP, a state-space model, or a Kolmogorov-Arnold network (KAN). These variants still evaluate that operator on the encoder's fixed, downsampled grid. We present UKAD, a deformable KAN encoder-decoder. Its Kolmogorov-Arnold Displacement (KAD) bottleneck predicts a content-adaptive grid, resamples the bottleneck features on that grid, and then applies a learnable activation. That activation is a low-degree group-rational function, and we share it within each channel group, so its parameter count does not grow with network width. UKAD also uses it in the skip gates and in a hybrid decoder. We train UKAD in one stage, without pretraining or distillation, and with one hyperparameter setting across datasets, resolutions, and widths. We retrain fourteen U-Net, MLP, state-space, and KAN baselines on four public datasets under one protocol. UKAD-L has the highest mean IoU and Dice, 83.11 and 90.25. AdaKAN reaches 82.36 and 89.78, and Rolling-UNet reaches 82.25 mean IoU. UKAD-B reaches 82.59 mean IoU with 45% fewer parameters than AdaKAN. We release the code at https://anonymous.4open.science/r/UKAD.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Segmentation, Kolmogorov-Arnold Networks, Deformable Sampling, Group-Rational Activation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iPa6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes UKAD, a deformable KAN encoder-decoder for medical image segmentation, addressing the limitation that existing KAN-based U-Nets evaluate learnable activation functions on the encoder's fixed downsampled grid. The primary contribution is the KAD bottleneck block, which predicts a content-adaptive displacement grid, resamples the bottleneck features on that grid, and applies low-degree group-rational activation functions whose parameter count does not scale with channel width. Furthermore, UKAD reuses this rational activation across skip-connection gates and a hybrid decoder, accompanied by an uncertainty-gated boundary refinement head. Under a unified training and splitting protocol, the authors retrain fourteen baseline models on four public 2D medical image segmentation benchmarks for comparison.

Critical strengths
  1. Instead of copying metrics from previous literature, the authors rigorously retrained fourteen mainstream and recent baseline models under a unified codebase, shared data splits, and identical training hyperparameters, complemented by paired statistical significance tests and bootstrap confidence intervals, providing substantial empirical reference value.
  2. Adopting low-degree Padé-style rational functions shared across channel groups requires only 48 additional parameters per layer, effectively addressing the quadratic parameter explosion and initialization instability inherent in conventional B-spline KANs.
Critical weaknesses
  1. The two-by-two ablation in supplementary Table 22 shows that the default UKAD-B combining both deformable displacement and group-rational activation achieves a four-dataset mean IoU of 82.59%, which is lower than using rational activation alone 82.69% and lower than using displacement alone 82.67%, demonstrating that combining displacement resampling with KAN yields negative synergy and directly undermines the core empirical hypothesis.
  2. The primary experiments in Tables 1 and 2 report performance directly on the validation partition used for checkpoint selection, conflating model selection with unbiased evaluation and lacking independent held-out test sets to prove true generalization on three of the four datasets.
  3. Beyond the KAD bottleneck, the architecture stacks numerous confounding add-ons including dilated context branches, adaptive context mixers, a hybrid KAN decoder, and an uncertainty-gated boundary head, and Table 21 shows that removing these modules causes noticeable performance drops, making it difficult to disentangle whether the gains stem from the core KAN formulation or from extensive architectural engineering.
  4. Although the parameter count is constrained to 7.77M, spatial bilinear resampling and hybrid decoding incur a high computational cost of 74.46 GFLOPs, which is 8.5 times that of U-KAN, along with substantially higher latency on an A100 GPU, weakening its practical clinical utility.
Questions for authors
  1. How do the authors explain the negative interaction in supplementary Table 22, where combining deformable displacement with group-rational activation yields a lower average IoU than using either component individually?
  2. Why were strictly independent held-out test sets omitted for BUSI, CVC-ClinicDB, and GlaS, and how do the authors justify the potential optimistic bias of reporting metrics on the checkpoint-selection validation splits?
  3. What is the net performance gain of the deformable KAN bottleneck over a standard convolutional baseline when all confounding add-ons, including the boundary head, auxiliary losses, and context branches, are completely removed?
  4. Given that clinical medical imaging extensively relies on 3D volumetric scans, how can this 2D deformable grid sampling operator be practically generalized to 3D convolutions and anisotropic voxel resolutions?
Reviewer_EmNqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a U-shaped encoder-decoder network that uses a Kolmogorov-Arnold Displacement (KAD) block as the model’s bottleneck. The KAD block consists of three branches. The displacement branch computes a raw offset and a gate logit for every pixel, leading to an adaptive grid. Paired with a context and a local branch, the output of the KAD block is the residual sum of all three branches. The authors show that the proposed approach outperforms U-Net, MLP, state-space and KAN baselines.

Critical strengths

The authors numerically verify the proposed approach for an extensive set of experiments.

Critical weaknesses
  • The paper is hard to understand as too much necessary information has been moved to the appendix even though the page limit has not been exploited. The paper lacks explanations about important details of the proposed method. The different branches, particularly the context and local branch, are not explained with enough rigour. While the displacement branch is highlighted, the other two branches are not defined properly which makes it hard to understand the overall approach. The different terms of the loss function in Section 3.7 used for training the network are not properly defined. A motivation for why each of the non-standard terms are required is not given.
  • Hyperparameter choices are not explained. The weighting of the loss terms in Equation 6 needs further detailing as well as the choice of the parameters p,qp, q and GG for the group-rational activations.
Questions for authors
  • Can you motivate and detail why an adaptive grid rather than a fixed grid seems beneficial in this case? What is the purpose of the local branch?
  • Can you add mathematical definitions of the context and the local branch?
  • Please also properly define the training objective. Why did you incorporate the additional loss terms L∂\mathcal{L}_{\partial}, L(3)\mathcal{L}^{(3)}, L(2)\mathcal{L}^{(2)} and how are they mathematically defined? How are z’,Y,b,Y∂z’, Y, b, Y_{\partial} defined?
  • Specifically the choices for p,q,Gp, q, G in the group-rational activations should be motivated. An ablation study showing the influence of different values for those hyperparameters would be nice to include.
Reviewer_CmqxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents UKAD, a deformable Kolmogorov-Arnold Network for medical image segmentation. The main idea is to combine content-adaptive feature sampling with group-rational activation functions in the bottleneck. The method also includes KAN-based skip gates, a hybrid decoder, and a boundary refinement head. The authors evaluate UKAD on four medical image segmentation datasets and compare it with fourteen baseline methods.

Critical strengths
  1. The authors conduct experiments on four datasets and retrain fourteen baselines under a unified protocol, which makes the comparison more consistent. E.g., section 4.1 and Table 7 specify sources and retained settings for fourteen baselines, while Table 1 reports results over the same three split manifests.

  2. The paper provides relatively detailed ablation studies and reports different metrics, including IoU, Dice, HD95, FLOPs and inference speed. The authors also release code, configurations and split manifests.

Critical weaknesses
  1. My main concern is the limited dataset diversity. All four datasets focus on binary segmentation, while more challenging multi-organ datasets, such as Synapse, are not included. Since multi-class segmentation is a common task in medical image segmentation, it remains unclear whether the proposed method can effectively handle more complex anatomical structures.

  2. Section 4.1 states that each model checkpoint is selected using its validation partition and metrics are reported on that same partition. More importantly, the authors state "We chose the UKAD design on these same splits", which may lead to overly optimistic results. Moreover, although UKAD-L achieves higher mean IoU than AdaKAN, the improvement is not statistically significant in the run-level comparison (pHolm=0.092 in Table 3). I suggest evaluating the final model on independent test sets, such as the official GlaS test split, and conducting experiments with more training seeds (currently only one) to better verify the reliability of the reported improvements.

  3. According to Table 1, UKAD-B achieves slightly higher mean IoU than AdaKAN (82.59 vs. 82.36), but requires nearly twice the FLOPs (74.46 vs. 38.54 GFLOPs). It is unclear whether such small performance gain justifies the additional computational cost. A more detailed discussion of the accuracy-efficiency trade-off would be helpful.

  4. Although fourteen methods are compared, some representative recent methods are missing, including KAN-based methods such as U-GRKAN and KANSeg, and Vision Mamba-based methods such as VM-UNet and VM-UNetV2. Including these baselines would make the comparison more comprehensive.

Questions for authors
  1. Could the authors provide additional results on more challenging multi-organ segmentation datasets, such as Synapse, to demonstrate the generalization ability of UKAD?

  2. Could the authors clarify how the UKAD architecture was selected using the validation splits? Is it possible to evaluate the final model on independent test sets, such as the official GlaS test set, and report results with multiple training seeds?

  3. Given the relatively small performance improvement over AdaKAN, how do the authors justify the additional computational cost of UKAD? Could the authors provide a more detailed analysis of the accuracy-efficiency trade-off?

  4. Could the authors include comparisons with more recent KAN-based and Vision Mamba-based segmentation methods, such as U-GRKAN, KANSeg, VM-UNet, and VM-UNetV2, under the same experimental settings?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.