pith. machine review for the scientific record. sign in

arxiv: 2604.20286 · v1 · submitted 2026-04-22 · 💻 cs.CV · cs.AI

Recognition: unknown

MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation

Md Maklachur Rahman , Soon Ki Jung , Tracy Hammond
Authors on Pith no claims yet

Pith reviewed 2026-05-10 00:20 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords skin lesion segmentationMamba state space modelsU-Net architecturefeature fusionmedical image analysisdomain generalizationcomputational efficiency
0
0 comments X

The pith

MambaLiteUNet embeds Mamba state-space modeling inside a U-Net and adds three fusion and gating modules to reach higher accuracy on skin lesion boundaries with far fewer parameters and operations than prior models.

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

The paper presents MambaLiteUNet as a compact segmentation network that places Mamba blocks inside a U-Net backbone and introduces Adaptive Multi-Branch Mamba Feature Fusion, Local-Global Feature Mixing, and Cross-Gated Attention. These additions are meant to strengthen local-global feature exchange and preserve fine spatial details during skip connections. On four standard skin-lesion benchmarks the model records an average IoU of 87.12 percent and Dice of 93.09 percent, exceeding earlier networks while cutting parameters by 93.6 percent and GFLOPs by 97.6 percent relative to a plain U-Net. A reader would care because precise lesion outlines support earlier skin-cancer diagnosis and because low compute makes the method usable in clinics with modest hardware. The authors further report that the same network generalizes to six unseen lesion categories with 77.61 percent IoU.

Core claim

By integrating Mamba state-space modeling into the U-Net encoder-decoder and equipping it with the AMF, LGFM, and CGA modules, the authors obtain average IoU of 87.12 percent and Dice of 93.09 percent across ISIC2017, ISIC2018, HAM10000, and PH2 while reducing parameters by 93.6 percent and GFLOPs by 97.6 percent compared with a standard U-Net; the same model also leads all tested networks on domain-generalization tests with six unseen lesion types.

What carries the argument

The Adaptive Multi-Branch Mamba Feature Fusion (AMF), Local-Global Feature Mixing (LGFM), and Cross-Gated Attention (CGA) modules that improve local-global interaction and skip-connection quality inside the Mamba-UNet hybrid.

Load-bearing premise

The reported accuracy and efficiency gains arise primarily from the AMF, LGFM, and CGA modules rather than from dataset-specific training schedules or post-training model selection.

What would settle it

A controlled ablation that removes the AMF, LGFM, and CGA modules from the identical Mamba-UNet backbone and retrains on the same four benchmarks, then checks whether IoU and Dice fall to levels comparable with prior state-of-the-art models.

Figures

Figures reproduced from arXiv: 2604.20286 by Md Maklachur Rahman, Soon Ki Jung, Tracy Hammond.

Figure 1
Figure 1. Figure 1: Complexity–performance trade-off of SOTA segmentation models based on average IoU and average DSC across ISIC2017, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Overall MambaLiteUNet pipeline. (b) Mamba block: Integrate SS2D [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Domain generalization ranking (average IoU vs. average [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on ISIC2017, ISIC2018, HAM10000, and PH2. Red outlines highlight segmentation errors and weak [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative visualization of module contributions. The [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Feature map visualization in MambaLiteUNet. Top row (left [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
read the original abstract

Recent segmentation models have demonstrated promising efficiency by aggressively reducing parameter counts and computational complexity. However, these models often struggle to accurately delineate fine lesion boundaries and texture patterns essential for early skin cancer diagnosis and treatment planning. In this paper, we propose MambaLiteUNet, a compact yet robust segmentation framework that integrates Mamba state space modeling into a U-Net architecture, along with three key modules: Adaptive Multi-Branch Mamba Feature Fusion (AMF), Local-Global Feature Mixing (LGFM), and Cross-Gated Attention (CGA). These modules are designed to enhance local-global feature interaction, preserve spatial details, and improve the quality of skip connections. MambaLiteUNet achieves an average IoU of 87.12% and average Dice score of 93.09% across ISIC2017, ISIC2018, HAM10000, and PH2 benchmarks, outperforming state-of-the-art models. Compared to U-Net, our model improves average IoU and Dice by 7.72 and 4.61 points, respectively, while reducing parameters by 93.6% and GFLOPs by 97.6%. Additionally, in domain generalization with six unseen lesion categories, MambaLiteUNet achieves 77.61% IoU and 87.23% Dice, performing best among all evaluated models. Our extensive experiments demonstrate that MambaLiteUNet achieves a strong balance between accuracy and efficiency, making it a competitive and practical solution for dermatological image segmentation. Our code is publicly available at: https://github.com/maklachur/MambaLiteUNet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript presents MambaLiteUNet, which combines Mamba state space modeling with a U-Net backbone and introduces three modules—Adaptive Multi-Branch Mamba Feature Fusion (AMF), Local-Global Feature Mixing (LGFM), and Cross-Gated Attention (CGA)—for improved skin lesion segmentation. The central claims are an average IoU of 87.12% and Dice score of 93.09% on four public benchmarks (ISIC2017, ISIC2018, HAM10000, PH2), outperforming prior methods while reducing parameters by 93.6% and GFLOPs by 97.6%, plus strong domain generalization results on unseen lesion categories.

Significance. Should the performance claims prove robust upon verification, the work offers a highly efficient segmentation model suitable for resource-constrained clinical settings in dermatology. The public code availability is a positive aspect that supports reproducibility. The integration of Mamba for medical imaging is timely given recent interest in state space models for vision tasks.

major comments (3)
  1. [Section 4] The results section reports substantial improvements but omits ablation studies that would isolate the effects of the AMF, LGFM, and CGA modules. For instance, performance with and without each module under fixed training conditions is not shown, which is necessary to confirm that the average 7.72 IoU and 4.61 Dice gains are attributable to these innovations rather than implementation or tuning differences.
  2. [Section 4.1] Insufficient details are provided on the experimental setup, including exact hyperparameters, whether all baseline models were retrained with the same data splits and augmentations, and the number of runs for averaging results. This information is critical to address concerns that the reported efficiency and accuracy benefits may arise from dataset-specific optimizations.
  3. [Tables in Section 4] The quantitative results lack error bars, standard deviations across multiple runs, or statistical significance testing (e.g., Wilcoxon signed-rank test), which are standard for establishing that MambaLiteUNet reliably outperforms the compared state-of-the-art models on the benchmarks.
minor comments (2)
  1. [Method section (Section 3)] The equations and diagrams for the proposed modules would benefit from more explicit notation, particularly for the cross-gating mechanism in CGA, to improve clarity for readers unfamiliar with Mamba adaptations.
  2. A few minor typographical inconsistencies in the abstract and introduction could be corrected during revision.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback, which has helped us strengthen the manuscript. We address each major comment below and have made revisions to improve the experimental rigor, clarity, and reproducibility of the work.

read point-by-point responses

  1. Referee: [Section 4] The results section reports substantial improvements but omits ablation studies that would isolate the effects of the AMF, LGFM, and CGA modules. For instance, performance with and without each module under fixed training conditions is not shown, which is necessary to confirm that the average 7.72 IoU and 4.61 Dice gains are attributable to these innovations rather than implementation or tuning differences.

    Authors: We agree that ablation studies are necessary to isolate the contributions of the AMF, LGFM, and CGA modules. In the revised manuscript, we have added a new subsection in Section 4 with comprehensive ablations. These experiments report IoU and Dice scores for the full MambaLiteUNet model as well as three variants (without AMF, without LGFM, and without CGA) trained under identical conditions, data splits, and hyperparameters. The results confirm incremental gains from each module, supporting that the reported improvements are attributable to the proposed components rather than other factors. revision: yes

  2. Referee: [Section 4.1] Insufficient details are provided on the experimental setup, including exact hyperparameters, whether all baseline models were retrained with the same data splits and augmentations, and the number of runs for averaging results. This information is critical to address concerns that the reported efficiency and accuracy benefits may arise from dataset-specific optimizations.

    Authors: We appreciate this observation and have substantially expanded Section 4.1 in the revision. The updated section now provides the complete list of hyperparameters (including optimizer, learning rate schedule, batch size, number of epochs, and loss function weights), explicitly states that all baseline models were retrained from scratch using the exact same data splits, augmentation pipelines, and training protocol as MambaLiteUNet, and clarifies that all quantitative results are averaged over five independent runs with different random seeds. revision: yes

  3. Referee: [Tables in Section 4] The quantitative results lack error bars, standard deviations across multiple runs, or statistical significance testing (e.g., Wilcoxon signed-rank test), which are standard for establishing that MambaLiteUNet reliably outperforms the compared state-of-the-art models on the benchmarks.

    Authors: We acknowledge the value of statistical reporting for establishing reliable superiority. In the revised tables of Section 4, we now include standard deviations computed across the five independent runs for both IoU and Dice scores. We have also added Wilcoxon signed-rank test p-values for pairwise comparisons between MambaLiteUNet and each baseline method, with significance levels indicated in the tables to demonstrate that the observed improvements are statistically significant. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark comparisons with no derivation chain

full rationale

The paper proposes three new modules (AMF, LGFM, CGA) inside a Mamba-UNet hybrid and reports average IoU/Dice on four public skin-lesion datasets plus a domain-generalization split. No equations, first-principles derivations, or fitted parameters are presented whose outputs are then relabeled as predictions; the performance numbers are direct empirical measurements against external baselines on fixed public benchmarks. The central claims therefore remain independent of any self-referential reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard supervised deep-learning assumptions (i.i.d. training and test images, cross-entropy or Dice loss, standard augmentations) plus the unproven premise that the three proposed modules are the decisive source of improvement. No new physical entities or non-standard mathematical axioms are introduced.

axioms (1)
  • domain assumption Standard supervised training on public dermatology datasets yields generalizable lesion boundaries
    Invoked implicitly when claiming outperformance and domain generalization on unseen lesion categories.

pith-pipeline@v0.9.0 · 5603 in / 1280 out tokens · 37369 ms · 2026-05-10T00:20:49.328856+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 22 canonical work pages · 5 internal anchors

  1. [1]

    Attention swin u-net: Cross-contextual at- tention mechanism for skin lesion segmentation

    Ehsan Khodapanah Aghdam, Reza Azad, Maral Zarvani, and Dorit Merhof. Attention swin u-net: Cross-contextual at- tention mechanism for skin lesion segmentation. In2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2023. 2, 5, 6, 7, 3

    2023

  2. [2]

    Dataset of breast ultrasound images.Data in brief, 28:104863, 2020

    Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images.Data in brief, 28:104863, 2020. 6, 2, 4

    2020

  3. [3]

    Loss functions in the era of semantic segmentation: A survey and outlook,

    Reza Azad, Moein Heidary, Kadir Yilmaz, Michael H¨uttemann, Sanaz Karimijafarbigloo, Yuli Wu, Anke Schmeink, and Dorit Merhof. Loss functions in the era of se- mantic segmentation: A survey and outlook.arXiv preprint arXiv:2312.05391, 2023. 4, 3

    work page arXiv 2023

  4. [4]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medi- cal image segmentation.arXiv preprint arXiv:2102.04306,

    work page internal anchor Pith review arXiv

  5. [5]

    Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imaging collaboration (isic).arXiv preprint arXiv:1902.03368, 2019. 2, 4, 1, 5

    work page Pith review arXiv 2018

  6. [6]

    Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kit- tler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomed- ical imaging (isbi), hosted by the international skin imaging collaboration (i...

    2017

  7. [7]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 4, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  8. [8]

    Sigmoid- weighted linear units for neural network function approxima- tion in reinforcement learning.Neural networks, 107:3–11,

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid- weighted linear units for neural network function approxima- tion in reinforcement learning.Neural networks, 107:3–11,

  9. [9]

    Dermatologist-level classification of skin cancer with deep neural networks.nature, 542(7639):115–118, 2017

    Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks.nature, 542(7639):115–118, 2017. 1

    2017

  10. [10]

    Wtcm-unet: A hybrid cnn-ssm framework combining wavelet transform for medical image segmentation.Biomedical Signal Processing and Control, 112:108525, 2026

    Zhihua Gan, Zhongxiang Xie, Yushu Zhang, Weihong Han, Bo Zhang, and Xiuli Chai. Wtcm-unet: A hybrid cnn-ssm framework combining wavelet transform for medical image segmentation.Biomedical Signal Processing and Control, 112:108525, 2026. 6, 1, 3

    2026

  11. [11]

    A Data-scalable Transformer for Medical Image Segmentation: Architecture , Model Efficiency, and Benchmark [preprint]

    Yunhe Gao, Mu Zhou, Di Liu, and Dimitris Metaxas. A multi-scale transformer for medical image segmentation: Architectures, model efficiency, and benchmarks.arXiv preprint arXiv:2203.00131, 2022. 2, 5, 6, 7, 3, 4

    work page arXiv 2022

  12. [12]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3, 5

    work page Pith review arXiv 2023

  13. [13]

    A survey, review, and future trends of skin lesion segmentation and classification.Computers in Biology and Medicine, 155:106624, 2023

    Md Kamrul Hasan, Md Asif Ahamad, Choon Hwai Yap, and Guang Yang. A survey, review, and future trends of skin lesion segmentation and classification.Computers in Biology and Medicine, 155:106624, 2023. 1

    2023

  14. [14]

    Gaussian Error Linear Units (GELUs)

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016. 4

    work page Pith review arXiv 2016

  15. [15]

    Devil is in channels: Contrastive single domain generalization for medical image segmentation

    Shishuai Hu, Zehui Liao, and Yong Xia. Devil is in channels: Contrastive single domain generalization for medical image segmentation. InInternational Conference on Medical Im- age Computing and Computer-Assisted Intervention, pages 14–23. Springer, 2023. 5, 6, 7, 2, 3, 4

    2023

  16. [16]

    Matchseg: To- wards better segmentation via reference image matching

    Jiayu Huo, Ruiqiang Xiao, Haotian Zheng, Yang Liu, S´ebastien Ourselin, and Rachel Sparks. Matchseg: To- wards better segmentation via reference image matching. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 2068–2073. IEEE, 2024. 4, 5, 6, 7

    2024

  17. [17]

    Lightm-unet: Mamba assists in lightweight unet for medical image segmentation.arXiv preprint arXiv:2403.05246, 2024

    Weibin Liao, Yinghao Zhu, Xinyuan Wang, Cehngwei Pan, Yasha Wang, and Liantao Ma. Lightm-unet: Mamba assists in lightweight unet for medical image segmentation.arXiv preprint arXiv:2403.05246, 2024. 2, 4, 5, 6, 7, 3

    work page arXiv 2024

  18. [18]

    Ds-transunet: Dual swin trans- former u-net for medical image segmentation.IEEE Trans- actions on Instrumentation and Measurement, 2022

    Ailiang Lin, Bingzhi Chen, Jiayu Xu, Zheng Zhang, Guang- ming Lu, and David Zhang. Ds-transunet: Dual swin trans- former u-net for medical image segmentation.IEEE Trans- actions on Instrumentation and Measurement, 2022. 2

    2022

  19. [19]

    arXiv preprint arXiv:2401.10166 , year=

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model.arXiv preprint arXiv:2401.10166,

    work page arXiv

  20. [20]

    SGDR: Stochastic Gradient Descent with Warm Restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts.arXiv preprint arXiv:1608.03983, 2016. 5

    work page Pith review arXiv 2016

  21. [21]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  22. [22]

    Ph 2-a dermoscopic image database for research and benchmarking

    Teresa Mendonc ¸a, Pedro M Ferreira, Jorge S Marques, Andr´e RS Marcal, and Jorge Rozeira. Ph 2-a dermoscopic image database for research and benchmarking. In2013 35th annual international conference of the IEEE engineer- ing in medicine and biology society (EMBC), pages 5437–

  23. [23]

    2, 4, 1, 3

    IEEE, 2013. 2, 4, 1, 3

    2013

  24. [24]

    Attention U-Net: Learning Where to Look for the Pancreas

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Atten- tion u-net: Learning where to look for the pancreas.arXiv preprint arXiv:1804.03999, 2018. 1, 2, 5

    work page internal anchor Pith review arXiv 2018

  25. [25]

    arXiv preprint arXiv:2410.03105 (2024)

    Md Maklachur Rahman, Abdullah Aman Tutul, Ankur Nath, Lamyanba Laishram, Soon Ki Jung, and Tracy Hammond. Mamba in vision: A comprehensive survey of techniques and applications.arXiv preprint arXiv:2410.03105, 2024. 2

    work page arXiv 2024

  26. [26]

    Aulunet: An adaptive ultra-lightweight u-net frame- work for efficient skin lesion segmentation in resource- constrained environments

    Md Maklachur Rahman, Soon Ki Jung, and Tracy Ham- mond. Aulunet: An adaptive ultra-lightweight u-net frame- work for efficient skin lesion segmentation in resource- constrained environments. In36th British Machine Vision 9 Conference 2025, BMVC 2025, Sheffield, UK, November 24- 27, 2025. BMV A, 2025. 1, 2

    2025

  27. [27]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InInternational Conference on Medical image com- puting and computer-assisted intervention, pages 234–241. Springer, 2015. 1, 2, 3, 5, 6, 7, 4

    2015

  28. [28]

    arXiv preprint arXiv:2402.02491 , year=

    Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation.arXiv preprint arXiv:2402.02491, 2024. 2, 4, 5, 6, 7, 3

    work page arXiv 2024

  29. [29]

    Malunet: A multi-attention and light- weight unet for skin lesion segmentation

    Jiacheng Ruan, Suncheng Xiang, Mingye Xie, Ting Liu, and Yuzhuo Fu. Malunet: A multi-attention and light- weight unet for skin lesion segmentation. In2022 IEEE In- ternational Conference on Bioinformatics and Biomedicine (BIBM), pages 1150–1156. IEEE, 2022. 2, 5, 6, 7, 3, 4

    2022

  30. [30]

    Ege-unet: an efficient group enhanced unet for skin lesion segmentation

    Jiacheng Ruan, Mingye Xie, Jingsheng Gao, Ting Liu, and Yuzhuo Fu. Ege-unet: an efficient group enhanced unet for skin lesion segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Interven- tion, pages 481–490. Springer, 2023. 1, 2, 5, 6, 7, 3, 4

    2023

  31. [31]

    Melanoma.The Lancet, 392(10151):971–984, 2018

    Dirk Schadendorf, Alexander CJ Van Akkooi, Carola Berk- ing, Klaus G Griewank, Ralf Gutzmer, Axel Hauschild, Andreas Stang, Alexander Roesch, and Selma Ugurel. Melanoma.The Lancet, 392(10151):971–984, 2018. 1

    2018

  32. [32]

    Gland segmentation in colon histology images: The glas challenge contest.Medical image analysis, 35:489–502,

    Korsuk Sirinukunwattana, Josien PW Pluim, Hao Chen, Xi- aojuan Qi, Pheng-Ann Heng, Yun Bo Guo, Li Yang Wang, Bogdan J Matuszewski, Elia Bruni, Urko Sanchez, et al. Gland segmentation in colon histology images: The glas challenge contest.Medical image analysis, 35:489–502,

  33. [33]

    The ham10000 dataset, a large collection of multi-source der- matoscopic images of common pigmented skin lesions.Sci- entific data, 5(1):1–9, 2018

    Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source der- matoscopic images of common pigmented skin lesions.Sci- entific data, 5(1):1–9, 2018. 1, 2, 4

    2018

  34. [34]

    Unext: Mlp- based rapid medical image segmentation network.arXiv preprint arXiv:2203.04967, 2022

    Jeya Maria Jose Valanarasu and Vishal M Patel. Unext: Mlp- based rapid medical image segmentation network.arXiv preprint arXiv:2203.04967, 2022. 2, 5, 6, 7, 3, 4

    work page arXiv 2022

  35. [35]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 4

    2017

  36. [36]

    Precise yet efficient semantic calibration and refine- ment in convnets for real-time polyp segmentation from colonoscopy videos

    Huisi Wu, Jiafu Zhong, Wei Wang, Zhenkun Wen, and Jing Qin. Precise yet efficient semantic calibration and refine- ment in convnets for real-time polyp segmentation from colonoscopy videos. InProceedings of the AAAI conference on artificial intelligence, pages 2916–2924, 2021. 2, 5, 6, 7, 3

    2021

  37. [37]

    Ultralight vm-unet: Parallel vision mamba significantly reduces parameters for skin lesion segmentation.arXiv preprint arXiv:2403.20035, 2024

    Renkai Wu, Yinghao Liu, Pengchen Liang, and Qing Chang. Ultralight vm-unet: Parallel vision mamba significantly reduces parameters for skin lesion segmentation.arXiv preprint arXiv:2403.20035, 2024. 2, 4, 5, 6, 7, 3

    work page arXiv 2024

  38. [38]

    H-vmunet: High-order vision mamba unet for medical image segmentation.Neurocomputing, 624:129447, 2025

    Renkai Wu, Yinghao Liu, Pengchen Liang, and Qing Chang. H-vmunet: High-order vision mamba unet for medical image segmentation.Neurocomputing, 624:129447, 2025. 6, 1, 3

    2025

  39. [39]

    Aggregated residual transformations for deep neural networks

    Saining Xie, Ross Girshick, Piotr Doll ´ar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500,

  40. [40]

    Lb-unet: A lightweight boundary-assisted unet for skin lesion segmentation

    Jiahao Xu and Lyuyang Tong. Lb-unet: A lightweight boundary-assisted unet for skin lesion segmentation. InIn- ternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 361–371. Springer,

  41. [41]

    Mucm- net: A mamba powered ucm-net for skin lesion segmenta- tion.arXiv preprint arXiv:2405.15925, 2024

    Chunyu Yuan, Dongfang Zhao, and Sos S Agaian. Mucm- net: A mamba powered ucm-net for skin lesion segmenta- tion.arXiv preprint arXiv:2405.15925, 2024. 4, 5

    work page arXiv 2024

  42. [42]

    Vm-unet-v2 rethinking vision mamba unet for medical image segmentation.arXiv preprint arXiv:2403.09157, 2024

    Mingya Zhang, Yue Yu, Limei Gu, Tingsheng Lin, and Xianping Tao. Vm-unet-v2 rethinking vision mamba unet for medical image segmentation.arXiv preprint arXiv:2403.09157, 2024. 2, 5, 6, 7, 3

    work page arXiv 2024

  43. [43]

    Transfuse: Fus- ing transformers and cnns for medical image segmentation

    Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fus- ing transformers and cnns for medical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 14–24. Springer,

  44. [44]

    Unet++: A nested u-net architecture for medical image segmentation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. InDeep learn- ing in medical image analysis and multimodal learning for clinical decision support, pages 3–11. Springer, 2018. 1, 2, 3

    2018

  45. [45]

    Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model.arXiv preprint arXiv:2401.09417, 2024. 1, 2 10 MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation Supplementary Material

    work page internal anchor Pith review arXiv 2024

  46. [46]

    The Intersection over Union (IoU), also known as the Jac- card index, calculates the ratio of the intersection between the predicted and ground truth masks relative to their union

    Evaluation Metrics We evaluate our segmentation performance using overlap- based, boundary-based, and classification-based metrics. The Intersection over Union (IoU), also known as the Jac- card index, calculates the ratio of the intersection between the predicted and ground truth masks relative to their union. The Dice similarity coefficient (DSC), which...

  47. [47]

    These cover boundary-focused met- rics such as HD95, cross-dataset generalization, and tests on non-dermoscopic datasets

    Additional Experiments and Results To complement the main paper, we present additional ex- periments that further broaden the evaluation and justifica- tion of our framework. These cover boundary-focused met- rics such as HD95, cross-dataset generalization, and tests on non-dermoscopic datasets. 7.1. HD95 Evaluation across Four Datasets To evaluate bounda...

  48. [48]

    Additional Ablation Study This section presents additional ablation studies to evaluate the impact of our design decisions further. 2 Model P(M)↓F(G)↓ ISIC2017 ISIC2018 HAM10000 PH2 Ours−Model(Avg.) IoU/DSC/HD95 Cost vs Ours Params×/GFLOPs×IoU↑DSC↑HD95↓ IoU↑DSC↑HD95↓ IoU↑DSC↑HD95↓ IoU↑DSC↑HD95↓ H-vmunet [37] 8.97 0.742 84.22 91.43 12.81 81.78 89.98 14.67 ...

    work page arXiv 1992

  49. [49]

    Table 18 summarizes each module’s de- sign goal, mechanism, nearest prior, and the architectural differences that lead to the expected improvements in lesion segmentation

    Comparative Analysis of Module Designs To further clarify the novelty of our proposed AMF, LGFM, and CGA, we provide a detailed comparison with their clos- est prior designs. Table 18 summarizes each module’s de- sign goal, mechanism, nearest prior, and the architectural differences that lead to the expected improvements in lesion segmentation. 3 Model BU...

    work page arXiv

  50. [50]

    with” denotes the full model con- taining the module, and “w/o

    Module-wise Feature Map Visualization Figure 5 provides a qualitative comparison of representative feature maps with and without the key modules in Mam- baLiteUNet. The top row presents the feature responses from the full model with AMF, LGFM, and CGA, while the bottom row shows the corresponding feature responses af- ter removing each module. This compar...

  51. [51]

    Figure 6 il- lustrates how our proposed MambaLiteUNet progressively transforms feature representations

    Stage-wise Feature Map Visualization This section provides stage-wise qualitative evidence of how MambaLiteUNet processes lesion images throughout its encoder–decoder pipeline, complementing the quantita- tive results presented in the main manuscript. Figure 6 il- lustrates how our proposed MambaLiteUNet progressively transforms feature representations. W...

    work page arXiv 1940

  52. [52]

    The top row shows the input image, followed by activation maps from each en- coder stage (Encoder1–Encoder5) and the bottleneck

    and tested on a held-out image. The top row shows the input image, followed by activation maps from each en- coder stage (Encoder1–Encoder5) and the bottleneck. As depth increases, the model learns progressively more ab- stract and localized features that emphasize lesion bound- aries and suppress background noise. The bottom row (left→right) shows the gr...

  53. [53]

    VM-UNet (0.1718 Sec/Image, 582.5 MB) and VM-UNet2 (0.1836 Sec/Image, 613.7 MB) are the most computationally ex- pensive

    Inference Time and Memory Usage Table 19 presents a comparative analysis of inference time and memory for Mamba-based models. VM-UNet (0.1718 Sec/Image, 582.5 MB) and VM-UNet2 (0.1836 Sec/Image, 613.7 MB) are the most computationally ex- pensive. LightM-UNet is considerably lighter (0.0194 5 Figure 6. Feature map visualization in MambaLiteUNet. Top row (l...