Pith. sign in

REVIEW 3 major objections 7 minor 25 references

KAN-Mamba FusionNet: Redefining Medical Image Segmentation with Non-Linear Modeling

T0 review · 3 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read KAN-Mamba FusionNet claims top IoU and F1 across BUSI, Kvasir-Seg, and GlaS by combining learnable nonlinearities with state-space long-range modeling.

desk verdict KAMBA block is a real but modest new assembly of known parts; the central performance claim rests on a single uncontrolled validation split. read the letter →

arxiv 2411.11926 v2 pith:QKTYJAFK submitted 2024-11-18 cs.CV

Akansh Agrawal , Akshan Agrawal , Shashwat Gupta , Priyanka Bagade This is my paper · ORCID
classification cs.CV
keywords medicalimagesegmentationKolmogorov-ArnoldnetworksstatespacemodelsMambaU-KANbagofactivationfunctionsspatialattentionKAMBAblock
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that medical image segmentation can be improved by joining two ideas that have complementary weaknesses: Kolmogorov-Arnold networks capture nonlinear structure through learnable activations but miss long-range context, while state-space models capture long-range dependencies but rely on fixed activations. It assembles both, together with spatial attention and a learned mixture of activation functions, into a KAMBA block placed inside a U-KAN-style encoder-decoder, yielding KAN-Mamba FusionNet. On breast ultrasound (BUSI), polyp (Kvasir-Seg), and gland (GlaS) datasets, the model reports the highest IoU and F1 among U-Net, U-NeXt, Rolling-UNet, U-Mamba, and Seg. U-KAN. The practical interest is that the gains come at a modest computational cost compared with transformer- or Mamba-heavy alternatives.

What carries the argument

The load-bearing object is the KAMBA block, whose output is $Z''_{\mathrm{out}} \oplus \Psi(Z_{\mathrm{in}}) \oplus Z_{\mathrm{in}}$, where $Z''_{\mathrm{out}}$ is the result of KAN, activation mixture, state-space model, and spatial attention in sequence. The activation mixture, called bag of activation functions (BoA), is $\Psi(Z) = \sum_{p=1}^{P} \alpha_p \psi_p(Z)$ with $\sum_p \alpha_p = 1$ and learnable weights, replacing a single fixed activation. The state-space model supplies linear-complexity long-range dependency modeling, and the spatial attention module uses a $7\times7$ convolution over pooled features to emphasize boundaries of irregular structures. Together these components are what the paper credits for the reported accuracy gains.

What would settle it

Train the five compared methods and KAN-Mamba FusionNet on identical multi-seed splits of BUSI, Kvasir-Seg, and GlaS and compute IoU and F1 on held-out test folds; if Seg. U-KAN or U-Mamba reaches or exceeds the reported numbers on those folds, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a single block can jointly handle the nonlinear appearance of anatomical structures and the long-range spatial context needed to segment them, and that doing so improves measured segmentation quality. The KAMBA block processes features through a KAN layer, a learned weighted sum of five activation functions (ReLU, tanh, Softplus, GELU, SiLU), a state-space model, and a spatial attention module, then fuses the result with a parallel activation mixture and the original input through a skip connection. On the three datasets the paper reports IoU/F1 of 65.84/79.17, 76.95/86.66, and 85.73/92.29, each above the corresponding values for all five compared methods. Ablation results show that adding KAN, BoA, and spatial attention to a classical Mamba backbone raises IoU from 60.87 to 65.84 on BUSI.

Load-bearing premise

The claimed superiority rests on validation scores from a single unseeded 4:1 random split of each dataset, so if that split is unrepresentative, the reported advantages may not reproduce.

Editorial extensions

If this is right

  • The KAMBA block can be inserted into a U-KAN pipeline and improve validation IoU and F1 on ultrasound, endoscopic, and histology images without adding much computational cost (13.96 GFLOPs and 6.37 million parameters).
  • The reported margins, from 0.3 to 2.2 IoU points over Seg. U-KAN, suggest that the main benefit is a consistent small gain rather than a large architectural reset.
  • The ablation evidence implies that each added component, KAN, BoA, and spatial attention, contributes positively, with the full KAMBA block outperforming every intermediate configuration.
  • Because the model uses a linear-complexity state-space model, it is a candidate for segmenting large images or 3D volumes where transformer attention would be too expensive.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: the paper's own comparison is on validation splits from a single random 4:1 split, so the most direct extension is to measure test-set performance over multiple fixed seeds to see whether the reported margins persist out-of-sample.
  • Inference: the BoA weights could be inspected after training; if one activation dominates on a given dataset, the bag may be acting mainly as a regularizer rather than as a genuinely diversified nonlinearity.
  • Inference: the same block structure should transfer to 3D segmentation or video frames, where long-range temporal or volumetric context is even more important and the state-space component has a larger relative role.
  • Inference: replacing the fixed sigmoid in the spatial attention module with a KAN-style learnable activation would remove the last static nonlinearity, a direct test of whether the nonlinearity argument extends to attention itself.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes KAN-Mamba FusionNet, a U-Net-like segmentation architecture that inserts a newly designed KAMBA block—comprising a KAN layer, a bag-of-activations (BoA) module, a state-space model (SSM), a spatial attention module, and a skip connection—into the U-KAN pipeline. The model is evaluated on three public medical segmentation datasets (BUSI, Kvasir-Seg, GlaS) and compared against U-Net, U-NeXt, Rolling-UNet, U-Mamba, and Seg. U-KAN using IoU and F1 scores (Table 1). The paper also reports model complexity (Table 2) and ablation studies on the BUSI dataset (Tables 3-4). The central claim is that the proposed architecture consistently outperforms state-of-the-art methods on these datasets.

Significance. If the reported gains hold under a statistically controlled evaluation, the architecture would be a useful incremental contribution to the line of work combining Kolmogorov-Arnold Networks and Mamba for medical image segmentation. The paper's strengths are the use of three public datasets, a clear description of the proposed KAMBA block, and the inclusion of complexity comparisons and ablations. However, the empirical claim currently rests entirely on validation scores from a single unseeded split per dataset, with no significance testing and no evidence that the baselines were tuned with the same care. The paper does not provide code or pretrained models, which limits reproducibility. These issues are load-bearing because every quantitative conclusion in the abstract and conclusion depends on the validity of the Table 1 comparison.

major comments (3)
  1. [Sections 4.1-4.2, Table 1] The central claim that KAN-Mamba FusionNet 'consistently outperforms' the baselines is not established by the reported protocol. Each dataset is split once into a 4:1 train/validation ratio with no fixed seed and no separate held-out test set, and the three 'independent runs' repeat the same split, so the reported plus-or-minus values capture only initialization noise, not split variability. On Kvasir-Seg and GlaS, the IoU advantage over Seg. U-KAN is 1.10 and 0.28 points, respectively, while the reported standard deviations for the baseline are 0.53 and 0.34; without a paired significance test, these margins are not sufficient to support the claim of consistent superiority. I recommend reporting test-set metrics from multiple independent splits (e.g., five random seeds or k-fold cross-validation) with paired bootstrap or Wilcoxon signed-rank tests, and clearly distinguishing validation from test performance.
  2. [Section 4.1, Table 1] The comparison against baselines is uncontrolled with respect to training protocol. The paper does not state whether U-Net, U-NeXt, Rolling-UNet, U-Mamba, and Seg. U-KAN were trained with the same optimizer, learning rate schedule, loss function, data augmentation, and number of epochs, nor whether their hyperparameters were tuned on the same validation splits. Without this information, part of the reported margin may be attributable to under-tuned baselines rather than to the proposed architecture. The authors should provide the exact training configuration used for every baseline, ideally with the same hyperparameter search budget, so that the comparison is fair.
  3. [Section 4.4, Tables 3-4] The ablation studies are performed only on the BUSI dataset and report single-run values without standard deviations. For example, the improvement of BoA over a single ReLU is +1.23 IoU (from 64.61 to 65.84), which is comparable to the run-to-run noise reported for the full model in Table 1 (±0.54 IoU on BUSI). Additionally, the 'Classical Mamba' baseline in Table 3 is not defined (which Mamba variant, what backbone), making the ablation irreproducible. Please repeat the ablations over multiple seeds, report means and standard deviations, and define the classical Mamba reference precisely.
minor comments (7)
  1. [Section 3] The BUSI dataset description is ambiguous: 'We utilized the images representing breast cancer' could mean excluding the 133 normal images, but this is not stated explicitly; please specify the exact number of images used per class and how the split was applied.
  2. [Section 2.2, Eq. (6)] The symbol ⊕ in Equation (6) is not defined; please state that it denotes concatenation and clarify how the channel dimensions of Z''_out, Ψ(Zin), and Zin are aligned.
  3. [Section 2.2, Eq. (3)] The summation notation in Equation (3) appears garbled in the typeset version ('P ∑ p=1'); please correct the notation and state how the constraint Σαp=1 is enforced during optimization (e.g., softmax normalization) and how αp is initialized.
  4. [Section 4.4] The term 'Classical Mamba' in Table 3 is not defined; specify which Mamba architecture and backbone are used so the ablation can be reproduced.
  5. [References] Reference [18] points to Kiu-Net rather than the original GlaS dataset paper (Sirinukunwattana et al., MICCAI 2015); please cite the original dataset source.
  6. [Table 2] The GFLOPs numbers depend on the input resolution and batch size; please specify these settings so that the complexity comparison is meaningful.
  7. [Throughout] The terms 'consistently outperforms' (abstract and conclusion) and 'significantly fewer false positive regions' (Section 4.3 and Fig. 2 caption) are stronger than the statistical evidence supports; please soften these statements or provide quantitative false-positive analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported results are direct empirical measurements on public benchmarks against external baselines.

full rationale

The paper makes no formal derivation or prediction from first principles; its central claim is that the proposed KAN-Mamba FusionNet achieves higher IoU and F1 scores than five external baselines on three public datasets (BUSI, Kvasir-Seg, GlaS). The KAMBA block is explicitly defined in Equations (5) and (6) as a composition of a KAN block, a bag-of-activations layer, an SSM layer, a spatial attention module, and a skip connection; this is an architectural construction, not a quantity fitted to the reported metrics. The BoA layer in Equation (3) is a learnable convex combination of standard activation functions and is evaluated by ablation against no-activation and single-ReLU alternatives, so it is not a hidden way of encoding the target scores. All performance numbers in Table 1 are measured validation scores on public data, and lower-scoring baselines are independently published methods, not renormalized versions of the proposed model. The cited works on KAN, U-KAN, Mamba, and U-Mamba are external prior art, and there is no self-citation chain that carries the argument. The main limitations are statistical—a single unseeded 4:1 split with no separate test set and no significance testing—but those concerns affect generalizability and correctness risk, not circularity. The superiority claim is therefore not equivalent to its inputs by construction, and no circular step can be identified from the paper's own equations or citations.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The paper's claims rest on domain assumptions from the cited literature (KANs capture non-linearity, SSMs handle long-range dependencies, the U-KAN backbone is suitable) and on an evaluation protocol with an unspecified random split. No new physical or mathematical entities are introduced, and no constants are fitted outside standard model training.

free parameters (1)
  • BCE loss weight = 0.5
    Hand-chosen weight in the combined loss (Eq. 9); the paper does not ablate this value, so the final scores are conditional on this choice.
assumptions (5)
  • domain assumption KAN layers capture non-linearity better than conventional activations in segmentation.
    Invoked in Section 2.1 and relied on for the KAN block in the KAMBA design; supported only by citations [11,14].
  • domain assumption SSMs efficiently capture long-range dependencies with linear complexity.
    Invoked in Section 2.2 to justify adding an SSM stage; supported by citations [6,15].
  • domain assumption The U-KAN encoder-decoder pipeline is an appropriate backbone for medical image segmentation.
    The architecture is built by inserting KAMBA into U-KAN (Section 2); this base is taken as given.
  • domain assumption A 4:1 random split of each dataset provides a valid evaluation protocol.
    Stated in Section 4.1 without seed or stratification; all reported comparisons depend on this split.
  • domain assumption Spatial attention (CBAM formulation) improves local feature extraction.
    Used in Equation (4) and incorporated in the KAMBA block; no ablation isolates its contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KAN-Mamba FusionNet: Redefining Medical Image Segmentation with Non-Linear Modeling." pith.science (2026). https://pith.science/paper/QKTYJAFK

@misc{pith2026241111926,
  author       = {Pith},
  title        = {Pith review of: KAN-Mamba FusionNet: Redefining Medical Image Segmentation with Non-Linear Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QKTYJAFK}},
  note         = {Machine review of arXiv:2411.11926}
}
read the original abstract

Medical image segmentation is essential for applications like robotic surgeries, disease diagnosis, and treatment planning. Recently, various deep-learning models have been proposed to enhance medical image segmentation. One promising approach utilizes Kolmogorov-Arnold Networks (KANs), which better capture non-linearity in input data. However, they are unable to effectively capture long-range dependencies, which are required to accurately segment complex medical images and, by that, improve diagnostic accuracy in clinical settings. Neural networks such as Mamba can handle long-range dependencies. However, they have a limited ability to accurately capture non-linearities in the images as compared to KANs. Thus, we propose a novel architecture, the KAN-Mamba FusionNet, which improves segmentation accuracy by effectively capturing the non-linearities from input and handling long-range dependencies with the newly proposed KAMBA block. We evaluated the proposed KAN-Mamba FusionNet on three distinct medical image segmentation datasets: BUSI, Kvasir-Seg, and GlaS - and found it consistently outperforms state-of-the-art methods in IoU and F1 scores. Further, we examined the effects of various components and assessed their contributions to the overall model performance via ablation studies. The findings highlight the effectiveness of this methodology for reliable medical image segmentation, providing a unique approach to address intricate visual data issues in healthcare.

Figures

Figures reproduced from arXiv: 2411.11926 by the authors.

Figure 1
Figure 1. An overview of the proposed KAN-Mamba FusionNet Architecture with the de [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualized segmentation results on three distinct datasets: BUSI, Kvasir-Seg, and [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 13 canonical work pages

  1. [1]

    Al-Dhabyani, M

    W. Al-Dhabyani, M. Gomaa, H. Khaled, and A. Fahmy. Dataset of breast ultrasound images. Data in Brief, 28:104863, 2020. doi: 10.1016/j.dib.2019.104863

    arXiv 2020

  2. [2]

    Amo Alonso, J

    C. Amo Alonso, J. Sieber, and M. N. Zeilinger. State space models as foundation models: A control theoretic overview, 2024. arXiv preprint arXiv:2403.16899

    arXiv 2024

  3. [3]

    Bertels, D

    J. Bertels, D. Robben, R. Lemmens, and D. Vandermeulen. Convolutional neural net- works for medical image segmentation, 2022. arXiv preprint arXiv:2211.09562

    arXiv 2022

  4. [4]

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou. Transunet: Transformers make strong encoders for medical image segmentation, 2021. arXiv preprint arXiv:2102.04306

    arXiv 2021

  5. [5]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. arXiv preprint arXiv:2010.11929

    arXiv 2021

  6. [6]

    Gu and T

    A. Gu and T. Dao. Mamba: Linear-time sequence modeling with selective state spaces,

    work page

  7. [7]

    Hesamian, W

    M.H. Hesamian, W. Jia, X. He, et al. Deep learning techniques for medical image segmentation: Achievements and challenges. Journal of Digital Imaging, 32:582–596, 2019

    work page 2019

  8. [8]

    Huang, L

    H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y . Iwamoto, X. Han, Y .-W. Chen, and J. Wu. Unet 3+: A full-scale connected unet for medical image segmentation, 2020. arXiv preprint arXiv:2004.08790

    arXiv 2020

Show all 25 references
  1. [9]

    D. Jha, P. H. Smedsrud, M. A. Riegler, P.Å. Halvorsen, T. de Lange, D. Johansen, and H.Å. Johansen. Kvasir-seg: A segmented polyp dataset. In International Conference on Multimedia Modeling, pages 451–462. Springer, 2020

    2020

  2. [10]

    LeCun, L

    Y . LeCun, L. Bottou, G.B. Orr, and K.-R. Müller. Efficient backprop. In Neural net- works: Tricks of the trade, pages 9–50. Springer, 1998

    1998

  3. [11]

    C. Li, X. Liu, W. Li, C. Wang, H. Liu, Y . Liu, Z. Chen, and Y . Yuan. U-kan makes strong backbone for medical image segmentation and generation, 2024. arXiv preprint arXiv:2406.02918. A. AGRAW AL, A. AGRAW AL, S. GUPTA, P . BAGADE: KAN-MAMBA FUSIONNET 11

    2024 arXiv

  4. [12]

    Litjens, T

    G. Litjens, T. Kooi, B. Ehteshami Bejnordi, A. Adiyoso Setio, F. Ciompi, M. Ghafoo- rian, J. A.W.M. van der Laak, B. van Ginneken, and C. I. Sánchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42:60–88, 2017

    2017

  5. [13]

    Y . Liu, H. Zhu, M. Liu, H. Yu, Z. Chen, and J. Gao. Rolling-unet: Revitalizing mlp’s ability to efficiently extract long-distance dependencies for medical image segmenta- tion. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 3819–3827, 2024

    2024

  6. [14]

    Z. Liu, Y . Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Solja ˇci´c, T. Y . Hou, and M. Tegmark. Kan: Kolmogorov-arnold networks, 2024. arXiv preprint arXiv:2404.19756

    2024 arXiv

  7. [15]

    J. Ma, F. Li, and B. Wang. U-mamba: Enhancing long-range dependency for biomedi- cal image segmentation, 2024. arXiv preprint arXiv:2401.04722

    2024 arXiv

  8. [16]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI 2015, pages 234–241. Springer, 2015

    2015

  9. [17]

    J. M. J. Valanarasu and V . M. Patel. Unext: Mlp-based rapid medical image seg- mentation network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 23–33. Springer, 2022

    2022

  10. [18]

    J. M. J. Valanarasu, V . A. Sindagi, I. Hacihaliloglu, and V . M. Patel. Kiu-net: Towards accurate segmentation of biomedical images using over-complete representations. In MICCAI 2020 Lecture Notes in Computer Science , pages 363–373. Springer Interna- tional Publishing, 2020

    2020

  11. [19]

    J. M. J. Valanarasu, P. Oza, I. Hacihaliloglu, and V .M. Patel. Medical transformer: Gated axial-attention for medical image segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2021, pages 24–36, Strasbourg, France,

    2021

  12. [20]

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19. Springer, 2018

    2018

  13. [21]

    Z. Xing, T. Ye, Y . Yang, G. Liu, and L. Zhu. Segmamba: Long-range sequen- tial modeling mamba for 3d medical image segmentation, 2024. arXiv preprint arXiv:2401.13560

    2024 arXiv

  14. [22]

    Yamashita, M

    R. Yamashita, M. Nishio, R.K.G. Do, et al. Convolutional neural networks: an overview and application in radiology. Insights Imaging, 9:611–629, 2018

    2018

  15. [23]

    Zhou, M.M

    Z. Zhou, M.M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Im- age Analysis and Multimodal Learning for Clinical Decision Support , pages 3–11. Springer, 2018

    2018

  16. [24]

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang. Vision mamba: Efficient vi- sual representation learning with bidirectional state space model, 2024. arXiv preprint arXiv:2401.09417

    2024 arXiv

  17. [2023]

    arXiv preprint arXiv:2312.00752

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.