ACM-UNet: Adaptive Integration of CNNs and Mamba for Efficient Medical Image Segmentation

Cheng Chen; Jing Huang; Qiying Lai; Yongkang Zhao; Yuhan Li; Zhitao Dai

REVIEW 4 major objections 5 minor 45 references

ACM-UNet: Adaptive Integration of CNNs and Mamba for Efficient Medical Image Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ACM-UNet claims that pretrained CNN and Mamba backbones can be spliced into a U-Net with lightweight adapters, yielding 85.12% average Dice on Synapse and 92.29% on ACDC at 16.48M parameters.

desk verdict Useful hybrid CNN-Mamba UNet with an honest efficiency profile, but the central SOTA claim is undercut by the paper's own ablation table. read the letter →

arxiv 2505.24481 v1 pith:FJHHPGHI submitted 2025-05-30 cs.CV

Jing Huang , Yongkang Zhao , Yuhan Li , Zhitao Dai , Cheng Chen , Qiying Lai This is my paper

classification cs.CV

keywords medicalimagesegmentationU-NetMambastatespacemodelsconvolutionalneuralnetworkslightweightadapterwavelettransformpretrainedbackbones

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 tries to establish that the structural mismatch blocking reuse of pretrained CNN and state-space backbones in medical segmentation can be resolved by a lightweight adapter, so a simple U-Net can keep fine local detail and long-range context at once. It presents ACM-UNet, in which the first three encoder stages are pretrained ResNet-50 layers, the fourth stage combines a ResNet layer with two pretrained Visual State Space blocks joined by adapters, and each decoder up-sample block includes two multi-scale wavelet transform modules. On the Synapse CT benchmark the reported result is 85.12% average Dice and 13.89 mm HD95; on the ACDC cardiac MRI benchmark it is 92.29% average Dice. These numbers come with 16.48M parameters and 17.93G FLOPs, which the paper holds up as evidence that pretraining-guided hybrids can be both accurate and deployable. If the claim holds, building a new segmenter could start from off-the-shelf backbones instead of training a bespoke hybrid from scratch.

What carries the argument

The load-bearing mechanism is the lightweight adapter, a short chain of linear operations inserted between the ResNet-50 layer and the pretrained visual state-space (VSS) blocks from VMamba; its job is to make the two backbones' feature spaces commensurate so pretrained weights continue to work. The second mechanism is the MSWT module, which uses three parallel wavelet-transform convolutions with kernels of different sizes to separate and recombine high- and low-frequency content, followed by batch norm, ReLU, and a residual connection. These two components carry the argument: the adapter makes reuse possible, and the wavelet module makes the decoder refine rather than merely upsample.

What would settle it

Retrain ACM-UNet on Synapse with the ResNet-50 layers and the VSS blocks randomly initialized instead of ImageNet-pretrained, keeping all other settings identical; if the 85.12% Dice and 13.89 mm HD95 are unchanged, the pretrained-asset-reuse claim is not what carries the result.

Watch

Extended reading notes

Core claim

The central discovery claimed is that pretrained weights survive an architecture change when the splice point is a deliberately minimal adapter. ACM-UNet's encoder keeps the first three ResNet-50 stages for local detail, then feeds a fourth stage made of one ResNet layer, two lightweight adapters, and two pretrained visual state-space (VSS) blocks from VMamba; the adapters align the feature dimensions and semantics so the Mamba branch's global-context representations can act on CNN features. In the decoder, each up-sampling block runs two multi-scale wavelet transform (MSWT) modules, each of which decomposes its input into high- and low-frequency parts, applies 1x1, 3x3, and 5x5 convolutions to those parts in parallel, sums them, and adds the input back through a residual connection. The paper reports that this configuration reaches 85.12% DSC and 13.89 mm HD95 on Synapse and 92.29% average DSC on ACDC, with per-organ gains that the authors attribute to the CNN-Mamba combination plus wavelet-domain refinement.

Load-bearing premise

The whole design rests on the assumption that the ImageNet-trained weights of ResNet-50 and the Mamba VSS blocks remain useful after the lightweight adapters splice them together, even though the paper does not directly measure whether the adapted features stay aligned.

Editorial extensions

If this is right

Pretrained ResNet and VMamba weights can be reused in a U-Net without retraining the backbone, so future segmenters can be assembled from off-the-shelf vision models.
The MSWT module is a plug-in: adding it alone raises Synapse Dice from 83.52% to 84.68% and cuts HD95 from 25.95 mm to 13.80 mm in the paper's ablation.
Two VSS blocks are the reported sweet spot; using four or eight blocks reduces both Dice and HD95 while increasing parameters, so the global branch should stay small.
At 16.48M parameters and 17.93G FLOPs, the model is light enough for the paper to claim it is suitable for real-time or resource-constrained clinical use.
The same architecture works on both CT and MRI benchmarks, which the paper takes as evidence of cross-modality generality.

Reading between the lines

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

Editorial inference: The same two-adapter splice could likely port other pretrained backbones, such as vision transformer blocks, into a U-Net encoder, making the recipe a general plug-and-play pattern rather than a ResNet-VMamba special case.
Editorial inference: Because MSWT separates frequencies, a useful stress test is to add Gaussian noise or blur at inference time beyond the training augmentations; if boundary accuracy degrades sharply, the wavelet refinement is not doing the explanatory work the paper assigns it.
Editorial inference: The ablation varies only 2, 4, and 8 VSS blocks; fitting the full trade-off curve across 1, 2, 3, 4, and 6 blocks on both datasets would show whether the sweet spot is a stable property of the architecture or a dataset-specific tuning choice.
Editorial inference: A direct diagnostic of feature alignment—comparing, say, the cosine similarity or distribution statistics of the CNN and Mamba branch outputs before and after the adapter—would test the core reuse mechanism and could be reported as a quantitative justification for the adapter design.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper does one thing well: it shows a clean way to bolt pretrained ResNet-50 and VMamba VSS blocks into a UNet using lightweight adapters, and it ships code. The efficiency numbers are real and worth attention—16.48M parameters and 17.93G FLOPs is genuinely light for a hybrid model, and the adapter idea is simple enough that practitioners could reuse it. The MSWT module is a minor adaptation of WTConv, but that is not a sin; the paper is honest about the lineage.

The soft spots are real, and one is load-bearing. Table 4 shows the full model with MSWT reaches HD95 of 13.89 mm, while the variant with VSS blocks alone reaches 11.14 mm. The paper explicitly claims MSWT preserves organ boundaries, but the boundary metric moves in the opposite direction. This is not a small inconsistency; it is the central claim of MSWT's contribution being unsupported by the paper's own numbers. On top of that, Table 1 lists MERIT-GCASCADE at 10.38 mm HD95, so ACM-UNet does not achieve the best HD95 on Synapse despite the abstract and Section 4.3.1 saying it does. The DSC improvement is marginal (85.12 vs. 85.00 for MSVM-UNet) and comes with no error bars or significance testing.

The number of VSS blocks is chosen by comparing test-set results, which is a methodological red flag. Hyperparameter selection should not be done on the test set unless the paper says so and accounts for it. The ACDC results are honestly reported as slightly below MSVM-UNet, which is fine, but it undercuts the "outperforms SOTA" framing.

None of this makes the paper worthless. The adapter-based integration is a reasonable engineering contribution, the code availability lowers the bar for others to build on it, and the efficiency analysis is straightforward. The problems are fixable: rerun the MSWT ablation with proper validation-based selection, add uncertainty estimates, and correct the SOTA claims in the abstract and text. If the authors do that, the paper would be a solid contribution to the medical segmentation toolbox.

I would send this to peer review. The architecture is plausible, the code is public, and the core issues are about presentation and experimental discipline rather than a dead end. A good referee can push the authors to make the claims match the evidence. Without those fixes, I would not rely on the SOTA claim, but I would still cite the adapter design if I were building a hybrid encoder.

Referee Report

4 major / 5 minor

Summary. The paper proposes ACM-UNet, a U-shaped encoder–decoder for medical image segmentation that combines a pretrained ResNet-50 and pretrained VMamba VSS blocks through lightweight adapters, together with a multi-scale wavelet transform (MSWT) decoder module. The authors report state-of-the-art results on Synapse (85.12% DSC, 13.89 mm HD95) and competitive results on ACDC (92.29% DSC) with 16.48M parameters and 17.93G FLOPs, and they provide ablation studies for the VSS blocks and MSWT modules. The central claim is that the proposed architecture achieves the best segmentation accuracy and efficiency among the compared methods while reusing large-scale pretrained vision backbones.

Significance. If the reported results and the contribution of each component were firmly established, the paper would offer a useful recipe for integrating pretrained CNNs and Mamba-style models in medical image segmentation with low parameter overhead. The strengths include public code release, experiments on two benchmark datasets, and a concrete efficiency comparison. However, the current manuscript contains internal inconsistencies between the stated claims and the numbers in its own tables, and the ablation evidence does not support the claimed benefit of the MSWT module for boundary preservation. The potential significance is real, but the evidence as presented does not yet justify the state-of-the-art claims.

major comments (4)

[Section 4.3.1, Table 1] The text states that ACM-UNet achieves the best average DSC of 85.12% and HD95 of 13.89 mm on Synapse, but Table 1 itself reports MERIT-GCASCADE with HD95 10.38 mm and MixFormer with HD95 12.67 mm, both lower and therefore better than 13.89 mm. The claim of best HD95 is not supported by the authors' own comparison table; the claim should be restricted to DSC or the discrepancy should be explained.
[Table 4, Section 4.6.1] The ablation table shows that adding MSWT to the VSS-only configuration changes HD95 from 11.14 mm to 13.89 mm, a 2.75 mm degradation, while DSC improves only from 84.86% to 85.12%. The text says the full model achieves its best performance and that MSWT retains details critical to identifying organ margins, but the reported boundary metric moves in the opposite direction. No error bars or repeated-run statistics are provided, so the claimed boundary-preservation benefit of MSWT is not established.
[Section 4.6.2, Table 5] The number of VSS blocks (N=2) is selected by directly comparing test-set DSC and HD95 values on the Synapse test partition (85.12% vs. 84.43% and 84.86%), with no validation split, cross-validation, or significance testing. Tuning a hyperparameter on the test set makes the reported test numbers optimistic and compromises the fairness of the comparison against prior methods in Table 1.
[Section 4.3.2, Table 2; Introduction contributions] The paper claims in the Introduction that ACM-UNet outperforms existing state-of-the-art methods in segmentation accuracy, but Table 2 shows MSVM-UNet achieving a higher average DSC (92.58%) than ACM-UNet (92.29%) on ACDC. The per-method comparison in the body acknowledges this for individual classes, but the broad claim about surpassing state-of-the-art methods is too strong and should be qualified.

minor comments (5)

[Section 4.2.1, Eq. (2)] The value of the loss weight α is never specified; please report the exact value used in the experiments.
[Section 4.2.1] The package name 'calflop' appears to be a typo for 'calflops' or a similarly named library; please correct the spelling.
[Figure 3 caption] The phrase 'organ regions circles by orange rectangles' should read 'organ regions circled by orange rectangles'.
[Section 4.6.2] The text says that with four VSS blocks DSC decreases to 84.43%, and then with eight VSS blocks 'the DSC further decreases to 84.86%'; since 84.86% is larger than 84.43%, the wording should be revised, for example to 'increases relative to the four-block configuration'.
[Section 4.3.1] The reference number for PVT-EMCAD-B2 is given as [42] in the text but as [43] in Table 1; please harmonize the citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported results are external-benchmark measurements, not derived from their own outputs.

full rationale

The paper's central claims are empirical comparisons on the public Synapse and ACDC benchmarks. The method is a composition of pretrained ResNet-50/VSS encoders with adapters and a wavelet decoder defined by Eq. (1); no module or parameter is defined in terms of the reported DSC/HD95, and the reported numbers are measurements on the designated test splits, not outputs of a derivation that encodes the target. The closest issues are methodological, not definitional: Table 5 selects the number of VSS blocks using the same Synapse test metrics later reported as the final result (test-set selection bias), and Table 4's HD95 values (full model 13.89 mm vs. VSS-only 11.14 mm) contradict the claim that MSWT improves boundary preservation. These are correctness/fairness concerns, not reductions of a prediction to its inputs by construction. The citation to MSVM-UNet is used for baseline comparison and weight-decay settings and does not carry the main argument. Thus no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters and untested transfer assumptions. No new physical or conceptual entities are introduced; the adapter and MSWT are engineering components built from existing operations.

free parameters (4)

loss weight alpha = not reported
In Eq. 2, the Dice/CE combination weight alpha is not given, so the exact training objective is under-specified.
number of VSS blocks = 2
Selected in Table 5 based on test performance on Synapse; no validation set was used for this hyperparameter.
weight decay = 1e-3 (Synapse), 1e-4 (ACDC)
Set differently per dataset to reduce overfitting, following MSVM-UNet.
MSWT kernel scales = 1x1, 3x3, 5x5
Chosen by design for multi-scale decomposition, not justified by experiments.

assumptions (3)

domain assumption Pretrained ImageNet weights of ResNet-50 and VMamba transfer to medical imaging data.
The entire training strategy relies on the benefit of large-scale pretraining; no validation is provided for this transfer.
domain assumption The standard splits and preprocessing of Synapse and ACDC are comparable to those used by prior methods.
Performance numbers are compared directly across papers, which assumes identical evaluation protocols.
domain assumption The SS2D selective-scan implementation from VMamba remains correct when embedded in the new encoder.
The paper reuses third-party code without modification and does not test its behavior in the new context.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ACM-UNet: Adaptive Integration of CNNs and Mamba for Efficient Medical Image Segmentation." pith.science (2026). https://pith.science/paper/FJHHPGHI

@misc{pith2026250524481,
  author       = {Pith},
  title        = {Pith review of: ACM-UNet: Adaptive Integration of CNNs and Mamba for Efficient Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FJHHPGHI}},
  note         = {Machine review of arXiv:2505.24481}
}

read the original abstract

The U-shaped encoder-decoder architecture with skip connections has become a prevailing paradigm in medical image segmentation due to its simplicity and effectiveness. While many recent works aim to improve this framework by designing more powerful encoders and decoders, employing advanced convolutional neural networks (CNNs) for local feature extraction, Transformers or state space models (SSMs) such as Mamba for global context modeling, or hybrid combinations of both, these methods often struggle to fully utilize pretrained vision backbones (e.g., ResNet, ViT, VMamba) due to structural mismatches. To bridge this gap, we introduce ACM-UNet, a general-purpose segmentation framework that retains a simple UNet-like design while effectively incorporating pretrained CNNs and Mamba models through a lightweight adapter mechanism. This adapter resolves architectural incompatibilities and enables the model to harness the complementary strengths of CNNs and SSMs-namely, fine-grained local detail extraction and long-range dependency modeling. Additionally, we propose a hierarchical multi-scale wavelet transform module in the decoder to enhance feature fusion and reconstruction fidelity. Extensive experiments on the Synapse and ACDC benchmarks demonstrate that ACM-UNet achieves state-of-the-art performance while remaining computationally efficient. Notably, it reaches 85.12% Dice Score and 13.89mm HD95 on the Synapse dataset with 17.93G FLOPs, showcasing its effectiveness and scalability. Code is available at: https://github.com/zyklcode/ACM-UNet.

Figures

Figures reproduced from arXiv: 2505.24481 by the authors.

**Figure 1.** The overall architecture of our proposed ACM-UNet. Left: (a) illustrates the structure of individual VSS block in its encoder, where LN and DWConv [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** The scan expanding and scan merging process of medical images via 2D-selective-scan strategy. (a) presents the detail of scan expanding in our SS2D [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Visual comparison of different methods on the Synapse multi-organ dataset. The first column represents the ground truth, and the following columns represent the segmentation predictions of ACM-UNet, MSVM-UNet, Swin-UNet, and TransUNet, respectively. It can be observed that our ACM-UNet is of superiority in those organ regions circles by orange rectangles. 17.93G) and number of parameters (16.48M). In contrast, given… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 23 canonical work pages

[1]

Ronneberger, P

O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Springer, 2015, pp. 234–241

2015
[2]

Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, J. Liang, Unet ++: A nested u-net architecture for medical image segmentation, in: Deep learn- ing in medical image analysis and multimodal learning for clinical deci- sion support: 4th international workshop, DLMIA 2018, and 8th interna- tional workshop, ML-CDS 2018, held in conjunction with MICCAI 2018, Gra...

work page 2018
[3]

J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, Y . Zhou, Transunet: Transformers make strong encoders for medical im- age segmentation, ArXiv abs/2102.04306 (2021)

arXiv 2021
[4]

H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, M. Wang, Swin- unet: Unet-like pure transformer for medical image segmentation, in: Eu- ropean conference on computer vision, Springer, 2022, pp. 205–218

work page 2022
[5]

J. Ruan, S. Xiang, Vm-unet: Vision mamba unet for medical image seg- mentation, ArXiv abs/2402.02491 (2024)

arXiv 2024
[6]

C. Chen, L. Yu, S. Min, S. Wang, Msvm-unet: Multi-scale vision mamba unet for medical image segmentation, 2024 IEEE International Confer- ence on Bioinformatics and Biomedicine (BIBM) (2024) 3111–3114

work page 2024
[7]

F. Shen, H. Ye, J. Zhang, C. Wang, X. Han, W. Yang, Advancing pose- guided image synthesis with progressive conditional di ffusion models, arXiv preprint arXiv:2310.06313 (2023)

arXiv 2023
[8]

F. Shen, J. Tang, Imagpose: A unified conditional framework for pose- guided person generation, Advances in neural information processing sys- tems 37 (2024) 6246–6266

work page 2024

Show all 45 references

[9]

F. Shen, X. Jiang, X. He, H. Ye, C. Wang, X. Du, Z. Li, J. Tang, Imagdressing-v1: Customizable virtual dressing, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 39, 2025, pp. 6795– 6804

2025
[10]

F. Shen, J. Yu, C. Wang, X. Jiang, X. Du, J. Tang, Imaggarment-1: Fine- grained garment generation for controllable fashion design, arXiv preprint arXiv:2504.13176 (2025)

2025 arXiv
[11]

F. Shen, C. Wang, J. Gao, Q. Guo, J. Dang, J. Tang, T.-S. Chua, Long- term talkingface generation via motion-prior conditional diffusion model, arXiv preprint arXiv:2502.09533 (2025)

2025 arXiv
[12]

L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille, Seman- tic image segmentation with deep convolutional nets and fully connected crfs, arXiv preprint arXiv:1412.7062 (2014)

2014 arXiv
[13]

L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille, Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE transactions on pat- tern analysis and machine intelligence 40 (4) (2017) 834–848

2017
[14]

L.-C. Chen, G. Papandreou, F. Schro ff, H. Adam, Rethinking atrous convolution for semantic image segmentation, arXiv preprint arXiv:1706.05587 (2017)

2017 arXiv
[15]

L.-C. Chen, Y . Zhu, G. Papandreou, F. Schro ff, H. Adam, Encoder- decoder with atrous separable convolution for semantic image segmen- tation, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 801–818

2018
[16]

J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, Y . Wei, Deformable con- volutional networks, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 764–773

2017
[17]

A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H. Adam, Mobilenets: E fficient convolutional neural net- works for mobile vision applications, ArXiv abs/1704.04861 (2017)

2017 arXiv
[18]

B. Yang, G. Bender, Q. V . Le, J. Ngiam, Condconv: Conditionally param- eterized convolutions for efficient inference, Advances in neural informa- tion processing systems 32 (2019)

2019
[19]

Y . Chen, X. Dai, M. Liu, D. Chen, L. Yuan, Z. Liu, Dynamic convolu- tion: Attention over convolution kernels, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 11027–11036

2019
[20]

M. Tan, Q. V . Le, Mixconv: Mixed depthwise convolutional kernels, arXiv preprint arXiv:1907.09595 (2019)

2019 arXiv
[21]

J. Yang, S. Liu, J. Wu, X. Su, N. Hai, X. Huang, Pinwheel-shaped convo- lution and scale-based dynamic loss for infrared small target detection, Proceedings of the AAAI Conference on Artificial Intelligence 39 (9) (2025) 9202–9210

2025
[22]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems 30 (2017)

2017
[23]

A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, ArXiv abs/2312.00752 (2023)

2023 arXiv
[24]

Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE /CVF international conference on computer vi- sion, 2021, pp. 10012–10022

2021
[25]

Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, Y . Liu, Vmamba: Visual state space model, Advances in neural information pro- cessing systems 37 (2024) 103031–103063

2024
[26]

Zhang, H

Y . Zhang, H. Liu, Q. Hu, Transfuse: Fusing transformers and cnns for medical image segmentation, ArXiv abs/2102.08005 (2021)

2021 arXiv
[27]

Xu, Hc-mamba: Vision mamba with hybrid convolutional techniques for medical image segmentation, ArXiv abs/2405.05007 (2024)

J. Xu, Hc-mamba: Vision mamba with hybrid convolutional techniques for medical image segmentation, ArXiv abs/2405.05007 (2024)

2024 arXiv
[28]

K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recog- nition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

2016
[29]

Dosovitskiy, L

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An im- age is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

2020 arXiv
[30]

A. Gu, K. Goel, C. R’e, E fficiently modeling long sequences with struc- tured state spaces, ArXiv abs/2111.00396 (2021)

2021 arXiv
[31]

A. Gu, A. Gupta, K. Goel, C. R ´e, On the parameterization and initializa- tion of diagonal state space models, ArXiv abs/2206.11893 (2022)

2022 arXiv
[32]

L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, X. Wang, Vision mamba: Efficient visual representation learning with bidirectional state space model, ArXiv abs/2401.09417 (2024)

2024 arXiv
[33]

S. E. Finder, R. Amoyal, E. Treister, O. Freifeld, Wavelet convolutions 9 for large receptive fields, in: European Conference on Computer Vision, Springer, 2024, pp. 363–380

2024
[34]

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conference on computer vision and pattern recognition, Ieee, 2009, pp. 248–255

2009
[35]

Landman, Z

B. Landman, Z. Xu, J. Igelsias, M. Styner, T. Langerak, A. Klein, Mic- cai multi-atlas labeling beyond the cranial vault–workshop and challenge, in: Proc. MICCAI multi-atlas labeling beyond cranial vault—workshop challenge, V ol. 5, Munich, Germany, 2015, p. 12

2015
[36]

Bernard, A

O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P.-A. Heng, I. Cetin, K. Lekadir, O. Camara, M. A. G. Ballester, G. Sanrom ´a, S. Napel, S. E. Petersen, G. Tziritas, E. Grinias, M. Khened, V . A. Kollerathu, G. Krishnamurthi, M.-M. Roh ´e, X. Pennec, M. Sermesant, F...

2018
[37]

M. M. Rahman, R. Marculescu, Medical image segmentation via cas- caded attention decoding, in: Proceedings of the IEEE /CVF winter con- ference on applications of computer vision, 2023, pp. 6222–6231

2023
[38]

G. Xu, X. Zhang, X. He, X. Wu, Levit-unet: Make faster encoders with transformer for medical image segmentation, in: Chinese Conference on Pattern Recognition and Computer Vision (PRCV), Springer, 2023, pp. 42–53

2023
[39]

R. Azad, R. Arimond, E. K. Aghdam, A. Kazerouni, D. Merhof, Dae- former: Dual attention-guided e fficient transformer for medical image segmentation, in: International workshop on predictive intelligence in medicine, Springer, 2023, pp. 83–95

2023
[40]

Huang, S

H. Huang, S. Xie, L. Lin, Y . Iwamoto, X. Han, Y .-W. Chen, R. Tong, Scaleformer: revisiting the transformer-based backbones from a scale-wise perspective for medical image segmentation, arXiv preprint arXiv:2207.14552 (2022)

2022 arXiv
[41]

R. Azad, L. Niggemeier, M. Huttemann, A. Kazerouni, E. K. Aghdam, Y . Velichko, U. Bagci, D. Merhof, Beyond self-attention: Deformable large kernel attention for medical image segmentation, 2024 IEEE /CVF Winter Conference on Applications of Computer Vision (W ACV) (2023) 1276–1286

2023
[42]

M. M. Rahman, R. Marculescu, G-cascade: E fficient cascaded graph con- volutional decoding for 2d medical image segmentation, 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV) (2023) 7713–7722

2023
[43]

W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, L. Shao, Pvt v2: Improved baselines with pyramid vision transformer, Computational visual media 8 (3) (2022) 415–424

2022
[44]

J. Liu, H. Yang, H.-Y . Zhou, Y . Xi, L. Yu, C. Li, Y . Liang, G. Shi, Y . Yu, S. Zhang, et al., Swin-umamba: Mamba-based unet with imagenet-based pretraining, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, 2024, pp. 615–625

2024
[45]

J. Liu, K. Li, C. Huang, H. Dong, Y . Song, R. Li, Mixformer: A mixed cnn–transformer backbone for medical image segmentation, IEEE Trans- actions on Instrumentation and Measurement 74 (2025) 1–20. 10

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Ronneberger, P

O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Springer, 2015, pp. 234–241

2015

[2] [2]

Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, J. Liang, Unet ++: A nested u-net architecture for medical image segmentation, in: Deep learn- ing in medical image analysis and multimodal learning for clinical deci- sion support: 4th international workshop, DLMIA 2018, and 8th interna- tional workshop, ML-CDS 2018, held in conjunction with MICCAI 2018, Gra...

work page 2018

[3] [3]

J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, Y . Zhou, Transunet: Transformers make strong encoders for medical im- age segmentation, ArXiv abs/2102.04306 (2021)

arXiv 2021

[4] [4]

H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, M. Wang, Swin- unet: Unet-like pure transformer for medical image segmentation, in: Eu- ropean conference on computer vision, Springer, 2022, pp. 205–218

work page 2022

[5] [5]

J. Ruan, S. Xiang, Vm-unet: Vision mamba unet for medical image seg- mentation, ArXiv abs/2402.02491 (2024)

arXiv 2024

[6] [6]

C. Chen, L. Yu, S. Min, S. Wang, Msvm-unet: Multi-scale vision mamba unet for medical image segmentation, 2024 IEEE International Confer- ence on Bioinformatics and Biomedicine (BIBM) (2024) 3111–3114

work page 2024

[7] [7]

F. Shen, H. Ye, J. Zhang, C. Wang, X. Han, W. Yang, Advancing pose- guided image synthesis with progressive conditional di ffusion models, arXiv preprint arXiv:2310.06313 (2023)

arXiv 2023

[8] [8]

F. Shen, J. Tang, Imagpose: A unified conditional framework for pose- guided person generation, Advances in neural information processing sys- tems 37 (2024) 6246–6266

work page 2024