Pith. sign in

REVIEW 3 major objections 4 minor 29 references

Mamba Guided Boundary Prior Matters: A New Perspective for Generalized Polyp Segmentation

T0 review · 3 major / 4 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 SAM-MaGuP claims that Mamba-guided boundary distillation lets a Segment Anything Model beat existing polyp segmentation methods on both seen and unseen colonoscopy datasets.

desk verdict Solid architecture idea, but the paper's own training objective omits the boundary distillation loss it credits for all its gains, so the central claim is not supported as written. read the letter →

arxiv 2507.01509 v1 pith:NOKK2XWL submitted 2025-07-02 cs.CV cs.LG

Tapas K. Dutta , Snehashis Majhi , Deepak Ranjan Nayak , Debesh Jha This is my paper
classification cs.CVcs.LG
keywords polypsegmentationSegmentAnythingModelMambaboundarydistillationgeneralizedcolonoscopyweak-boundarystate-spacemodels
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 proposes SAM-MaGuP, an adapter-based fine-tuning of the Segment Anything Model for polyp segmentation in colonoscopy images. The authors claim that adding a Mamba-guided boundary prior, composed of a train-only boundary distillation component and a 1D-2D Mamba block, lets SAM resolve weak, blurry polyp boundaries while retaining strong generalization. On five benchmarks, trained on Kvasir-SEG and CVC-ClinicDB, they report consistently higher mean Dice and mean IoU than prior state-of-the-art models, including on three unseen datasets. If the claim holds, it would give clinicians a segmentation model that is both more accurate on ambiguous boundaries and more robust across different imaging setups.

What carries the argument

The central mechanism is the MaGuP (Mamba Guided Boundary Prior) module, a plug-in adapter that injects polyp-specific cues into SAM's frozen image encoder. It contains the 1D-2D Mamba block, which decomposes features with three parallel convolutions of receptive fields 3, 5, and 7 and then runs 1D channel-wise and 2D spatial state-space models with a gated fusion, and the Boundary Distillation Component (BDC), which computes polyp and non-polyp embedding regions from the ground-truth mask, refines them by cross-attention, and minimizes an $L_2$ distillation loss against the encoder embedding. BDC is used only during training, so the final network keeps SAM-Mamba's inference-time parameters while learning boundary-aware features.

What would settle it

Run the full SAM-MaGuP pipeline and an identical run without BDC across repeated seeds on an independent colonoscopy dataset; if removing BDC leaves boundary metrics such as $F_{\beta}^{w}$ and $S_{\alpha}$ statistically unchanged, the claimed boundary-prior benefit is not established.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that SAM's frozen ViT features can be steered toward polyp-relevant, boundary-sensitive representations by a small trainable adapter, the MaGuP module, without full fine-tuning. The 1D-2D Mamba block fuses channel-level and spatial state-space interactions over multi-scale convolutions, while the boundary distillation component uses the ground-truth mask to define polyp and non-polyp embedding regions, refines them with cross-attention, and distills the refined embedding back into SAM's encoder. This sharpens the model's handling of polyps with indistinct boundaries, and because the boundary distillation component is removed at inference, the deployed model keeps SAM-Mamba's inference-time parameter count while reporting higher accuracy on seen and unseen datasets.

Load-bearing premise

The boundary distillation component works by aligning SAM's encoder features with a boundary-enhanced embedding during training, and the model assumes this alignment still sharpens boundaries after the component is removed at inference, a transfer evidenced only by single-run ablations without error bars.

Editorial extensions

If this is right

  • On the five benchmark datasets, SAM-MaGuP reports the best mean Dice and IoU among compared methods, including 94.7% mDice on Kvasir-SEG and 95.3% on CVC-ClinicDB.
  • On unseen datasets, it reports 92.7%, 85.9%, and 85.4% mDice on CVC-300, ColonDB, and ETIS, indicating that the boundary prior transfers beyond the training distribution.
  • The boundary distillation component is discarded at inference, so no ground-truth mask is needed at deployment and the trainable-parameter overhead over SAM-Mamba is only about 3 million parameters (106M versus 103M).
  • The two-stage training schedule and pseudo-mask prompt allow SAM's mask decoder to be adapted without manual prompt engineering, unlike the original SAM.

Reading between the lines

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

  • Because BDC is discarded at inference, it acts as a training-time regularizer; a natural test would be whether simpler boundary-supervised losses achieve the same gain, a comparison the paper does not report.
  • The 1D-2D Mamba adapter is not specific to polyps, so the same adapter-plus-distillation recipe could plausibly transfer to other structures with weak boundaries, such as skin lesions or optic discs, though the paper reports no such experiments.
  • The reported gains over SAM-Mamba are modest relative to the added three million trainable parameters, so a reader should weigh whether the improvement justifies the extra cost on their own data.
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 / 4 minor

Summary. The paper proposes SAM-MaGuP, a SAM-based polyp segmentation framework that combines a 1D-2D Mamba adapter with a boundary distillation component (BDC). The authors claim that the Mamba-guided boundary prior improves weak-boundary segmentation and that SAM-MaGuP consistently outperforms existing state-of-the-art methods on seen and unseen datasets. The method is evaluated on Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, and ETIS, with quantitative comparisons, qualitative examples, and an ablation study. The authors also report a modest parameter overhead over SAM-Mamba (106M vs. 103M trainable parameters).

Significance. If the reported results are reproducible and the claims are correctly scoped, the paper offers a useful contribution: a lightweight adapter-based approach to adapting SAM for polyp segmentation, with a boundary-prior mechanism that is removed at inference time. The datasets are standard, the comparison set is broad, and the authors provide a code link and report inference-time parameter parity with a strong prior baseline. The main value would be in showing that boundary distillation can transfer from training-time ground-truth supervision to inference-time boundary precision. However, the current manuscript has two load-bearing problems: the training objective as written does not include the distillation loss, and the headline claim of outperforming all methods on all metrics is contradicted by the paper's own tables. These issues must be resolved before the contribution can be assessed.

major comments (3)
  1. [Section 2.4, Eq. (2)] The training objective is incomplete: Section 2.4 defines the total loss as L_D = L_Dice^w + L_BCE^w for both Stage I and Stage II, with no mention of L_distill from Eq. (2) or any weighting coefficient. As written, the BDC computes E* in the forward pass, but no gradient from L_distill flows to the encoder, so the BDC cannot influence the learned representation. The ablation gains attributed to BDC in Table 3 (e.g., Kvasir mDice from 93.1 to 94.7, ETIS from 84.9 to 85.4) are therefore not reproducible from the described method. Please specify the complete objective, the loss weight, the training stage(s) in which L_distill is used, and how the pseudo-mask and decoder interact with it.
  2. [Section 3.2, Tables 1 and 2] The claim that SAM-MaGuP "consistently outperforms all existing methods" and achieves the "highest scores across all metrics" is contradicted by the paper's own tables: on Kvasir-SEG, E_max is 98.1 for SAM-MaGuP versus 98.3 for Polyp-Mamba; on ClinicDB, E_max is 98.8 versus 99.3 for Polyp-Mamba; on CVC-300, F_beta is 90.1 versus 93.8 for CFA-Net; and there are ties in MAE on both seen and unseen datasets (e.g., ETIS MAE 1.0 for both SAM-Mamba and SAM-MaGuP). Please revise the claims to match the reported numbers and, if the claim of overall superiority is retained, support it with paired significance tests or a consistent aggregation rule across metrics.
  3. [Section 3.4, Table 3] The ablation study is reported for single runs without error bars, standard deviations, or significance testing. Several incremental gains are small, for example ColonDB mDice 85.5 to 85.9 after adding BDC, and CVC-300 92.2 to 92.7. Without multiple seeds or statistical tests, these differences are within plausible run-to-run variation, and the conclusion that each component, particularly BDC, contributes reliably is not supported. Please report mean and standard deviation over at least three runs, or provide a paired test across the evaluation folds, for both ablation and main tables.
minor comments (4)
  1. [Section 3.2, text after Table 1] The sentence "Its S_alpha (96.5%) and E_max (98.3%) reflect exceptional boundary localization" appears in the Kvasir-SEG paragraph, but neither value matches the Kvasir row of Table 1 (S_alpha 94.2, E_max 98.1) or the ClinicDB row (S_alpha 96.3, E_max 98.8); please correct this inconsistency.
  2. [Abstract and Introduction] Terms such as "groundbreaking," "unmatched," and "set a new benchmark" are not supported by the quantitative comparison as reported; please moderate the language to match the actual differences and acknowledge ties and metrics where the method is not best.
  3. [Figure 1] The architecture diagram is dense and some notation is undefined in the caption or text, including "C_sigma," the exact location of the "Saliency Pool," and the meaning of the "f(theta)" block; please add a legend or descriptive caption.
  4. [Section 2.2, Eq. (1)] The notation E, E_P, E_NP, and E* is introduced informally; please define the dimensions and the cross-attention operation explicitly so that Eq. (2) is self-contained.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claim is supported by held-out evaluation against external baselines, though the BDC loss is omitted from the stated training objective.

full rationale

The paper's derivation is a standard supervised segmentation pipeline: SAM-MaGuP is trained with mask supervision on Kvasir-SEG and ClinicDB training images, then tested on held-out splits and three unseen datasets. The BDC constructs E_P and E_NP from the ground-truth mask, computes a boundary-enhanced E* by cross-attention, and aligns E with E* via L_distill (Eq. 2); this is ordinary mask-supervised representation learning, and removing BDC at inference means the test-time model is not reusing the ground truth. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. The only self-citation, SAM-Mamba [6], serves as a published baseline and architectural starting point; the increments (2D-Mamba, BDC) are evaluated against it and against external methods on held-out data, so the central claim does not reduce to the citation. A non-circularity reproducibility issue must be flagged: Section 2.4 defines the total loss as L_D = L_Dice_w + L_BCE_w for both stages and never includes L_distill, so as written the BDC cannot influence training and Table 3's BDC gain is not reproducible from the stated objective; this weakens the evidence for the boundary-prior claim but is an omitted training term, not a circular derivation.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced; the 'Mamba-guided boundary prior' is a training-time supervision mechanism using ground truth masks, not an independent model or entity. The key free parameters are training hyperparameters and the unspecified distillation loss weight.

free parameters (6)
  • learning rate = 1e-5
    Chosen for Adam optimizer; no sensitivity analysis.
  • batch size = 8
    Training choice; not justified.
  • epochs = 200
    Training duration; no early stopping or convergence analysis.
  • input resolution = 352x352
    Uniform resize for all images; follows prior work.
  • multi-scale scaling factors = {0.75, 1, 1.25}
    Data augmentation choice.
  • distillation loss weight = not reported
    The paper does not specify how L_distill (Eq. 2) is combined with L_D; this is a missing free parameter that affects the method.
assumptions (5)
  • domain assumption The five benchmark datasets and the PraNet training/test split provide a fair and representative evaluation of polyp segmentation.
    The method is judged by this protocol (Section 3.1); any bias in the split affects the central claim.
  • domain assumption Ground truth annotations in the training data accurately locate polyp boundaries.
    The BDC uses the ground truth mask (Eq. 1) as the source of boundary supervision; incorrect labels would corrupt the 'boundary prior'.
  • standard math SAM ViT-H/16 pretrained weights, Mamba/SSM formulations, and the SAM decoder are taken as given from cited prior work.
    The method reuses these components without re-deriving them (Sections 2.1-2.2).
  • ad hoc to paper The boundary distillation transfers from training with ground truth to inference without ground truth.
    The BDC is discarded at inference, so the method assumes the encoder has been permanently improved by the distillation (Section 2.2).
  • domain assumption The improvements over baselines are not due to run-to-run variance.
    No error bars or significance tests are provided, so the small margins depend on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mamba Guided Boundary Prior Matters: A New Perspective for Generalized Polyp Segmentation." pith.science (2026). https://pith.science/paper/NOKK2XWL

@misc{pith2026250701509,
  author       = {Pith},
  title        = {Pith review of: Mamba Guided Boundary Prior Matters: A New Perspective for Generalized Polyp Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOKK2XWL}},
  note         = {Machine review of arXiv:2507.01509}
}
read the original abstract

Polyp segmentation in colonoscopy images is crucial for early detection and diagnosis of colorectal cancer. However, this task remains a significant challenge due to the substantial variations in polyp shape, size, and color, as well as the high similarity between polyps and surrounding tissues, often compounded by indistinct boundaries. While existing encoder-decoder CNN and transformer-based approaches have shown promising results, they struggle with stable segmentation performance on polyps with weak or blurry boundaries. These methods exhibit limited abilities to distinguish between polyps and non-polyps and capture essential boundary cues. Moreover, their generalizability still falls short of meeting the demands of real-time clinical applications. To address these limitations, we propose SAM-MaGuP, a groundbreaking approach for robust polyp segmentation. By incorporating a boundary distillation module and a 1D-2D Mamba adapter within the Segment Anything Model (SAM), SAM-MaGuP excels at resolving weak boundary challenges and amplifies feature learning through enriched global contextual interactions. Extensive evaluations across five diverse datasets reveal that SAM-MaGuP outperforms state-of-the-art methods, achieving unmatched segmentation accuracy and robustness. Our key innovations, a Mamba-guided boundary prior and a 1D-2D Mamba block, set a new benchmark in the field, pushing the boundaries of polyp segmentation to new heights.

Figures

Figures reproduced from arXiv: 2507.01509 by the authors.

Figure 1
Figure 1. Overall framework of the SAM-MaGuP for weak-boundary polyp seg￾mentation. It incorporated the MaGuP Adapter-based fine-tuning into the SAM back￾bone to enhance its representation ability for weak-boundary polyp segmentation. In MaGuP, the BDC is only used during training to refine SAM’s feature learning ability pertaining to diverse polyp shapes and sizes. accurate prompts or define subtle boundaries. Thus, using th… view at source ↗
Figure 2
Figure 2. Comparison of qualitative results on both seen and unseen datasets demon￾strates the SAM-MaGuP’s learning ability to segment polyps with shape, size, texture variations and intricate boundaries, and showcases strong generalization. and reducing false positives by accurately localizing transitions between polyps and the background. These results underscore the critical role of our boundary￾focused modules in deliveri… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 18 canonical work pages

  1. [1]

    Gut and Liver6(1), 64 (2012)

    Ahn, S.B., Han, D.S., et al.: The miss rate for colorectal adenoma determined by quality-adjusted, back-to-back colonoscopies. Gut and Liver6(1), 64 (2012)

    work page 2012

  2. [2]

    saliency maps from physicians

    Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilar- iño, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics 43, 99–111 (2015)

    2015

  3. [3]

    Pattern Recognition45(9), 3166–3182 (2012)

    Bernal, J., Sánchez, J., Vilarino, F.: Towards automatic polyp detection with a polyp appearance model. Pattern Recognition45(9), 3166–3182 (2012)

    2012

  4. [4]

    Bui, N.T., Hoang, D.H., Nguyen, Q.T., Tran, M.T., Le, N.: Meganet: Multi-scale edge-guidedattentionnetworkforweakboundarypolypsegmentation.In:Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 7985–7994 (2024)

    work page 2024

  5. [5]

    arXiv preprint arXiv:2205.08534 (2022)

    Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)

    arXiv 2022

  6. [6]

    In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV)

    Dutta, T.K., Majhi, S., Nayak, D.R., Jha, D.: Sam-mamba: Mamba guided sam architecture for generalized zero-shot polyp segmentation. In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV). pp. 4655–4664 (February 2025)

    work page 2025

  7. [7]

    In: Proceedings of IEEE International Conference on Computer Vision

    Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of IEEE International Conference on Computer Vision. pp. 4548–4557 (2017) 10 Dutta et al

    work page 2017

  8. [8]

    arXiv preprint arXiv:1805.10421 (2018)

    Fan,D.P.,Gong,C.,Cao,Y.,Ren,B.,Cheng,M.M.,Borji,A.:Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)

    arXiv 2018

Show all 29 references
  1. [9]

    In: Medical Image Computing and Computer Assisted Intervention (MICCAI)

    Fan, D.P., Ji, G.P., Zhou, T., Chen, G., Fu, H., Shen, J., Shao, L.: Pranet: Parallel reverse attention network for polyp segmentation. In: Medical Image Computing and Computer Assisted Intervention (MICCAI). pp. 263–273. Springer (2020)

    2020

  2. [10]

    In: 22nd International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2019)

    Fang, Y., Chen, C., Yuan, Y., Tong, K.y.: Selective feature aggregation network with area-boundary constraints for polyp segmentation. In: 22nd International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2019). pp. 302–310. Springer (2019)

    2019

  3. [11]

    In: MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26

    Jha, D., Smedsrud, P.H., Riegler, M.A., Halvorsen, P., De Lange, T., Johansen, D., Johansen, H.D.: Kvasir-seg: A segmented polyp dataset. In: MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26. pp. 451...

    2020

  4. [12]

    In: 2019 IEEE International Symposium on Multimedia (ISM)

    Jha, D., Smedsrud, P.H., Riegler, M.A., Johansen, D., De Lange, T., Halvorsen, P., Johansen, H.D.: Resunet++: An advanced architecture for medical image seg- mentation. In: 2019 IEEE International Symposium on Multimedia (ISM). pp. 225–2255. IEEE (2019)

    2019

  5. [13]

    arXiv:2304.02643 (2023)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)

    2023 arXiv

  6. [14]

    In: Advances in Neural Information Processing systems

    Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Jiao, J., Liu, Y.: Vmamba: Visual state space model. In: Advances in Neural Information Processing systems. pp. 103031–103063 (2025)

    2025

  7. [15]

    Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. pp. 248–255 (2014)

    2014

  8. [16]

    In: Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision

    Rahman, M.M., Marculescu, R.: Medical image segmentation via cascaded atten- tion decoding. In: Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision. pp. 6222–6231 (2023)

    2023

  9. [17]

    In: 18th international conference on Medical image com- puting and computer-assisted intervention (MICCAI 2015)

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: 18th international conference on Medical image com- puting and computer-assisted intervention (MICCAI 2015). pp. 234–241. Springer (2015)

    2015

  10. [18]

    CA: A Cancer Journal for Clinicians73(3), 233–254 (2023)

    Siegel, R.L., Wagle, N.S., Cercek, A., Smith, R.A., Jemal, A.: Colorectal cancer statistics, 2023. CA: A Cancer Journal for Clinicians73(3), 233–254 (2023)

    2023

  11. [19]

    International Journal of Computer Assisted Radiology and Surgery9, 283–293 (2014)

    Silva, J., Histace, A., Romain, O., Dray, X., Granado, B.: Toward embedded detec- tion of polyps in wce images for early diagnosis of colorectal cancer. International Journal of Computer Assisted Radiology and Surgery9, 283–293 (2014)

    2014

  12. [20]

    IEEE Transactions on Medical Imag- ing 35(2), 630–644 (2015)

    Tajbakhsh, N., Gurudu, S.R., Liang, J.: Automated polyp detection in colonoscopy videos using shape and context information. IEEE Transactions on Medical Imag- ing 35(2), 630–644 (2015)

    2015

  13. [21]

    Journal of Healthcare Engineering2017(1), 4037190 (2017)

    Vázquez, D., Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., López, A.M., Romero, A., Drozdzal, M., Courville, A.: A benchmark for endoluminal scene segmentation of colonoscopy images. Journal of Healthcare Engineering2017(1), 4037190 (2017)

    2017

  14. [22]

    arXiv preprint arXiv:2304.12620 (2023)

    Wu, J., Ji, W., Liu, Y., Fu, H., Xu, M., Xu, Y., Jin, Y.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)

    2023 arXiv

  15. [23]

    IEEE Transactions on Cybernetics (2024) Title Suppressed Due to Excessive Length 11

    Xiao, B., Hu, J., Li, W., Pun, C.M., Bi, X.: Ctnet: Contrastive transformer network for polyp segmentation. IEEE Transactions on Cybernetics (2024) Title Suppressed Due to Excessive Length 11

    2024

  16. [24]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Xu, Z., Tang, F., Chen, Z., Zhou, Z., Wu, W., Yang, Y., Liang, Y., Jiang, J., Cai, X., Su, J.: Polyp-mamba: Polyp segmentation with visual mamba. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 510–521. Springer (2024)

    2024

  17. [25]

    In: 24th International Conference on Medical Image Computing and Computer Assisted Intervention–MICCAI 2021

    Zhao, X., Zhang, L., Lu, H.: Automatic polyp segmentation via multi-scale sub- traction network. In: 24th International Conference on Medical Image Computing and Computer Assisted Intervention–MICCAI 2021. pp. 120–130. Springer (2021)

    2021

  18. [26]

    Zhou, T., Zhang, Y., Zhou, Y., Wu, Y., Gong, C.: Can sam segment polyps? arXiv preprint arXiv:2304.07583 (2023)

    2023 arXiv

  19. [27]

    Pattern Recognition 140, 109555 (2023)

    Zhou, T., Zhou, Y., He, K., Gong, C., Yang, J., Fu, H., Shen, D.: Cross-level feature aggregation network for polyp segmentation. Pattern Recognition 140, 109555 (2023)

    2023

  20. [28]

    IEEE Transac- tions on Medical Imaging39(6), 1856–1867 (2019)

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE Transac- tions on Medical Imaging39(6), 1856–1867 (2019)

    2019

  21. [29]

    arXiv preprint arXiv:2401.09417 (2024)

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

    2024 arXiv

Pith tools

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