Pith. sign in

REVIEW 3 major objections 5 minor 21 references

DM-SegNet: Dual-Mamba Architecture for 3D Medical Image Segmentation with Global Context Modeling

T0 review · 3 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 A dual-Mamba architecture for 3D medical segmentation claims state-of-the-art Dice on Synapse and BraTS2023.

desk verdict Plausible Mamba-based segmentation architecture, but the SOTA claim rests on a nonstandard random split and single-run results; needs official benchmarks and error bars before it can be taken at face value. read the letter →

arxiv 2506.05297 v1 pith:6MMIC2Z7 submitted 2025-06-05 eess.IV cs.CV

Hangyu Ji This is my paper
classification eess.IVcs.CV
keywords 3DmedicalimagesegmentationStateSpaceModelsMambaquadri-directionalspatialscanninggatedconvolutionmulti-scalefusiondecoderSynapsedatasetBraTS2023
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 dual-Mamba architecture can resolve the encoder-decoder incompatibility that limits prior state-space-model segmenters for 3D medical images. It reports state-of-the-art Dice similarity coefficients of 85.44% on the Synapse abdominal CT dataset and 90.22% on the BraTS2023 brain tumor MRI dataset. The proposed design scans each volume in four directions to preserve anatomical continuity, applies gated spatial convolution to restore local detail, and uses a Mamba-based decoder to propagate state information across scales. A sympathetic reader would care because these results, if correct, show that linear-time sequence models can match or beat transformer and CNN baselines on clinically relevant volumetric segmentation.

What carries the argument

The load-bearing mechanism is the quadri-directional spatial Mamba (QSM) module: instead of flattening a 3D volume into one 1D sequence, it runs four selective-scan passes over the volume—forward, reverse, inter-slice, and reverse inter-slice—under the Mamba state-space update, then concatenates and gated-fuses their outputs. This preserves anatomical coherence along the axial, sagittal, and coronal axes. Two supporting components carry the rest of the argument: the gated spatial convolution (GSC), which uses a gating branch of stacked 3×3×3 convolutions to add local spatial detail before state modeling, and the multi-scale fusion Mamba decoder (MFMamba), which fuses features from multiple encoder scales and synchronizes encoder and decoder states across resolutions.

What would settle it

Run DM-SegNet and the five baselines on the official Synapse challenge test split and the official BraTS2023 validation split using the paper's training recipe; if DM-SegNet no longer beats SegMamba (or SwinUNETR) by the reported margins, or if published leaderboard numbers already exceed these DSC scores, the state-of-the-art claim fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the two structural weaknesses of Mamba-based medical segmentation—1D flattening destroying spatial topology and decoders discarding learned state—can be fixed together. The quadri-directional spatial Mamba module runs four selective scans (forward, reverse, inter-slice, and reverse inter-slice) and fuses their outputs, so the state-space model retains anatomical continuity while still capturing global context. Gated spatial convolution layers placed before each Mamba block re-inject local spatial cues through a gating branch. The multi-scale fusion Mamba decoder then aligns encoder states with decoder feature maps bidirectionally, preventing context dissipation during upsampling. On the paper's evaluation, this produces the top DSC and lowest HD95 among the compared methods on both Synapse and BraTS2023, with the ablation attributing a 1.73-point Dice gain and a 43.7% HD95 reduction to the new components.

Load-bearing premise

The reported state-of-the-art claim rests on the assumption that a randomly allocated 70/10/20 training/validation/test split of each dataset is comparable to the official Synapse and BraTS2023 benchmark splits, so the Dice and Hausdorff margins reflect genuine architectural gains rather than an easier test distribution.

Editorial extensions

If this is right

  • If the reported gains hold, Mamba-based models become a credible substitute for transformers in 3D segmentation, keeping linear-time scaling while matching or beating UNETR and SwinUNETR.
  • The four-directional scanning strategy is a general remedy for the topology-loss problem in any 3D SSM, not only this architecture.
  • The paper's ablation isolates GSC and QSM as complementary: together they add 1.73 DSC points and cut HD95 by 43.7%, so both components pull weight.
  • A Mamba decoder that synchronizes states with the encoder offers a template for other encoder-decoder tasks where upsampling traditionally discards sequence-modeling information.

Reading between the lines

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

  • The paper reports a single run per configuration; a natural follow-up, not provided in the paper, would be repeated-seed evaluation to see whether the DSC margins over SegMamba survive run-to-run variance.
  • Because the evaluation uses a 70/10/20 random split rather than the official challenge splits, a direct extension would be rerunning the same comparison on the official Synapse and BraTS2023 test sets to see if the state-of-the-art claim transfers.
  • The 'bidirectional state synchronization' in the decoder is described functionally; a testable extension would be measuring how much of the encoder's hidden state survives into the decoder outputs, e.g., via state-alignment probes.
  • The quadri-directional scanning could plausibly transfer to other volumetric sequence tasks such as video prediction or 3D object reconstruction, though the paper does not test those settings.
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

3 major / 5 minor

Summary. The paper proposes DM-SegNet, a 3D medical image segmentation architecture built around a dual-Mamba design. The encoder combines gated spatial convolutions with a quadri-directional spatial Mamba scanning module, and the decoder uses a Mamba-based multi-scale fusion module. The authors evaluate on the Synapse multi-organ CT dataset and the BraTS2023 brain tumor MRI dataset, reporting a Dice Similarity Coefficient of 85.44% on Synapse and 90.22% on BraTS2023, which they claim is state-of-the-art. They also provide a small ablation study on Synapse showing incremental gains from the proposed components.

Significance. If the reported results were obtained under the standard benchmark protocols, DM-SegNet would be a useful addition to the growing family of state-space-model-based 3D segmentation networks, with a plausible efficiency argument over transformers and a technically interesting decoder design. The paper gives credit for describing the architecture in enough detail to be reimplemented and for retraining all comparison methods under an internally consistent protocol, which makes the relative ordering of methods in Tables 1-3 meaningful as an internal comparison. The ablation study (Table 3) is directionally consistent: each added component improves DSC and reduces HD95. However, the central state-of-the-art claim is not presently supported because the evaluation uses a nonstandard random split that is not comparable to published benchmark numbers, and all results are single-run point estimates without error bars or significance tests. The paper releases no code, which further limits reproducibility.

major comments (3)
  1. [Section 4.2, Section 4.3, Tables 1-2] The evaluation protocol is not comparable to the published benchmarks the paper claims to beat. Section 4.2 states that for each dataset the authors 'randomly allocated 70% of the 3D data for training, 10% for validation, and the remaining 20% for testing.' For Synapse's 30 volumes this is a 21/3/6 split, whereas the established Synapse benchmark convention is a fixed split (typically 18 training and 12 test volumes) used by the cited baselines and published state-of-the-art reports. For BraTS2023, the challenge defines fixed training/validation/test partitions, and a random split changes the case mix. Consequently, the claimed margins in Table 1 (e.g., 85.44% versus 85.09% for SegMamba) and the BraTS2023 numbers in Table 2 are not comparable to the cited baseline results or to published leaderboard numbers. Section 4.3 notes that all models were retrained under the same split, which makes the internal comparison fair, but internal fairness does not establish a state-of-the-art claim. The authors should either report results on the official benchmark splits or clearly reframe the claims as results under a new, nonstandard protocol and avoid direct comparison with published numbers.
  2. [Tables 1-3] All tables report a single run with no error bars, confidence intervals, or significance tests. This is especially problematic for the Synapse result, where the test set has only 6 volumes: the 0.35-point DSC advantage over SegMamba in Table 1 is within plausible seed-to-seed variability, and the per-organ differences are even smaller. The same concern applies to the BraTS2023 results in Table 2 and to the ablation study in Table 3. Without multiple random seeds or a paired statistical test, the reported margins do not support the claim of consistent superiority. The authors should report mean and standard deviation over at least three independent runs, or otherwise quantify the variability.
  3. [Section 3.2, Eq. (4)] The definition of the quadri-directional spatial Mamba module is internally inconsistent. In the text the authors describe 'channel-wise concatenation and gated fusion' of the four directional paths, but Eq. (4) defines QSMamba(z) as the simple sum Mamba(zf) + Mamba(zr) + Mamba(zi) + Mamba(zri). Concatenation and summation produce different feature representations, and the gated fusion step mentioned in the text does not appear in the equation. This discrepancy must be resolved for the architecture description to be reproducible: either the equation should include concatenation followed by a fusion projection, or the text should describe additive combination.
minor comments (5)
  1. [Table 1] The 'AG' column for SegResNet lists 0.00 DSC. A completely zero Dice for the adrenal gland is implausible unless the class was missing from all test predictions; please verify this entry and, if it is a typo or an artifact of the random split, correct it or explain it. The column header 'HD' should also be 'HD95' for consistency with the text.
  2. [Section 3.2, Eq. (1)] Eq. (1) uses the same weight symbol W_g^1 for both sequential convolutions in the gated branch. If the two convolution layers have different weights, the notation should distinguish them (e.g., W_g^1 and W_g^2); otherwise, the paper appears to share weights unintentionally.
  3. [Section 4.3] The paper states that all compared models were retrained with 'author-official implementations under identical training protocols,' but it does not list the exact hyperparameters used for the baselines. Please provide a table or supplement with the training configuration for each baseline to support the reproducibility claim.
  4. [Limitations and Future Work] The Limitations section candidly acknowledges computational time and memory constraints, which is appropriate. It does not, however, mention the main limitation identified above: that the reported results were obtained under a self-defined split and are therefore not directly comparable to published benchmark numbers. Please add this caveat explicitly.
  5. [Abstract and Introduction] The phrase 'state-of-the-art DSC' appears in the Abstract and Introduction before the benchmark protocol is defined. Until the official-split evaluation is provided, the claims should be qualified as 'under our evaluation protocol' or 'compared to methods retrained on our split.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the architecture modules are defined constructively and the reported DSC/HD95 numbers are external benchmark evaluations, not outputs of a fitted equation or a self-citation chain.

full rationale

DM-SegNet does not contain a circular step in the sense targeted by this analysis. The method sections define the GSC, QSMamba, and Mamba decoder modules as explicit compositions of convolutions and Mamba scans (e.g., Eq. 4 defines QSMamba as a sum of four directional Mamba processes), but this is a constructive architectural definition, not a derivation of the reported accuracy from the definition. The load-bearing claims are empirical: Table 1 and Table 2 report Dice and HD95 on Synapse and BraTS2023, and Table 3 reports an ablation where modules are removed and re-added. None of these quantities is a fitted parameter renamed as a prediction, and no equation in the paper maps an input directly to the target metric by construction. The one substantive concern is that Section 4.2 states a nonstandard random 70/10/20 split and Section 4.3 claims all baselines were retrained under this split, which weakens comparability with published state-of-the-art numbers but does not constitute circularity. The paper also contains no load-bearing self-citation: the reference list is entirely external prior work, and no uniqueness theorem or prior result by this author is invoked to force a choice. Therefore, while the SOTA claim carries correctness and protocol risk, the derivation chain is not circular.

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

The headline empirical claims rest on hand-chosen training and evaluation choices (data split, crop size, epochs, learning rate) and on the correctness and fairness of the baseline retraining. The paper introduces no new physical entities; the QSMamba and MFMamba components are software modules. The listed axioms are domain assumptions about the benchmark protocol and the sufficiency of the architectural description.

free parameters (3)
  • Training/validation/test split = 70% train / 10% validation / 20% test (random)
    Selected in Section 4.2; every reported DSC and HD95 depends on this split, which differs from official benchmark splits.
  • Training hyperparameters = Crop 96x96x96 or 128x128x128; 1000 or 2000 epochs; batch size 1; SGD lr 1e-4; poly decay 1e-5
    Chosen in Section 4.2 without sensitivity analysis; benchmark scores are tied to them.
  • Network width and depth = not reported
    The number of QSMamba blocks and feature channels per stage is not stated, so the model capacity behind the scores is an undisclosed choice.
assumptions (3)
  • domain assumption A random 70/10/20 split of Synapse and BraTS2023 is a valid substitute for the official benchmark splits.
    Section 4.2 defines this split and Tables 1-2 compare against published methods on that basis. If the split differs from official protocols, the headline SOTA comparison is invalid.
  • domain assumption All baselines were retrained under identical protocols, so the reported margins are due to architecture rather than training detail.
    Stated in Section 4.3, but no seeds, training logs, or code are provided to verify this claim.
  • domain assumption Equations (1)-(4) and Figures 2-3 contain enough detail to implement the QSMamba encoder and MFMamba decoder.
    Implementation details such as number of blocks, channel widths, and the exact four-direction scanning mechanism are unspecified; reproducibility assumes the text suffices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DM-SegNet: Dual-Mamba Architecture for 3D Medical Image Segmentation with Global Context Modeling." pith.science (2026). https://pith.science/paper/6MMIC2Z7

@misc{pith2026250605297,
  author       = {Pith},
  title        = {Pith review of: DM-SegNet: Dual-Mamba Architecture for 3D Medical Image Segmentation with Global Context Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MMIC2Z7}},
  note         = {Machine review of arXiv:2506.05297}
}
read the original abstract

Accurate 3D medical image segmentation demands architectures capable of reconciling global context modeling with spatial topology preservation. While State Space Models (SSMs) like Mamba show potential for sequence modeling, existing medical SSMs suffer from encoder-decoder incompatibility: the encoder's 1D sequence flattening compromises spatial structures, while conventional decoders fail to leverage Mamba's state propagation. We present DM-SegNet, a Dual-Mamba architecture integrating directional state transitions with anatomy-aware hierarchical decoding. The core innovations include a quadri-directional spatial Mamba module employing four-directional 3D scanning to maintain anatomical spatial coherence, a gated spatial convolution layer that enhances spatially sensitive feature representation prior to state modeling, and a Mamba-driven decoding framework enabling bidirectional state synchronization across scales. Extensive evaluation on two clinically significant benchmarks demonstrates the efficacy of DM-SegNet: achieving state-of-the-art Dice Similarity Coefficient (DSC) of 85.44% on the Synapse dataset for abdominal organ segmentation and 90.22% on the BraTS2023 dataset for brain tumor segmentation.

Figures

Figures reproduced from arXiv: 2506.05297 by the authors.

Figure 1
Figure 1. Overview of DM-SegNet architecture. The encoder architecture comprises a stem convolutional block [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. QSMamba Encoder structure diagram. It consists of a Gated Spatial Convolution module, multiple Quadri [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The Mamba decoder architecture with Mamba-based Multi-Scale Feature Fusion and Extraction. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visual comparison of the segmentation of Synapse multi-organ CT. The red boxes highlight the key details. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of the segmentation of brain tumor segmentation. White boxes indicate the focus position [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 4 canonical work pages

  1. [1]

    Niyas, S

    S. Niyas, S. J. Pawan, M. Anand Kumar, J. Rajan, Medical image segmentation with 3d convolutional neural networks: A survey, Neurocomputing 493 (2022) 397–413. doi:10.1016/j.neucom.2022.04.065

    work page doi:10.1016/j.neucom.2022.04.065 2022

  2. [2]

    Litjens, T

    G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, C. I. Sánchez, A survey on deep learning in medical image analysis, Medical Image Analysis 42 (2017) 60–88. doi:10.1016/j.media.2017.07.005

    doi:10.1016/j.media.2017.07.005 2017

  3. [3]

    Shorten, T

    C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of Big Data 6 (1) (2019) 60. doi:10.1186/s40537-019-0197-0

    doi:10.1186/s40537-019-0197-0 2019

  4. [4]

    Çiçek, A

    Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, O. Ronneberger, 3d u-net: Learning dense volumetric segmen- tation from sparse annotation, in: S. Ourselin, L. Joskowicz, M. R. Sabuncu, G. Unal, W. Wells (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016, Springer International Publishing, Cham, 2016, pp. 424–432. doi:10.100...

    doi:10.1007/978-3-319-46723-8_49 2016

  5. [5]

    Hatamizadeh, Y

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Landman, H. Roth, D. Xu, Unetr: Transformers for 3d medical image segmentation (Oct. 2021). arXiv:2103.10504, doi:10.48550/arXiv.2103.10504

    arXiv doi:10.48550/arxiv.2103.10504 2021

  6. [6]

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, M. Wang, Swin-unet: Unet-like pure transformer for medical image segmentation (May 2021). arXiv:2105.05537, doi:10.48550/arXiv.2105.05537

    arXiv doi:10.48550/arxiv.2105.05537 2021

  7. [7]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need (Aug. 2023). arXiv:1706.03762, doi:10.48550/arXiv.1706.03762. 8 A PREPRINT - AUGUST 6, 2025

    arXiv doi:10.48550/arxiv.1706.03762 2023

  8. [8]

    A. Gu, K. Goel, C. Ré, Efficiently modeling long sequences with structured state spaces (Aug. 2022). arXiv: 2111.00396, doi:10.48550/arXiv.2111.00396

    arXiv doi:10.48550/arxiv.2111.00396 2022

Show all 21 references
  1. [9]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces (May 2024). arXiv: 2312.00752, doi:10.48550/arXiv.2312.00752

    2024 doi arXiv

  2. [10]

    J. Ma, F. Li, B. Wang, U-mamba: Enhancing long-range dependency for biomedical image segmentation (Jan. 2024). arXiv:2401.04722, doi:10.48550/arXiv.2401.04722

    2024 doi arXiv

  3. [11]

    A. A. Taha, A. Hanbury, Metrics for evaluating 3d medical image segmentation: Analysis, selection, and tool, BMC Medical Imaging 15 (1) (2015) 29. doi:10.1186/s12880-015-0068-x

    2015 doi

  4. [12]

    H. H. Lee, S. Bao, Y . Huo, B. A. Landman, 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation (Mar. 2023). arXiv:2209.15076, doi:10.48550/arXiv.2209. 15076

    2023 doi arXiv

  5. [13]

    Z. Xing, T. Ye, Y . Yang, G. Liu, L. Zhu, Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation (Sep. 2024). arXiv:2401.13560, doi:10.48550/arXiv.2401.13560

    2024 doi arXiv

  6. [14]

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, Y . Liu, Vmamba: Visual state space model (Dec. 2024). arXiv:2401.10166, doi:10.48550/arXiv.2401.10166

    2024 doi arXiv

  7. [15]

    C. Yang, Z. Chen, M. Espinosa, L. Ericsson, Z. Wang, J. Liu, E. J. Crowley, Plainmamba: Improving non- hierarchical mamba in visual recognition (Aug. 2024). arXiv:2403.17695, doi:10.48550/arXiv.2403. 17695

    2024 doi arXiv

  8. [16]

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, X. Wang, Vision mamba: Efficient visual representation learning with bidirectional state space model, in: Proceedings of the 41st International Conference on Machine Learning, PMLR, 2024, pp. 62429–62442

    2024

  9. [17]

    Landman, Z

    B. Landman, Z. Xu, J. Igelsias, M. Styner, T. Langerak, A. Klein, Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge, in: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, V ol. 5, 2015, p. 12

    2015

  10. [18]

    A. F. Kazerooni, N. Khalili, X. Liu, D. Haldar, Z. Jiang, S. M. Anwar, J. Albrecht, M. Adewole, U. Anazodo, H. Anderson, S. Bagheri, U. Baid, T. Bergquist, A. J. Borja, E. Calabrese, V . Chung, G.-M. Conte, F. Dako, J. Eddy, I. Ezhov, A. Familiar, K. Farahani, S. Haldar, J. E....

    2023 doi arXiv

  11. [19]

    B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, Y . Burren, N. Porz, J. Slotboom, R. Wiest, L. Lanczi, E. Gerstner, M.-A. Weber, T. Arbel, B. B. Avants, N. Ayache, P. Buendia, D. L. Collins, N. Cordier, J. J. Corso, A. Criminisi, T. Das, H. Delinget...

    2015

  12. [20]

    Myronenko, 3d mri brain tumor segmentation using autoencoder regularization (Nov

    A. Myronenko, 3d mri brain tumor segmentation using autoencoder regularization (Nov. 2018). arXiv:1810. 11654, doi:10.48550/arXiv.1810.11654

    2018 doi arXiv

  13. [21]

    Perera, P

    S. Perera, P. Navard, A. Yilmaz, Segformer3d: An efficient transformer for 3d medical image segmentation (Apr. 2024). arXiv:2404.10156, doi:10.48550/arXiv.2404.10156. 9

    2024 doi arXiv

Pith tools

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