Pith. sign in

REVIEW 4 major objections 7 minor 72 references

MambaVesselNet++: A Hybrid CNN-Mamba Architecture for Medical Image Segmentation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MambaVesselNet++ claims that a U-Net with convolutions throughout and vision Mamba only at the bottleneck beats CNN, transformer, and pure Mamba baselines on six medical segmentation datasets.

desk verdict Plausible incremental architecture with broad benchmarks, but the central outperformance claim is undercut by internal table-text inconsistencies and missing 2D baseline protocol details. read the letter →

arxiv 2507.19931 v1 pith:TCFDII42 submitted 2025-07-26 cs.CV

Qing Xu , Yanming Chen , Yue Li , Ziyu Liu , Zhenye Lou , Yixuan Zhang , Xiangjian He This is my paper
classification cs.CV
keywords medicalimagesegmentationhybridCNN-MambaselectivestatespacemodelMambaU-Netsemanticinstancecerebrovascular
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

This paper claims that a U-shaped network that keeps convolution in the encoder and decoder, and inserts a vision Mamba block only at the bottleneck, outperforms pure CNN, transformer, and pure Mamba segmentation models on six public medical datasets covering dermoscopy, polyp, retinal vessel, ultrasound, histopathology nuclei, and 3D MRA cerebrovascular images. The motivation is that convolutions preserve the local texture needed for fine boundaries, while Mamba's selective state space mechanism supplies long-range context with linear rather than quadratic cost. The authors present the model as an extension of their earlier 3D cerebrovascular MambaVesselNet, adding adaptive 2D/3D convolutions and a multi-branch decoder that switches between semantic and instance segmentation. For the 3D MRA results, the reported Dice, training time, and memory use improve on the compared CNN, transformer, and Mamba baselines under the stated protocol. If the comparisons hold up, the design offers a middle path for high-resolution medical volumes where transformers are too costly and pure CNNs miss distant vessel continuity.

What carries the argument

The load-bearing object is the Hi-Encoder and BF-Decoder pair. The Hi-Encoder alternates texture-aware layers — residual blocks made of two convolutions ($3\times3$ and $2\times2$ stride 2) with layer normalization and LeakyReLU — with vision Mamba layers built from the VSS block of VMamba, whose SS2D mechanism scans feature tokens in four directions and processes them through the S6 selective state space block. The BF-Decoder upsamples the global feature with $2\times2$ deconvolutions, adds element-wise the local features from the last texture-aware layer, then concatenates earlier encoder features through skip connections and refines them with $3\times3$ convolutions. What carries the argument is the placement: Mamba appears only at the bottleneck, so long-range dependencies are modeled at linear cost while convolutions throughout the network preserve local texture; adaptive convolution switches the same design between 2D and 3D inputs, and a three-head decoder (nuclei prediction, horizontal-vertical distance, nuclei type) switches the same backbone to instance segmentation.

What would settle it

Recompute the reported metrics from shared prediction masks: if mIoU is computed per image for the foreground class as Equation (14) suggests, U-Mamba's PH2 mIoU of 0.936 cannot exceed its Dice of 0.904, so that row alone signals an evaluation inconsistency. Also check Table 4, where the text gives the highest SQ to MambaVesselNet++ but the table assigns 0.746 to CellViT and 0.729 to MambaVesselNet++; rerunning all models in one harness would settle which margins survive.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that MambaVesselNet++ achieves the best accuracy among current convolution-, transformer-, and Mamba-based segmentation models on every task it tests: skin lesion (PH2), polyp (CVC-ClinicDB), retinal vessel (DRIVE), breast lesion (UDIAT), nuclei instance (TNBC), and cerebrovascular volume (IXI). The mechanism is a hybrid encoder in which texture-aware convolutional layers extract local features, then vision Mamba blocks model global dependencies with linear complexity, and a bifocal fusion decoder combines both streams through skip connections. On the reported numbers the model reaches Dice 0.953 on PH2, 0.711 on DRIVE, 0.849 on UDIAT, 0.870 on IXI, and AJI 0.534 on TNBC, while using 2.1 minutes per epoch and 6.9 GB memory on the 3D task. For nuclei instance segmentation the paper claims the best overall performance, though its own Table 4 lists CellViT ahead on PQ (0.481 vs 0.479) and SQ (0.746 vs 0.729).

Load-bearing premise

The load-bearing premise is that every baseline model was trained and evaluated under the same protocol with identically computed metrics; the paper states this explicitly only for the 3D comparisons that use public implementations, so an unverified mismatch in the 2D or instance evaluations would collapse the claimed superiority.

Editorial extensions

If this is right

  • If the reported comparisons hold, one hybrid backbone covers 2D semantic, 3D volumetric, and instance segmentation without an architectural change, only different decoding heads.
  • The reported 3D cost of 2.1 minutes per epoch and 6.9 GB memory is lower than every compared 3D baseline, so the bottleneck-Mamba design is a viable path to high-resolution volume segmentation on limited hardware.
  • The domain-transfer tables imply the same model generalizes to unseen datasets from the same anatomy — PH2 to ISIC2018, CVC-ClinicDB to ColonDB, DRIVE to STARE, UDIAT to BUSI — better than the listed baselines.
  • The largest relative gains in the tables appear on the thin-structure retinal vessel task, where local texture and global continuity both matter; this is consistent with the paper's over- versus under-segmentation argument.

Reading between the lines

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

  • Inference: the bottleneck-Mamba placement should transfer to other thin-structure segmentation tasks such as airway or coronary artery segmentation, which the paper does not test.
  • Inference: the reported margins suggest a testable hypothesis that the hybrid gain grows as target structures become thinner and more spatially dispersed; comparing DRIVE with the solid-lesion UDIAT task on the same protocol would test it.
  • Inference: the paper's interpretability claims about the selective parameters $B$ and $C$ are qualitative; a direct extension would be to correlate high-variance selective parameters with clinician-flagged uncertainty regions.
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.

Referee Report

4 major / 7 minor

Summary. This paper proposes MambaVesselNet++, a hybrid CNN-Mamba encoder-decoder for medical image segmentation. The Hi-Encoder stacks texture-aware convolutional layers for local feature extraction and vision Mamba blocks at the bottleneck for global dependencies; the BF-Decoder fuses multi-scale encoder features through skip connections. The model is adapted to 2D/3D inputs and to semantic or instance segmentation via a multi-branch decoding head with task-specific losses. Experiments compare with CNN-, transformer-, and Mamba-based baselines on PH2, CVC-ClinicDB, DRIVE, UDIAT, TNBC, and IXI, and include domain-generalization results on ISIC2018, ColonDB, STARE, and BUSI. The abstract and conclusion claim state-of-the-art performance across all these tasks.

Significance. If the reported results were reliable, the contribution would be practically useful: the architecture is simple, builds on established components, and shows a favorable accuracy/efficiency trade-off, particularly on 3D cerebrovascular segmentation where MambaVesselNet++ reports the best Dice (0.870) with lower training time and memory than the baselines. The MONAI-based 3D comparison and the ablation study are strengths. However, the central comparative claim is currently under-supported because of internal inconsistencies in Tables 2 and 4 and the absence of a documented protocol for the 2D and instance-segmentation baselines. The significance is therefore conditional on correcting the evidence.

major comments (4)
  1. [§4.4, Table 4] The text states that MambaVesselNet++ 'achieves the highest value of 0.746' for SQ and a PQ of 0.481 that is 'slightly higher' than the alternatives, but Table 4 reports SQ = 0.729 for MambaVesselNet++ and SQ = 0.746 for CellViT, and PQ = 0.479 for MambaVesselNet++ with CellViT also at 0.481. These numbers contradict the claim that the proposed method outperforms all baselines on the TNBC instance-segmentation task, which is one of the paper's central claims. Please correct the table/text and re-state the conclusion accordingly.
  2. [§4.4, Table 2] Table 2 reports U-Mamba on PH2 with Dice = 0.904 and mIoU = 0.936. For a binary foreground/background segmentation, the per-image Dice is always at least the per-image IoU (Dice = 2IoU/(1+IoU) ≥ IoU), so the reported mean mIoU cannot exceed the reported mean Dice under the stated metric definitions. This indicates either a metric-implementation error, a mislabeled column, or a transcription error. Because the table is used to support the superiority claim, the authors must clarify the exact averaging and report corrected values.
  3. [§4.4, Tables 2–4 and §4.6, Tables 7–8] The 3D comparison in Table 5 is accompanied by a protocol statement ('All models are trained and tested under the same data augmentation settings, and the public implementations ... from MONAI are used'), but no equivalent statement is given for the 2D semantic, instance-segmentation, or domain-generalization comparisons. Since the central claim is that MambaVesselNet++ outperforms all baselines, the authors need to report, for each baseline and dataset: preprocessing, input resolution, training schedule, loss functions, post-processing, metric computation code, and whether results come from re-training or from published papers. Without this information the reported margins (e.g., DRIVE Dice 0.711 vs 0.614 for TransUNet) cannot be attributed to the architecture.
  4. [§4.4, Table 3] On UDIAT, MambaVesselNet++ has HD = 23.01, which is higher (worse) than TransUNet (18.68) and Swin-UMamba (18.83), despite the text's statement that the model 'also demonstrates lower HD values across most tasks'. The claim should be qualified, or the HD computation should be checked, because the current wording overstates the result on this dataset.
minor comments (7)
  1. [§4.4] The sentence introducing the nuclei instance-segmentation comparison says 'as shown in Table 3', but the results appear in Table 4; please correct the cross-reference.
  2. [Table 4] The citation keys in Table 4 are inconsistent with the bibliography: Hover-Net, StarDIST, CellPose, CPP-Net, and CellViT are cited as [28], [30], [23], [37], and [59], whereas the text and reference list identify them as [18], [48], [50], [9], and [26]; please fix the keys and also standardize the model names (e.g., HoVer-Net, StarDist, Cellpose).
  3. [§5] The interpretability and uncertainty-quantification claims about selective parameters B and C are speculative and not supported by any experiment or visualization in the paper; please either remove them or add quantitative evidence.
  4. [Front matter and References] The manuscript retains ACM template artifacts: the header says 'Trovato et al.', the ACM Reference Format line gives 2018, and the reference for the IXI cerebrovascular dataset [19] cites a PET study rather than the TOF-MRA dataset used; please correct these.
  5. [Fig. 6] Figure 6's caption calls the comparison '2D qualitative comparison ... on the IXI dataset'; since IXI is a 3D MRA volume, the caption should say that 2D slices from the 3D volume are shown.
  6. [§3.3 and Abstract] The decoder is referred to as both 'Bi-Decoder' (abstract, Eq. (7)) and 'BF-Decoder' (body, Section 3.3); please use one name consistently.
  7. [Abstract/Code availability] The code link points to https://github.com/CC0117/MambaVesselNet; please clarify whether this repository contains the MambaVesselNet++ implementation or only the earlier version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: results are external benchmark measurements; self-citations are contextual, not load-bearing.

full rationale

This is an empirical engineering paper rather than a derivation. The architecture is assembled from published, externally sourced components (Mamba/SS2D from refs [20,35], U-Net-style decoder design, MONAI baselines), and its results are direct benchmark measurements on public datasets (PH2, CVC-ClinicDB, DRIVE, UDIAT, TNBC, IXI). The paper's own equations (1)-(18) define standard SSM dynamics, losses, and metrics; no equation is used both as an input and as the claimed conclusion. The authors cite their prior MambaVesselNet [11] and other self-authored work [36], but these citations are contextual precursors for the extended framework, not the evidence for the novel claims; the new contributions are validated by external comparisons and ablations. Assertions such as "MambaVesselNet++ outperforms existing state-of-the-art methods" are empirical claims whose support is the reported tables. The internal inconsistencies in Table 4 (e.g., text assigning SQ 0.746 to MambaVesselNet++ while the table credits CellViT) and the mIoU > Dice anomaly on PH2 are metric-consistency and correctness concerns, not circular reasoning. Therefore no circular step meeting the quoted-reduction standard is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper is an empirical engineering contribution. It introduces no new theoretical entities. The central claim depends on standard Mamba/CNN machinery, hand-chosen hyperparameters, and fair baseline comparisons. The ledger is small because the paper's contribution is architectural and empirical rather than derivational.

free parameters (5)
  • Loss weights (lambda_Dice, lambda_CE) = 1, 1
    Section 4.3 sets both to 1 by default; no sensitivity analysis.
  • Instance loss weights (lambda_NP, lambda_HV, lambda_NT) = 1, 5, 1
    Section 4.3; lambda_HV=5 is hand-selected following prior work, no ablation.
  • Number of Mamba blocks n = 16
    Architecture description in Section 3.2 and Fig. 2; chosen without ablation.
  • 3D patch size = 64x64x64
    Section 4.3; no ablation on patch size.
  • Base channel width C_k = 48 * 2^k
    Section 3.3; design choice, not justified by experiments.
assumptions (4)
  • standard math Mamba selective state space model (Gu and Dao [20]) provides correct long-range dependency modeling with linear complexity.
    Section 3.1 relies on the published Mamba formulation.
  • domain assumption SS2D cross-scan from VMamba [35] is applied to 3D volumes without requiring a new 3D scanning mechanism.
    Section 3.2 and Fig. 3 describe 2D scanning; the paper does not specify how 3D patches are scanned, yet 3D experiments are reported.
  • domain assumption Ground-truth annotations in PH2, CVC-ClinicDB, DRIVE, UDIAT, TNBC, and IXI are correct.
    Datasets are used as training and test targets without annotation quality checks.
  • domain assumption Baseline models were trained and evaluated fairly and at their best configuration.
    Section 4.4 states this only for the 3D comparison with MONAI implementations; for 2D tasks this premise is implicit and unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MambaVesselNet++: A Hybrid CNN-Mamba Architecture for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/TCFDII42

@misc{pith2026250719931,
  author       = {Pith},
  title        = {Pith review of: MambaVesselNet++: A Hybrid CNN-Mamba Architecture for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TCFDII42}},
  note         = {Machine review of arXiv:2507.19931}
}
read the original abstract

Medical image segmentation plays an important role in computer-aided diagnosis. Traditional convolution-based U-shape segmentation architectures are usually limited by the local receptive field. Existing vision transformers have been widely applied to diverse medical segmentation frameworks due to their superior capabilities of capturing global contexts. Despite the advantage, the real-world application of vision transformers is challenged by their non-linear self-attention mechanism, requiring huge computational costs. To address this issue, the selective state space model (SSM) Mamba has gained recognition for its adeptness in modeling long-range dependencies in sequential data, particularly noted for its efficient memory costs. In this paper, we propose MambaVesselNet++, a Hybrid CNN-Mamba framework for medical image segmentation. Our MambaVesselNet++ is comprised of a hybrid image encoder (Hi-Encoder) and a bifocal fusion decoder (BF-Decoder). In Hi-Encoder, we first devise the texture-aware layer to capture low-level semantic features by leveraging convolutions. Then, we utilize Mamba to effectively model long-range dependencies with linear complexity. The Bi-Decoder adopts skip connections to combine local and global information of the Hi-Encoder for the accurate generation of segmentation masks. Extensive experiments demonstrate that MambaVesselNet++ outperforms current convolution-based, transformer-based, and Mamba-based state-of-the-arts across diverse medical 2D, 3D, and instance segmentation tasks. The code is available at https://github.com/CC0117/MambaVesselNet.

Figures

Figures reproduced from arXiv: 2507.19931 by the authors.

Figure 1
Figure 1. Illustration of local features versus global dependencies in medical vessel segmentation. Blue boxes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of MambaVesselNet++ architecture. The model takes 2D medical images or 3D patches as the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Detailed illustration of the 2D-Selective-Scan (SS2D) mechanism. The SS2D mechanism processes [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: 2D medical semantic comparison of different baselines on PH2, CVC-ClinicDB, DRIVE and UDIAT [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: 2D nuclei instance comparison of different baselines on the TNBC dataset. Our MambaVesselNet++ [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: 2D qualitative comparison of different baselines on the IXI dataset. The first image shows the ground [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: This figure shows a 3D qualitative comparison of segmentation results on the IXI dataset. The first image [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 58 canonical work pages

  1. [1]

    Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. 2020. Dataset of breast ultrasound images. Data in brief28 (2020), 104863

    work page 2020

  2. [2]

    Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. 2024. Medical image segmentation review: The success of u-net.IEEE Transactions on Pattern Analysis and Machine Intelligence(2024)

    work page 2024

  3. [3]

    Subhashis Banerjee, Fredrik Nysjö, Dimitrios Toumpanakis, Ashis Kumar Dhara, Johan Wikström, and Robin Strand

    work page

  4. [4]

    Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilariño

    work page

  5. [5]

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. 2022. Swin-unet: Unet-like pure transformer for medical image segmentation. InEuropean conference on computer vision. Springer, 205–218

    2022

  6. [6]

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou

    work page

  7. [7]

    Jieneng Chen, Jieru Mei, Xianhang Li, Yongyi Lu, Qihang Yu, Qingyue Wei, Xiangde Luo, Yutong Xie, Ehsan Adeli, Yan Wang, et al. 2024. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers.Medical Image Analysis97 (2024), 103280

    work page 2024

  8. [8]

    Li Chen, Yanjun Xie, Jie Sun, Niranjan Balu, Mahmud Mossa-Basha, Kristi Pimentel, Thomas S Hatsukami, Jenq-Neng Hwang, and Chun Yuan. 2017. 3D intracranial artery segmentation using a convolutional autoencoder. In2017 IEEE international conference on bioinformatics and biomedicine (BIBM). IEEE, 714–717

    work page 2017

Show all 72 references
  1. [9]

    Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. 2023. CPP-net: Context-aware polygon proposal network for nucleus segmentation.IEEE Transactions on Image Processing32 (2023), 980–994

    2023

  2. [10]

    Ying Chen, Darui Jin, Bin Guo, and Xiangzhi Bai. 2022. Attention-assisted adversarial model for cerebrovascular segmentation in 3D TOF-MRA volumes.IEEE Transactions on Medical Imaging41, 12 (2022), 3520–3532

    2022

  3. [11]

    Yanming Chen, Ziyu Liu, and Xiangjian He. 2024. MambaVesselNet: A Hybrid CNN-Mamba Architecture for 3D Cerebrovascular Segmentation. InProceedings of the 6th ACM International Conference on Multimedia in Asia. 1–7

    2024

  4. [12]

    Zhen Chen, Xiaoqing Guo, Peter YM Woo, and Yixuan Yuan. 2021. Super-resolution enhanced medical image diagnosis with sample affinity interaction.IEEE Transactions on Medical Imaging40, 5 (2021), 1377–1389

    2021

  5. [13]

    Zhen Chen, Qing Xu, Xinyu Liu, and Yixuan Yuan. 2024. Un-sam: Universal prompt-free segmentation for generalized nuclei images.arXiv preprint arXiv:2402.16663(2024)

    2024 arXiv

  6. [14]

    Zhen Chen, Qing Xu, Xinyu Liu, and Yixuan Yuan. 2025. UN-SAM: Domain-adaptive self-prompt segmentation for universal nuclei images.Medical Image Analysis(2025), 103607

    2025

  7. [15]

    Zhen Chen, Chen Yang, Meilu Zhu, Zhe Peng, and Yixuan Yuan. 2022. Personalized retrogress-resilient federated learning toward imbalanced medical data.IEEE Transactions on Medical Imaging41, 12 (2022), 3663–3674

    2022

  8. [16]

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. 2019. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin im...

    2019 arXiv

  9. [17]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint ...

    2020 arXiv

  10. [18]

    Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot

  11. [19]

    Katherine R Gray, Robin Wolz, Rolf A Heckemann, Paul Aljabar, Alexander Hammers, Daniel Rueckert, Alzheimer’s Disease Neuroimaging Initiative, et al. 2012. Multi-region analysis of longitudinal FDG-PET for the classification of Alzheimer’s disease.NeuroImage60, 1 (2012), 221–229

    2012

  12. [20]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752(2023)

    2023 arXiv

  13. [21]

    Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. 2021. Combining recurrent, convolutional, and continuous-time models with linear state space layers.Advances in neural information processing systems34 (2021), 572–585

    2021

  14. [22]

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. 2021. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. InInternational MICCAI Brainlesion Workshop. J. ACM, Vol. 37, No. 4, Article 111. Publication da...

    2021

  15. [23]

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. 2022. Unetr: Transformers for 3d medical image segmentation. InProceedings of the IEEE/CVF winter conference on applications of computer vision. 574–584

    2022

  16. [24]

    Along He, Kai Wang, Tao Li, Chengkun Du, Shuang Xia, and Huazhu Fu. 2023. H2former: An efficient hierarchical hybrid transformer for medical image segmentation.IEEE Transactions on Medical Imaging42, 9 (2023), 2763–2775

    2023

  17. [25]

    Adam Hoover and Michael Goldbaum. 2003. Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels.IEEE transactions on medical imaging22, 8 (2003), 951–958

    2003

  18. [26]

    Fabian Hörst, Moritz Rempe, Lukas Heine, Constantin Seibold, Julius Keyl, Giulia Baldini, Selma Ugurel, Jens Siveke, Barbara Grünwald, Jan Egger, et al. 2024. Cellvit: Vision transformers for precise cell segmentation and classification. Medical Image Analysis94 (2024), 103143

    2024

  19. [27]

    Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. 2019. Local relation networks for image recognition. InProceedings of the IEEE/CVF international conference on computer vision. 3464–3473

    2019

  20. [28]

    Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. 2020. Unet 3+: A full-scale connected unet for medical image segmentation. InICASSP 2020-2020 IEEE international conference on acoustics, speech and signa...

    2020

  21. [29]

    Nabil Ibtehaz and Daisuke Kihara. 2023. Acc-unet: A completely convolutional unet model for the 2020s. InInternational conference on medical image computing and computer-assisted intervention. Springer, 692–702

    2023

  22. [30]

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. 2021. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods18, 2 (2021), 203–211

    2021

  23. [31]

    Bo Li, Yong Zhang, Yunhan Ren, Chengyang Zhang, and Baocai Yin. 2024. Lite-UNet: A lightweight and efficient network for cell localization.Engineering Applications of Artificial Intelligence129 (2024), 107634

    2024

  24. [32]

    Chenxin Li, Wuyang Li, Hengyu Liu, Xinyu Liu, Qing Xu, Zhen Chen, Yue Huang, and Yixuan Yuan. 2024. Flaws can be applause: Unleashing potential of segmenting ambiguous objects in SAM.Advances in Neural Information Processing Systems37 (2024), 45578–45599

    2024

  25. [33]

    Ailiang Lin, Bingzhi Chen, Jiayu Xu, Zheng Zhang, Guangming Lu, and David Zhang. 2022. Ds-transunet: Dual swin transformer u-net for medical image segmentation.IEEE Transactions on Instrumentation and Measurement71 (2022), 1–15

    2022

  26. [34]

    Jiarun Liu, Hao Yang, Hong-Yu Zhou, Yan Xi, Lequan Yu, Cheng Li, Yong Liang, Guangming Shi, Yizhou Yu, Shaoting Zhang, et al. 2024. Swin-umamba: Mamba-based unet with imagenet-based pretraining. InInternational Conference on Medical Image Computing and Computer-Assisted Interv...

    2024

  27. [35]

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. 2024. VMamba: Visual State Space Model. arXiv:2401.10166 [cs.CV] https://arxiv.org/abs/2401.10166

    2024 arXiv

  28. [36]

    Ziyu Liu, Jiaxuan Li, Xiangjian He, Qing Xu, Xin Chen, and Shoujun Zhou. 2025. Swin-VasMamba: A Topologically Constrained Model For 3D Vascular Segmentation. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    2025

  29. [37]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international conference on computer vision. 10012–10022

    2021

  30. [38]

    Zhenye Lou, Qing Xu, Zekun Jiang, Xiangjian He, Zhen Chen, Yi Wang, Chenxin Li, Maggie M He, and Wenting Duan

  31. [39]

    Zhenye Lou, Qing Xu, Zekun Jiang, Xiangjian He, Chenxin Li, Zhen Chen, Yi Wang, Maggie M He, and Wenting Duan

  32. [40]

    Yuxiang Luo, Jinglei Feng, Guangwu Qian, Wenting Duan, et al. [n. d.]. Med-fastsam: Improving transfer efficiency of sam to domain-generalised medical image segmentation. ([n. d.])

  33. [42]

    Jun Ma, Feifei Li, and Bo Wang. 2024. U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmenta- tion. arXiv:2401.04722 [eess.IV]

    2024 arXiv

  34. [43]

    arXiv preprint arXiv:2408.11787(2024)

    Nusegdg: Integration of heterogeneous space and gaussian kernel for domain-generalized nuclei segmentation. arXiv preprint arXiv:2408.11787(2024)

    2024 arXiv

  35. [44]

    Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. 2018. Segmentation of nuclei in histopathology images by deep regression of the distance map.IEEE transactions on medical imaging38, 2 (2018), 448–459

    2018

  36. [45]

    Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu. 2024. Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and J. ACM, Vol. 37, No. 4, Article 111. Publication date: ...

    2024

  37. [46]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, ...

    2015

  38. [47]

    Jiacheng Ruan and Suncheng Xiang. 2024. Vm-unet: Vision mamba unet for medical image segmentation.arXiv preprint arXiv:2402.02491(2024)

    2024 arXiv

  39. [48]

    Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. 2018. Cell detection with star-convex polygons. InMedical image computing and computer assisted intervention–MICCAI 2018: 21st international conference, Granada, Spain, September 16-20, 2018, proceedings, part II 1...

    2018

  40. [49]

    Teresa Mendonça, Pedro M Ferreira, Jorge S Marques, André RS Marcal, and Jorge Rozeira. 2013. PH 2-A dermoscopic image database for research and benchmarking. InEMBC. IEEE, 5437–5440

    2013

  41. [50]

    Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. 2021. Cellpose: a generalist algorithm for cellular segmentation.Nature methods18, 1 (2021), 100–106

    2021

  42. [51]

    Giles Tetteh, Velizar Efremov, Nils D Forkert, Matthias Schneider, Jan Kirschke, Bruno Weber, Claus Zimmer, Marie Piraud, and Björn H Menze. 2020. Deepvesselnet: Vessel segmentation, centerline prediction, and bifurcation detection in 3-d angiographic volumes.Frontiers in Neur...

    2020

  43. [52]

    Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. 2018. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Scientific data5, 1 (2018), 1–9

    2018

  44. [53]

    Jeya Maria Jose Valanarasu and Vishal M Patel. 2022. Unext: Mlp-based rapid medical image segmentation network. In International conference on medical image computing and computer-assisted intervention. Springer, 23–33

    2022

  45. [54]

    David Vázquez, Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Antonio M López, Adriana Romero, Michal Drozdzal, and Aaron Courville. 2017. A benchmark for endoluminal scene segmentation of colonoscopy images. Journal of healthcare engineering2017, 1 (2017), 4037190

    2017

  46. [55]

    Joes Staal, Michael D Abràmoff, Meindert Niemeijer, Max A Viergever, and Bram Van Ginneken. 2004. Ridge-based vessel segmentation in color images of the retina.IEEE transactions on medical imaging23, 4 (2004), 501–509

    2004

  47. [56]

    Ziyang Wang, Jian-Qing Zheng, Yichi Zhang, Ge Cui, and Lei Li. 2024. Mamba-unet: Unet-like pure visual mamba for medical image segmentation.arXiv preprint arXiv:2402.05079(2024)

    2024 arXiv

  48. [57]

    Zhitao Xiao, Bowen Liu, Lei Geng, Fang Zhang, and Yanbei Liu. 2020. Segmentation of lung nodules using improved 3D-UNet neural network.Symmetry12, 11 (2020), 1787

    2020

  49. [58]

    Yutong Xie, Jianpeng Zhang, Chunhua Shen, and Yong Xia. 2021. Cotr: Efficiently bridging cnn and transformer for 3d medical image segmentation. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, Septembe...

    2021

  50. [59]

    Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. 2024. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 578–588

    2024

  51. [60]

    Qing Xu and Wenting Duan. 2021. An automatic nuclei image segmentation based on multi-scale split-attention U-Net. InMICCAI Workshop on Computational Pathology. PMLR, 236–245

    2021

  52. [61]

    Yaoqin Wang, Wenting Xie, Chenxin Li, Qing Xu, Zhongshi Du, Zhaoming Zhong, and Lina Tang. 2025. Automated microvascular invasion prediction of hepatocellular carcinoma via deep relation reasoning from dynamic contrast- enhanced ultrasound.Computerized Medical Imaging and Grap...

    2025

  53. [62]

    Qing Xu, Zhenye Lou, Chenxin Li, Xiangjian He, Rong Qu, Tesema Fiseha Berhanu, Yi Wang, Wenting Duan, and Zhen Chen. 2025. HRMedSeg: Unlocking High-resolution Medical Image segmentation via Memory-efficient Attention Modeling.arXiv preprint arXiv:2504.06205(2025)

    2025 arXiv

  54. [63]

    Qing Xu, Yuxiang Luo, Wenting Duan, and Zhen Chen. 2025. Co-Seg++: Mutual Prompt-Guided Collaborative Learning for Versatile Medical Segmentation.arXiv preprint arXiv:2506.17159(2025)

    2025

  55. [64]

    Qing Xu, Zhicheng Ma, HE Na, and Wenting Duan. 2023. DCSAU-Net: A deeper and more compact split-attention U-Net for medical image segmentation.Computers in Biology and Medicine154 (2023), 106626

    2023

  56. [65]

    Moi Hoon Yap, Gerard Pons, Joan Marti, Sergi Ganau, Melcior Sentis, Reyer Zwiggelaar, Adrian K Davison, and Robert Marti. 2017. Automated breast ultrasound lesions detection using convolutional neural networks.IEEE J. Biomed. Health Inform.22, 4 (2017), 1218–1226

    2017

  57. [66]

    Shaofeng Yuan and Feng Yang. 2023. Segmentation of Aortic Vessel Tree in CT Scans with Deep Fully Convolutional Networks.arXiv preprint arXiv:2305.09833(2023)

    2023 arXiv

  58. [67]

    Qing Xu, Jiaxuan Li, Xiangjian He, Ziyu Liu, Zhen Chen, Wenting Duan, Chenxin Li, Maggie M He, Fiseha B Tesema, Wooi P Cheah, et al. 2024. Esp-medsam: Efficient self-prompting sam for universal domain-generalized medical image segmentation.arXiv preprint arXiv:2407.14153(2024)

    2024 arXiv

  59. [73]

    Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. 2021. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. InProceedings of the IEEE/CVF conferenc...

    2021

  60. [2015]

    saliency maps from physicians

    WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Comput. Med. Imaging Graph.43 (2015), 99–111

    2015

  61. [2019]

    Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images.Medical image analysis58 (2019), 101563

    2019

  62. [2021]

    Transunet: Transformers make strong encoders for medical image segmentation.arXiv preprint arXiv:2102.04306 (2021)

    2021 arXiv

  63. [2024]

    Streamlining neuroradiology workflow with AI for improved cerebrovascular structure monitoring.Scientific Reports14, 1 (2024), 9245

    2024

  64. [2025]

    Knowledge-Based Systems(2025), 113641

    NuSegDG: Integration of heterogeneous space and Gaussian kernel for domain-generalized nuclei segmentation. Knowledge-Based Systems(2025), 113641

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.