Pith. sign in

REVIEW 4 major objections 6 minor 47 references

ASM-UNet: Adaptive Scan Mamba Integrating Group Commonalities and Individual Variations for Fine-Grained Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ASM-UNet claims that letting a Mamba-based U-Net choose its scan order per patient—via a group-plus-individual adaptive score—improves fine-grained medical image segmentation, reporting the best average Dice on ACDC (92.61), Synapse (87.27)

desk verdict The BTMS dataset is a real contribution; the adaptive-scan mechanism is plausible but underspecified in the paper, and the empirical margins are thin. read the letter →

arxiv 2508.07237 v1 pith:IZWFPJIA submitted 2025-08-10 cs.CV

Bo Wang , Mengyuan Xu , Yue Yan , Yuqun Yang , Kechen Shu , Wei Ping , Xu Tang , Wei Jiang
show 1 more author
Zheng You
This is my paper
classification cs.CV
keywords fine-grainedsegmentationmedicalimageMambastatespacemodeladaptivescanningU-Netbiliarytractanatomicalvariation
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 is trying to establish that fixed, hand-designed scan orders are the main bottleneck for Mamba-based segmentation of fine-grained anatomical structures, and that letting the network pick the scan order per patient removes that bottleneck. It proposes ASM-UNet, a U-Net whose adaptive scanning Mamba block computes a per-image adaptive scan score by adding a shared group scan score to an individual scan score, then uses that score to guide the order in which flattened feature tokens pass through stacked Mamba layers. On ACDC, Synapse, and the new eight-class BTMS biliary-tract dataset, the method reports the highest average Dice among the compared models. If the claim holds, content-aware scan ordering is a low-cost way to gain roughly 0.5–2 Dice points on fine-grained tasks without changing the loss or backbone. The paper also introduces BTMS as a benchmark that exposes how much accuracy is lost when moving from coarse organ detection to fine-grained structure delineation.

What carries the argument

The central object is the adaptive scan score, a per-spatial-location scalar formed as the sum of a group scan score and an individual scan score. The group scan score is a learnable embedding, resized to the spatial dimensions and passed through a sigmoid so its values lie in [0, 1]; it encodes common anatomical layout. The individual scan score is produced by a generator that first scans the patient's features in three fixed directions, mirroring how a clinician builds an initial global impression before looking at small structures. According to the paper, sorting the flattened tokens by the combined adaptive score determines the Mamba scanning order; the score is also concatenated with th

What would settle it

Inspect the released code path that consumes the adaptive scan score: if no permutation or indexing operation reorders the token sequence before the Mamba layers, then the reported Dice gains are not evidence for adaptive scan order. Alternatively, an experiment that replaces the learned ordering with random permutations, while keeping all other components, would show whether reordering itself—rather than extra parameters or the concatenated score—accounts for the improvement.

Watch

Extended reading notes

Core claim

The central claim is that the scanning order of a Mamba layer is not a neutral implementation detail: for small, variable anatomical structures, order determines what context the model can assemble, and a fixed horizontal, vertical, or diagonal scan cannot adapt to patients whose ducts join in different patterns. ASM-UNet accordingly replaces the fixed scan with an adaptive scan score at each spatial location. The score is the sum of a group scan score, a learnable embedding normalized by sigmoid that encodes the common layout of the anatomy, and an individual scan score, generated from the patient's own features after three fixed initial scans. Tokens are then reordered according to that sc

Load-bearing premise

The load-bearing premise, asserted in Section IV-B but not demonstrated there, is that sorting the adaptive scan score actually sets the token order fed to Mamba: the text says the score is sorted to determine the order and also says it is concatenated with the feature embedding, but no equation, pseudocode, or diagram shows the conversion from scalar score to permutation.

Editorial extensions

If this is right

  • If the adaptive scan score is doing the work, any Mamba-based segmentation model with fixed scan paths could adopt the same dual-score mechanism and should see similar fine-grained gains; the paper positions the mechanism as a replacement for fixed scans.
  • On the BTMS dataset, the gap between coarse (76.74) and fine (42.57) Dice shows that distinguishing the eight biliary structures, not just detecting the organ, is where existing methods fail; the paper's reported gains are concentrated there.
  • Ablations show the individual scan score and group scan score contribute additively on the hard categories (40.85 to 41.34 to 41.87 to 42.57), so both commonality and variation are needed for the best fine-grained result.
  • The hardest classes, CD and RHD, remain below 35% Dice for all methods, so the paper's claim is not that adaptive scanning solves fine-grained segmentation, only that it improves it.

Reading between the lines

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

  • Editorial extension: if the sorting mechanism is real, the learned scan order should be inspectable: visualizing which voxels are processed early versus late in different patients could show whether the order tracks known biliary variants without any extra labeling.
  • Editorial extension: the paper compares whole models, so the reported gains could in principle come from other ASM-UNet design details; mounting the same adaptive-score module on an existing fixed-scan Mamba backbone would isolate the contribution of adaptivity.
  • Editorial extension: because the only explicit operation involving the score is concatenation, an ablation that forces a true permutation (sort by score, feed reordered tokens) versus one that only concatenates the score would settle whether scan-path adaptivity or feature augmentation causes the Dice gains.
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 / 6 minor

Summary. The paper proposes ASM-UNet, a Mamba-based U-Net for 3D medical image segmentation, with the stated goal of improving fine-grained segmentation. The core novelty is an 'adaptive scan score' computed as the sum of a learnable group scan embedding and an image-specific individual scan score; the score is claimed to dynamically determine the token ordering processed by the Mamba layers. The method is evaluated on ACDC, Synapse, and a new BTMS dataset containing eight biliary-tract classes, reporting the highest average Dice on all three datasets. The paper also contributes the BTMS dataset and a GitHub link for code and data.

Significance. If the adaptive-scan mechanism works as described, it addresses a real limitation of Mamba-based segmenters, which typically use fixed scanning paths. The BTMS dataset is potentially a useful benchmark for fine-grained abdominal anatomy. Strengths of the paper include standard held-out evaluation, per-class Dice tables, an ablation study, and a parameter analysis; no definitional circularity is apparent. However, the mechanism that turns the scan score into a permutation is never specified, the key numerical comparisons lack variance estimates, and hyperparameters appear to be selected on test results. These are serious, fixable gaps.

major comments (4)
  1. [IV-B, Fig. 5, Table V] The central claim is that the adaptive scan score dynamically determines the Mamba scan order. Section IV-B says the scan order is determined by sorting the scores, but the next sentence says the score is concatenated with the feature embedding. No equation, pseudocode, or diagram shows how a scalar score per token is converted into a permutation of the W x H x D sequence; Fig. 5 is titled 'feature reordering' but shows no reordering operation. The ablations in Table V toggle IS/GS only and never remove the sorting while keeping the concatenation, so the gains could come from the added parameters or concatenated features rather than from adaptive scanning. Please specify the exact sorting/inverse operation and add an ablation that isolates the permutation from the score features.
  2. [V-C, Tables III-IV] All Dice values are from a single run. Several headline margins are under one point (0.55 on ACDC, 0.70 on Synapse, 0.78 on BTMS 'w/ Hard'). Without standard deviations, confidence intervals, or paired significance tests, the claimed superiority over nnFormer and U-Mamba is not established. Report mean +/- std over at least three seeds and, ideally, paired tests or bootstrap intervals.
  3. [V-E, Table VI, IV-B.1] The default number of Mamba branches appears to be selected on the test set: Table VI reports test Dice for 0-4 branches and the text chooses three branches because it gives the best hard-category score, while four branches are better on non-hard. No validation-based selection is described, and the group scan embedding length n_g is a hyperparameter with no sensitivity analysis. If test data was used for model selection, the reported performance is optimistic. Use a validation split for choosing hyperparameters and report test performance only for the final configuration.
  4. [III, Table II] BTMS is introduced as a new clinical benchmark, but the paper does not report ethics approval, acquisition details, annotation protocol or inter-observer agreement, inclusion/exclusion criteria, or whether the 4:3:3 split is at case level. These details are necessary for a dataset contribution to be reproducible and clinically credible.
minor comments (6)
  1. [II-B] References [24] and [37] both call their method Swin-UMamba; clarify the relationship or use distinct labels.
  2. [Table I] CB is used for gallbladder in Table I, but the rest of the paper and Table II use GB; unify the abbreviation.
  3. [IV-A] Six encoder blocks are described, but only five downsampling rates and five channel counts are listed; correct the count.
  4. [Table VI] 0 branches yields the same numbers as M1 in Table V, implying no Mamba at all; clarify what branch count 0 means.
  5. [IV-B.2] The individual scan score is said to use three fixed scan orders [21]; [21] is the Mamba paper and does not define those orders. Specify the three orders or give a proper reference.
  6. [Fig. 5] The labels in Figure 5 are too small to read; enlarge them.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional circularity; adaptive-scan mechanism is under-specified, not circular.

full rationale

ASM-UNet's central claim—adaptive scan scores, formed from group and individual scores, determine the Mamba scan order—is not circular by construction. The group scan score is a learnable embedding, the individual scan score is generated from the patient's features via fixed scans, and both are trained on the training split and evaluated on held-out test folds (ACDC/Synapse/BTMS). No parameter is fitted to the test statistic and then reported as a prediction; all performance numbers come from forward inference on data not used in training. Author self-citations ([5], [20], [23], [38]) appear only in related-work or baseline contexts and do not carry the argument. There is no imported uniqueness theorem and no ansatz smuggled in via citation. The genuine weakness is Section IV-B: the text states 'the scan order is dynamically determined by sorting these scores' but no equation, pseudocode, or figure shows the conversion of the score into a permutation, and the following sentence says the score is 'concatenated with the corresponding feature embedding.' Fig.5 is captioned 'feature reordering' but contains no reordering operation. That is under-specification/verifiability risk—the reported gains could conceivably come from added parameters/concatenation rather than from an actual learned scan path—but it is not circularity, because even if the claimed sorting were implemented exactly as stated, the derivation would not reduce to its inputs. Hyperparameter choices (e.g., three branches) come from validation performance, which is a selection issue rather than a circular derivation. The paper is benchmarked against external methods, and the code/dataset are promised publicly, so the central results are externally checkable. Score 0.

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

The method rests on a trained group embedding, a trained individual score generator, and a hand-picked branch count; each contributes to the final segmentation and each is fitted without an independent benchmark. No new physical or conceptual entities are postulated beyond the architecture itself. The domain assumptions are the clinical analogy and label reliability.

free parameters (3)
  • Group scan embedding (learnable vector, length n_g) = learned via gradient descent; n_g hyperparameter not reported
    Section IV-B.1: the group scan score is a learnable embedding initialized from a normal distribution; its values are fitted on training data to maximize Dice.
  • Individual scan score generator weights = learned via gradient descent
    Section IV-B.2: the individual scan score comes from a module processing three fixed-scan Mamba outputs; all of its weights are trained.
  • Number of Mamba branches = 3 branches used as default; 4 gave higher non-hard Dice
    Table VI: selected based on performance. The paper does not state whether the selection used the validation or test split.
assumptions (4)
  • standard math Mamba's selective state space model works as a sequence encoder for flattened image features.
    Section IV-B uses Mamba layers without deriving or questioning their sequence-modeling properties.
  • ad hoc to paper A scalar additive scan score, formed from a learned embedding plus image features, determines a good token ordering for Mamba.
    Section IV-B: the entire adaptive scanning mechanism relies on this premise, and no experiment isolates whether the score is used as an ordering or as a concatenated feature channel.
  • domain assumption Clinicians initially scan in fixed directions and then dynamically rescan; this analogy motivates the individual scan score design.
    Section IV-B.2: this is an observational analogy, not empirically validated in the paper.
  • domain assumption BTMS ground-truth labels are anatomically correct and complete.
    Section III: no annotation protocol, inter-observer agreement, or clinician validation is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ASM-UNet: Adaptive Scan Mamba Integrating Group Commonalities and Individual Variations for Fine-Grained Segmentation." pith.science (2026). https://pith.science/paper/IZWFPJIA

@misc{pith2026250807237,
  author       = {Pith},
  title        = {Pith review of: ASM-UNet: Adaptive Scan Mamba Integrating Group Commonalities and Individual Variations for Fine-Grained Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IZWFPJIA}},
  note         = {Machine review of arXiv:2508.07237}
}
read the original abstract

Precise lesion resection depends on accurately identifying fine-grained anatomical structures. While many coarse-grained segmentation (CGS) methods have been successful in large-scale segmentation (e.g., organs), they fall short in clinical scenarios requiring fine-grained segmentation (FGS), which remains challenging due to frequent individual variations in small-scale anatomical structures. Although recent Mamba-based models have advanced medical image segmentation, they often rely on fixed manually-defined scanning orders, which limit their adaptability to individual variations in FGS. To address this, we propose ASM-UNet, a novel Mamba-based architecture for FGS. It introduces adaptive scan scores to dynamically guide the scanning order, generated by combining group-level commonalities and individual-level variations. Experiments on two public datasets (ACDC and Synapse) and a newly proposed challenging biliary tract FGS dataset, namely BTMS, demonstrate that ASM-UNet achieves superior performance in both CGS and FGS tasks. Our code and dataset are available at https://github.com/YqunYang/ASM-UNet.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 34 canonical work pages

  1. [1]

    Prostate-specific membrane antigen-targeted surgery in prostate cancer: accurate identification, real-time diagnosis, and precise resection,

    J. Jiao, J. Zhang, W. Wen, W. Qin, and X. Chen, “Prostate-specific membrane antigen-targeted surgery in prostate cancer: accurate identification, real-time diagnosis, and precise resection,” Theranostics, vol. 14, no. 7, p. 2736, 2024

    work page 2024

  2. [2]

    Nir-ii aiegens nanosystem for fluorescence and chemiluminescence synergistic imaging-guided precise resection in osteosarcoma surgery,

    R. Li et al. , “Nir-ii aiegens nanosystem for fluorescence and chemiluminescence synergistic imaging-guided precise resection in osteosarcoma surgery,” Aggregate, vol. 6, no. 1, p. e658, 2025

    work page 2025

  3. [3]

    Medical image segmentation review: The success of u-net,

    R. Azad et al. , “Medical image segmentation review: The success of u-net,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 10076–10095, 2024

    work page 2024

  4. [4]

    Unetr++: delving into efficient and accurate 3d medical image segmentation,

    A. Shaker, M. Maaz, H. Rasheed, S. Khan, M.-H. Yang, and F. S. Khan, “Unetr++: delving into efficient and accurate 3d medical image segmentation,” IEEE Transactions on Medical Imaging , vol. 43, no. 9, pp. 3377–3390, 2024

    2024

  5. [5]

    Fsvs-net: A few-shot semi-supervised vessel segmentation network for multiple organs based on feature distillation and bidirectional weighted fusion,

    Y . Yang et al. , “Fsvs-net: A few-shot semi-supervised vessel segmentation network for multiple organs based on feature distillation and bidirectional weighted fusion,” Information Fusion, p. 103281, 2025

    work page 2025

  6. [6]

    Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction,

    X. Fang and P. Yan, “Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction,” IEEE Transactions on Medical Imaging , vol. 39, no. 11, pp. 3619–3629, 2020

    work page 2020

  7. [7]

    Abdomenct-1k: Is abdominal organ segmentation a solved problem?

    J. Ma et al., “Abdomenct-1k: Is abdominal organ segmentation a solved problem?” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6695–6714, 2021

    work page 2021

  8. [8]

    Cosst: Multi-organ segmentation with partially labeled datasets using comprehensive supervisions and self-training,

    H. Liu et al. , “Cosst: Multi-organ segmentation with partially labeled datasets using comprehensive supervisions and self-training,” IEEE Transactions on Medical Imaging , vol. 43, no. 5, pp. 1995–2009, 2024

    work page 1995

Show all 47 references
  1. [9]

    Anatomical variations and surgical strategies in right lobe living donor liver transplantation: lessons from 120 cases1,

    T. Nakamura et al. , “Anatomical variations and surgical strategies in right lobe living donor liver transplantation: lessons from 120 cases1,” Transplantation, vol. 73, no. 12, pp. 1896–1903, 2002

    1903

  2. [10]

    Anatomic variations in right liver living donors1,

    G. Varotti et al., “Anatomic variations in right liver living donors1,” Journal of the American College of Surgeons , vol. 198, no. 4, pp. 577–582, 2004

    2004

  3. [11]

    Cnn-based segmentation of medical imaging data,

    B. Kayalibay, G. Jensen, and P. van der Smagt, “Cnn-based segmentation of medical imaging data,” arXiv preprint arXiv:1701.03056 , 2017

    2017 arXiv

  4. [12]

    U-net: Convolutional networks for biomedical image segmentation,

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

    2015

  5. [13]

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

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th ...

    2018

  6. [14]

    Unet 3+: A full-scale connected unet for medical image segmentation,

    H. Huang et al. , “Unet 3+: A full-scale connected unet for medical image segmentation,” in ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2020, pp. 1055–1059. 14

    2020

  7. [15]

    Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer,

    H. Wang, P. Cao, J. Wang, and O. R. Zaiane, “Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer,” in Proceedings of the AAAI conference on artificial intelligence , vol. 36, no. 3, 2022, pp. 2441–2449

    2022

  8. [16]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems , vol. 30, 2017

    2017

  9. [17]

    Attention u-net: Learning where to look for the pancreas,

    O. Oktay et al., “Attention u-net: Learning where to look for the pancreas,” arXiv preprint arXiv:1804.03999 , 2018

    2018 arXiv

  10. [18]

    Missformer: An effective transformer for 2d medical image segmentation,

    X. Huang, Z. Deng, D. Li, X. Yuan, and Y . Fu, “Missformer: An effective transformer for 2d medical image segmentation,” IEEE Transactions on Medical Imaging , vol. 42, no. 5, pp. 1484–1494, 2022

    2022

  11. [19]

    nnformer: volumetric medical image segmentation via a 3d transformer,

    H.-Y . Zhou et al., “nnformer: volumetric medical image segmentation via a 3d transformer,” IEEE Transactions on Image Processing , vol. 32, pp. 4036–4045, 2023

    2023

  12. [20]

    BiCLIP-nnFormer: A Virtual Multimodal Instrument for Efficient and Accurate Medical Image Segmentation,

    B. Wang et al. , “BiCLIP-nnFormer: A Virtual Multimodal Instrument for Efficient and Accurate Medical Image Segmentation,” Instrumentation, vol. 12, no. 2, 2025

    2025

  13. [21]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  14. [22]

    nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model,

    H. Gong et al., “nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model,” in 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI) . IEEE, 2025, pp. 1–5

    2025

  15. [23]

    Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,

    Z. Xing, T. Ye, Y . Yang, G. Liu, and L. Zhu, “Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2024, pp. 578–588

    2024

  16. [24]

    Swin-umamba†: Adapting mamba-based vision foundation models for medical image segmentation,

    J. Liu et al., “Swin-umamba†: Adapting mamba-based vision foundation models for medical image segmentation,” IEEE Transactions on Medical Imaging , early access, 2024

    2024

  17. [25]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

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

    2024 arXiv

  18. [26]

    Localmamba: Visual state space model with windowed selective scan,

    T. Huang, X. Pei, S. You, F. Wang, C. Qian, and C. Xu, “Localmamba: Visual state space model with windowed selective scan,” arXiv preprint arXiv:2403.09338, 2024

    2024 arXiv

  19. [27]

    Mambacsr: Dual-interleaved scanning for compressed image super-resolution with ssms,

    Y . Ren, X. Li, M. Guo, B. Li, S. Zhao, and Z. Chen, “Mambacsr: Dual-interleaved scanning for compressed image super-resolution with ssms,” arXiv preprint arXiv:2408.11758 , 2024

    2024 arXiv

  20. [28]

    Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge,

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

    2015

  21. [29]

    Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?

    O. Bernard et al. , “Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?” IEEE Transactions on Medical Imaging , vol. 37, no. 11, pp. 2514–2525, 2018

    2018

  22. [30]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 3431–3440

    2015

  23. [31]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

  24. [32]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11 976–11 986

    2022

  25. [33]

    Cotr: Efficiently bridging cnn and transformer for 3d medical image segmentation,

    Y . Xie, J. Zhang, C. Shen, and Y . Xia, “Cotr: Efficiently bridging cnn and transformer for 3d medical image segmentation,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 20...

    2021

  26. [34]

    Unetr: Transformers for 3d medical image segmentation,

    A. Hatamizadeh et al., “Unetr: Transformers for 3d medical image segmentation,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2022, pp. 574–584

    2022

  27. [35]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,

    A. Hatamizadeh, V . Nath, Y . Tang, D. Yang, H. R. Roth, and D. Xu, “Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,” in International MICCAI brainlesion workshop . Springer, 2021, pp. 272–284

    2021

  28. [36]

    Vm-unet: Vision mamba unet for medical image segmentation,

    J. Ruan and S. Xiang, “Vm-unet: Vision mamba unet for medical image segmentation,” arXiv preprint arXiv:2402.02491 , 2024

    2024 arXiv

  29. [37]

    Swin-umamba: Mamba-based unet with imagenet-based pretraining,

    J. Liu 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

  30. [38]

    Vivim: a video vision mamba for medical video object segmentation,

    Y . Yang, Z. Xing, and L. Zhu, “Vivim: a video vision mamba for medical video object segmentation,” arXiv preprint arXiv:2401.14168, 2024

    2024 arXiv

  31. [39]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation,

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

    2024 arXiv

  32. [40]

    nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model,

    H. Gong, L. Kang, Y . Wang, X. Wan, and H. Li, “nnmamba: 3d biomedical image segmentation, classification and landmark detection with state space model,” arXiv preprint arXiv:2402.03526 , 2024

    2024 arXiv

  33. [41]

    Lkm-unet: Large kernel vision mamba unet for medical image segmentation,

    J. Wang, J. Chen, D. Chen, and J. Wu, “Lkm-unet: Large kernel vision mamba unet for medical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2024, pp. 360–370

    2024

  34. [42]

    3d image reconstruction for comparison of algorithm database,

    L. Soler et al., “3d image reconstruction for comparison of algorithm database,” URL: https://www. ircad. fr/research/data-sets/liver- segmentation-3d-ircadb-01, 2010

    2010

  35. [43]

    Word: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from ct image,

    X. Luo et al., “Word: A large scale dataset, benchmark and clinical applicable study for abdominal organ segmentation from ct image,” Medical Image Analysis , vol. 82, p. 102642, 2022

    2022

  36. [44]

    Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers,

    J. Chen et al., “Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers,” Medical Image Analysis , vol. 97, p. 103280, 2024

    2024

  37. [45]

    Swin-unet: Unet-like pure transformer for medical image segmentation,

    H. Cao et al., “Swin-unet: Unet-like pure transformer for medical image segmentation,” in European conference on computer vision . Springer, 2022, pp. 205–218

    2022

  38. [46]

    Levit-unet: Make faster encoders with transformer for medical image segmentation,

    G. Xu, X. Zhang, X. He, and 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. [47]

    Segformer3d: an efficient transformer for 3d medical image segmentation,

    S. Perera, P. Navard, and A. Yilmaz, “Segformer3d: an efficient transformer for 3d medical image segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 4981–4988

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.