Pith. sign in

REVIEW 5 major objections 6 minor 39 references

MSA2-Net: Utilizing Self-Adaptive Convolution Module to Extract Multi-Scale Information in Medical Image Segmentation

T0 review · 5 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 MSA2-Net claims that letting convolution kernel sizes be chosen per dataset from quartile statistics is what lets it reach the best reported Dice scores on Synapse, ACDC, Kvasir, and ISIC2017.

desk verdict A promising kernel-selection idea undermined by an undefined QS function and inconsistent reported numbers; worth a careful referee, not acceptance. read the letter →

arxiv 2509.01498 v2 pith:Y4WZZHB2 submitted 2025-09-01 cs.CV cs.AI

Chao Deng , Xiaosen Li , Xiao Qin This is my paper
classification cs.CVcs.AI
keywords medicalimagesegmentationself-adaptiveconvolutionmulti-scalefeatureextractiontransformerU-NetCSWinencoderquartilestatisticsDicecoefficientablationstudy
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 show that convolution kernel size—normally a fixed design choice—should be treated as a trainable, dataset-dependent decision. It introduces a Self-Adaptive Convolution Module that derives candidate kernels from quartile statistics of the dataset and uses a backprop-updated selection matrix to choose among them. The module is inserted in both the skip-connection bridge (MSConvBridge) and the decoder (MSADecoder) of a CSWin-Transformer-based U-Net, giving MSA2-Net. On the four public medical segmentation benchmarks named in the paper it reports average Dice scores of 86.49%, 92.56%, 93.37%, and 92.98%, and the ablation study attributes the largest gains to having both the bridge and decoder modules in place. If this works, it offers a reusable way to make segmentation networks adapt their receptive fields to each dataset rather than relying on a human-set kernel size.

What carries the argument

Self-Adaptive Convolution Module (Eqs. 1-4): a kernel selector whose candidate sizes come from a baseline vector [1,3,5,7] multiplied by a quartile shift matrix computed from the dataset, with the actual choice made by a probability matrix Ws trained by backpropagation. The MSConvBridge applies this module in skip connections together with squeeze-and-excitation and dense connectivity; the MSADecoder splits features into four channel groups, applies adaptive convolutions, and fuses them with squeeze-and-excitation. Together they carry the multi-scale argument: the same network can use different kernel sizes for different datasets.

What would settle it

Train MSA2-Net twice on Synapse with everything identical except that the quartile shift matrix is replaced by a constant zero matrix; if Dice and HD95 stay within noise, dataset-derived kernel selection is not the cause of the reported gains. Separately, compute QS(Dataset) from the training split only and from the whole dataset and compare the selected kernel sizes; if they do not differ, any results depending on that statistic would be suspect.

Watch

Extended reading notes

Core claim

The central claim is that a convolution's receptive field can be auto-tuned to a dataset by combining a fixed baseline kernel set [1,3,5,7] with a quartile-derived shift matrix, then letting a learned selection matrix pick one kernel per operation. The candidate parameter matrix is WC = WB·(1+QS(Dataset)), where QS is a quartile statistical operation on the dataset; a selection probability matrix Ws is updated during backpropagation and selects the kernel. Placing this module in the bridge cleans redundant textures from encoder features, and placing it in the decoder preserves small organs while upsampling. The paper reports that this combination yields the best average Dice on Synapse and A

Load-bearing premise

The load-bearing premise is that QS(Dataset) is a well-defined statistic computed on training data only; the paper does not say what the statistic is or how the data is split, so this assumption cannot currently be checked.

Editorial extensions

If this is right

  • The self-adaptive module is architecture-agnostic: once its equations are fixed, any U-shaped segmentation network can attach it to skip connections or decoders without per-dataset kernel tuning.
  • On Synapse, MSA2-Net wins 5 of 8 per-organ Dice scores and the wins are small organs, suggesting the decoder's adaptive convolutions specifically recover small structures during upsampling.
  • The ablation shows removing both MSADecoder and MSConvBridge costs 10.10 Dice points on Synapse, so the module's placement, not the CSWin encoder alone, carries most of the reported improvement.
  • Kvasir results are weaker despite high overall scores, which the paper reads as evidence that the mechanism still struggles with organs whose shape and area vary widely.

Reading between the lines

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

  • An image-level rather than dataset-level variant, where the quartile statistic is computed per input organ-area distribution, would be a natural testable extension that might improve high-shape-variance cases like polyp datasets.
  • The dataset-fingerprint idea could transfer to instance segmentation outside medicine, where object scale varies within a single image; the module's current form only adapts globally per dataset.
  • The paper leaves QS unspecified; the first reproducibility check is to pin down its definition and which data split feeds it, and that check matters more than any architecture detail.
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

5 major / 6 minor

Summary. MSA2-Net proposes a transformer-based U-Net for medical image segmentation whose main novelty is a Self-Adaptive Convolution Module. The module is intended to select convolution kernel sizes from a candidate matrix derived from dataset 'quartile statistics,' and it is inserted into the MSConvBridge (skip connections) and MSADecoder (decoder). The paper reports state-of-the-art or competitive Dice scores on Synapse, ACDC, Kvasir-SEG, and ISIC skin-lesion datasets, and includes ablations of the bridge and decoder. The central presentation issue is that the adaptive mechanism is not fully specified: the key quantity QS(Dataset) in Eq. (4) is never defined, and the selection rule in Eq. (2) uses a non-differentiable max over a probability matrix. No code is provided, so the core mechanism cannot be reproduced from the paper as written.

Significance. If the Self-Adaptive Convolution Module worked as described, the contribution would be a simple, reusable way to adapt kernel sizes to dataset characteristics, with potential applicability to multi-scale medical image segmentation across CT, MRI, endoscopy, and dermoscopy. The paper's strengths are the breadth of the evaluation (four public datasets), the inclusion of an ablation study, and the qualitative comparisons. However, the central equations are incomplete or internally inconsistent, and the reported numbers are not statistically supported. As presented, the paper does not establish the validity of its main claimed mechanism, so the significance of the contribution remains unverified.

major comments (5)
  1. [III-A, Eq. (4)] QS(Dataset) is the load-bearing definition of the entire method, but it is never defined. The paper does not state which statistic is computed (e.g., first or third quartile of what quantity), on which data split (train/validation/test) it is computed, or how it maps to a 1×4 quartile shift matrix. Table 5 reports concrete kernel arrays for Q1/Q2/Q3 and 'Self-Adaptive,' but no formula connects those arrays to QS. As written, the claim of dataset-adaptive, leakage-free kernel selection is unverifiable and the implementation cannot be reproduced.
  2. [III-A, Eq. (2)] The selection rule is written as i_hat = max(Ws), where Ws is a selection probability matrix. Taking the maximum is non-differentiable with respect to Ws. The text states that Ws is updated during each backpropagation iteration, but no straight-through estimator, Gumbel-Softmax relaxation, or alternative is described. As written, the 'self-adaptive' selection either cannot be trained end-to-end or relies on an undisclosed relaxation. This is a technical gap in the core module, not a presentation issue.
  3. [Abstract vs. Tables 2 and 5] The Kvasir-SEG Dice score is reported as 93.37% in the abstract, but Table 2 and Table 5 report 91.49%. The headline number is therefore inconsistent with the experimental section. This discrepancy must be corrected before the reported gains over prior methods can be assessed.
  4. [IV-D, Tables 1–3] No error bars, standard deviations, or number of training runs are reported. Several claimed improvements over the best baseline are small (e.g., +0.75 Dice over AgileFormer on Synapse, +1.10 over CSwin-UNet on ACDC). Baseline results are taken from previously published papers rather than re-run under a common protocol. Statistical significance is not established, so the state-of-the-art claim is not supported.
  5. [III-A and Table 5] The term 'dynamic adjustment' is not supported by the evidence. Table 5 lists one fixed kernel array per dataset and component (e.g., Synapse MSADecoder Stage1 is [1,3,3,7]), with no indication that these values change during training. If the kernel sizes are selected once before training from dataset statistics, they are static hyperparameters. No experiment shows the kernels adapting over time, weakening the paper's claimed distinction from preset kernel-size schemes.
minor comments (6)
  1. [Eq. (13)] The Dice formula is written as 2 × (pred ∩ true)/(pred ∪ true). It should use cardinalities, e.g., 2|X∩Y|/(|X|+|Y|).
  2. [Eq. (14)] The equation gives the standard Hausdorff distance, not the 95th percentile Hausdorff distance (HD95) used in the experiments.
  3. [IV-B, Kvasir-SEG] The dataset description says there is a standardized split of 880 training and 120 validation images, then states 'Following the same approach as previous studies [18], we use 70 cases for training, 10 for validation, and 20 for testing.' These two statements are inconsistent.
  4. [IV-A] Input images are said to be resized to 224×244 pixels; this is likely a typo for 224×224.
  5. [Throughout] There are repeated typos, e.g., 'MSADeocder,' 'MSConvBrdige,' 'networkss,' and 'Specif icity.'
  6. [References] Reference [28] is cited for KDAS3 but the listed title is about e-government maturity models; this appears to be an incorrect reference and should be verified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found: the self-adaptive kernel selection is not defined in terms of the target Dice metric, and reported gains are compared against external baselines rather than derived from fitted constants.

full rationale

The central claim—that the Self-Adaptive Convolution Module adjusts kernel sizes using dataset fingerprints and thereby improves segmentation—does not reduce to its inputs by construction. The kernel selection in Eqs. 1–4 takes as input WB=[1,3,5,7] and a quartile statistic QS(Dataset); it does not take the reported Dice/HD95 scores as input, and no equation defines QS in terms of the target metric. The network is then evaluated against externally published baselines, not against any quantity fitted from the same test outputs. Table 5 compares the 'Self-Adaptive' configuration to fixed quartile-weighting schemes and to 'None', and although QS is never formally defined and the argmax in Eq. 2 is non-differentiable, those are reproducibility/correctness concerns (and possible test-set leakage), not circularity. There is no load-bearing self-citation chain: the paper's dependencies are external methods (CSWin, TransXNet/IDConv, MISSFormer, nnUNet), and the TransXNet reference is actually missing from the reference list rather than being a self-citation. The acknowledged limitation on Kvasir-SEG ('MSA2-Net still lacks the capability to handle organs with highly variable shapes') is an honest statement of scope, not evidence of circular derivation. Therefore the paper is not circular in the sense of predictions reducing to fitted inputs or definitions; score 0.

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

The paper contributes an architecture whose central mechanism depends on an undefined dataset statistic and a non-differentiable selection rule. The free parameters are the per-dataset QS values, the learned selection matrix, and the hand-picked baseline kernel pool. No new physical or conceptual entities are introduced.

free parameters (3)
  • QS(Dataset) quartile shift values = Not disclosed; resulting kernel arrays listed for Synapse in Table 5, no values for other datasets
    The self-adaptive kernel sizes are derived from an unspecified quartile statistic of each dataset. These per-dataset values directly determine the candidate kernel matrix WC (Eq. 3-4) and hence the receptive field, but are never defined or reported.
  • Selection probability matrix Ws = Learned during training, final values not reported
    Ws selects the kernel element via argmax (Eq. 2). Its initialization, update rule, and final values are not given, so the actual behavior of the module cannot be checked.
  • Baseline kernel matrix WB = [1,3,5,7] = [1,3,5,7]
    Adopted from the unreferenced IDConv/TransXNet configuration. This choice fixes the candidate kernel-size pool and is not derived from the data.
assumptions (4)
  • ad hoc to paper QS(Dataset), the quartile statistical operation, is well-defined and computable from the dataset before training.
    Introduced in Eq. 4 with no definition of the operation or the distribution it is applied to. Figure 2 suggests organ area proportions, but no algorithm is given.
  • domain assumption Argmax of the selection probability matrix Ws yields a kernel index that can be optimized by backpropagation.
    Eq. 2 uses max(Ws) to select a parameter from WC. The paper does not describe a differentiable relaxation such as softmax or Gumbel, so end-to-end training of the discrete selection is unsubstantiated.
  • domain assumption Kernel size governs capture of organ-scale features: large kernels amplify large organs, small kernels highlight small organs.
    Figure 1 and Section III-A assert this heuristic. It motivates the whole module but is not formally supported or validated independently.
  • domain assumption CSWin pretrained weights transfer to medical image segmentation.
    The encoder uses pre-trained CSWin weights. Transferability is assumed and not ablated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MSA2-Net: Utilizing Self-Adaptive Convolution Module to Extract Multi-Scale Information in Medical Image Segmentation." pith.science (2026). https://pith.science/paper/Y4WZZHB2

@misc{pith2026250901498,
  author       = {Pith},
  title        = {Pith review of: MSA2-Net: Utilizing Self-Adaptive Convolution Module to Extract Multi-Scale Information in Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y4WZZHB2}},
  note         = {Machine review of arXiv:2509.01498}
}
read the original abstract

The nnUNet segmentation framework adeptly adjusts most hyperparameters in training scripts automatically, but it overlooks the tuning of internal hyperparameters within the segmentation network itself, which constrains the model's ability to generalize. Addressing this limitation, this study presents a novel Self-Adaptive Convolution Module that dynamically adjusts the size of the convolution kernels depending on the unique fingerprints of different datasets. This adjustment enables the MSA2-Net, when equipped with this module, to proficiently capture both global and local features within the feature maps. Self-Adaptive Convolution Module is strategically integrated into two key components of the MSA2-Net: the Multi-Scale Convolution Bridge and the Multi-Scale Amalgamation Decoder. In the MSConvBridge, the module enhances the ability to refine outputs from various stages of the CSWin Transformer during the skip connections, effectively eliminating redundant data that could potentially impair the decoder's performance. Simultaneously, the MSADecoder, utilizing the module, excels in capturing detailed information of organs varying in size during the decoding phase. This capability ensures that the decoder's output closely reproduces the intricate details within the feature maps, thus yielding highly accurate segmentation images. MSA2-Net, bolstered by this advanced architecture, has demonstrated exceptional performance, achieving Dice coefficient scores of 86.49\%, 92.56\%, 93.37\%, and 92.98\% on the Synapse, ACDC, Kvasir, and Skin Lesion Segmentation (ISIC2017) datasets, respectively. This underscores MSA2-Net's robustness and precision in medical image segmentation tasks across various datasets.

Figures

Figures reproduced from arXiv: 2509.01498 by the authors.

Figure 1
Figure 1. FIGURE 1: Impact of Different Kernel Sizes on Feature [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIGURE 2: (A) Schematic diagram of the Self-Adaptive Convolution Module and its parameter auto-update workflow. (B) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIGURE 3: Architectural Overview of MSA [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FIGURE 4: MSConvBrdige Architecture Diagram [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: FIGURE 5: (a) CSWin feature maps without MSConvBridge [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: FIGURE 6: MSADecoder Architecture Diagram [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: FIGURE 7: Segmentation effects of MSA [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: FIGURE 8: the box plots of the area proportions of organs in [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [1]

    Deep semantic segmentation of natural and medical im- ages: a review,

    S. Asgari Taghanaki, K. Abhishek, J. P. Cohen, J. Cohen-Adad, and G. Hamarneh, “Deep semantic segmentation of natural and medical im- ages: a review,” Artificial Intelligence Review, vol. 54, pp. 137–178, 2021

    work page 2021

  2. [2]

    Medical image segmentation using deep semantic- based methods: A review of techniques, applications and emerging trends,

    I. Qureshi, J. Yan, Q. Abbas, K. Shaheed, A. B. Riaz, A. Wahid, M. W. J. Khan, and P. Szczuko, “Medical image segmentation using deep semantic- based methods: A review of techniques, applications and emerging trends,” Information Fusion, vol. 90, pp. 316–352, 2023

    work page 2023

  3. [3]

    Medical image segmentation using deep learning: A survey,

    R. Wang, T. Lei, R. Cui, B. Zhang, H. Meng, and A. K. Nandi, “Medical image segmentation using deep learning: A survey,” IET image processing, vol. 16, no. 5, pp. 1243–1267, 2022

    work page 2022

  4. [4]

    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, Conference Proceedings, pp. 3431–3440

    work page

  5. [5]

    On the texture bias for few-shot cnn segmentation,

    R. Azad, A. R. Fayjie, C. Kauffmann, I. Ben Ayed, M. Pedersoli, and J. Dolz, “On the texture bias for few-shot cnn segmentation,” in Pro- ceedings of the IEEE/CVF winter conference on applications of computer vision, Conference Proceedings, pp. 2674–2683

    work page

  6. [6]

    Transunet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,” arXiv preprint arXiv:2102.04306, 2021

    arXiv 2021

  7. [7]

    Short- term and long-term memory self-attention network for segmentation of tumours in 3d medical images,

    M. Wen, Q. Zhou, B. Tao, P. Shcherbakov, Y . Xu, and X. Zhang, “Short- term and long-term memory self-attention network for segmentation of tumours in 3d medical images,” CAAI Transactions on Intelligence Tech- nology, vol. 8, no. 4, pp. 1524–1537, 2023

    work page 2023

  8. [8]

    Missformer: An effective medical image segmentation transformer,

    X. Huang, Z. Deng, D. Li, and X. Yuan, “Missformer: An effective medical image segmentation transformer,” arXiv preprint arXiv:2109.07162, 2021

    arXiv 2021

Show all 39 references
  1. [9]

    Disegnet: A deep dilated convolutional encoder-decoder architecture for lymph node segmentation on pet/ct images,

    G. Xu, H. Cao, J. K. Udupa, Y . Tong, and D. A. Torigian, “Disegnet: A deep dilated convolutional encoder-decoder architecture for lymph node segmentation on pet/ct images,” Computerized Medical Imaging and Graphics, vol. 88, p. 101851, 2021

    2021

  2. [10]

    Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning,

    L. Chen, H. Zhang, J. Xiao, L. Nie, J. Shao, W. Liu, and T.-S. Chua, “Sca-cnn: Spatial and channel-wise attention in convolutional networks for image captioning,” in Proceedings of the IEEE conference on computer vision and pattern recognition, Conference Proceedings, pp. 5659–5667

  3. [11]

    Adaptive spatial pixel-level feature fusion network for multispectral pedestrian detection,

    L. Fu, W.-b. Gu, Y .-b. Ai, W. Li, and D. Wang, “Adaptive spatial pixel-level feature fusion network for multispectral pedestrian detection,” Infrared Physics & Technology, vol. 116, p. 103770, 2021

    2021

  4. [12]

    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

  5. [13]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows,

    X. Dong, J. Bao, D. Chen, W. Zhang, N. Yu, L. Yuan, D. Chen, and B. Guo, “Cswin transformer: A general vision transformer backbone with cross-shaped windows,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Conference Proceedings, pp. 12 1...

  6. [14]

    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 VOLUME 4, 2016 11 Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS Learning in Medical Image Analysis and Multimoda...

    2016

  7. [15]

    Age estimation from mr images via 3d convolutional neural network and densely connect,

    Q. Qi, B. Du, M. Zhuang, Y . Huang, and X. Ding, “Age estimation from mr images via 3d convolutional neural network and densely connect,” in International Conference on Neural Information Processing. Springer, 2018, pp. 410–419

    2018

  8. [16]

    Do- main adaptive relational reasoning for 3d multi-organ segmentation,

    S. Fu, Y . Lu, Y . Wang, Y . Zhou, W. Shen, E. Fishman, and A. Yuille, “Do- main adaptive relational reasoning for 3d multi-organ segmentation,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2...

    2020

  9. [17]

    Medical image segmentation via cascaded attention decoding,

    M. M. Rahman and R. Marculescu, “Medical image segmentation via cascaded attention decoding,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, Conference Proceedings, pp. 6222–6231

  10. [18]

    Transresu- net: Transformer based resu-net for real-time colonoscopy polyp segmen- tation,

    N. K. Tomar, A. Shergill, B. Rieders, U. Bagci, and D. Jha, “Transresu- net: Transformer based resu-net for real-time colonoscopy polyp segmen- tation,” arXiv preprint arXiv:2206.08985, 2022

    2022 arXiv

  11. [19]

    Hiformer: Hierarchical multi-scale representations using transformers for medical image segmentation,

    M. Heidari, A. Kazerouni, M. Soltany, R. Azad, E. K. Aghdam, J. Cohen- Adad, and D. Merhof, “Hiformer: Hierarchical multi-scale representations using transformers for medical image segmentation,” in Proceedings of the IEEE/CVF winter conference on applications of computer visi...

    2023

  12. [20]

    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 confer- ence, Munich, Germany, October 5-9, 2015, proceedings, part III 18....

    2015

  13. [21]

    Att-unet: Pixel-wise staircase attention for weed and crop detection,

    X.-Z. Hu, W.-S. Jeon, and S.-Y . Rhee, “Att-unet: Pixel-wise staircase attention for weed and crop detection,” in 2023 International Conference on Fuzzy Theory and Its Applications (iFUZZY). IEEE, 2023, pp. 1–5

    2023

  14. [22]

    Stepwise feature fusion: Local guides global,

    J. Wang, Q. Huang, F. Tang, J. Meng, J. Su, and S. Song, “Stepwise feature fusion: Local guides global,” in International Conference on Med- ical Image Computing and Computer-Assisted Intervention. Springer, Conference Proceedings, pp. 110–120

  15. [23]

    Polyp-pvt: Polyp segmentation with pyramid vision transformers,

    B. Dong, W. Wang, D.-P. Fan, J. Li, H. Fu, and L. Shao, “Polyp-pvt: Polyp segmentation with pyramid vision transformers,” arXiv preprint arXiv:2108.06932, 2021

    2021 arXiv

  16. [24]

    Mixed transformer u-net for medical image segmentation,

    H. Wang, S. Xie, L. Lin, Y . Iwamoto, X.-H. Han, Y .-W. Chen, and R. Tong, “Mixed transformer u-net for medical image segmentation,” in ICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, Conference Proceedings, pp. 2390–2394

    2022

  17. [25]

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

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

  18. [26]

    Class-aware adversarial transformers for medical image seg- mentation,

    C. You, R. Zhao, F. Liu, S. Dong, S. Chinchali, U. Topcu, L. Staib, and J. Duncan, “Class-aware adversarial transformers for medical image seg- mentation,” Advances in neural information processing systems, vol. 35, pp. 29 582–29 596, 2022

    2022

  19. [27]

    Bdg-net: boundary distribution guided network for accurate polyp segmentation,

    Z. Qiu, Z. Wang, M. Zhang, Z. Xu, J. Fan, and L. Xu, “Bdg-net: boundary distribution guided network for accurate polyp segmentation,” in Medical Imaging 2022: Image Processing, vol. 12032. SPIE, 2022, pp. 792–799

    2022

  20. [28]

    Identify- ing weaknesses for chilean e-government implementation in public agen- cies with maturity model,

    M. Solar, H. Astudillo, G. Valdes, M. Iribarren, and G. Concha, “Identify- ing weaknesses for chilean e-government implementation in public agen- cies with maturity model,” in Electronic Government: 8th International Conference, EGOV 2009, Linz, Austria, August 31-September 3,...

    2009

  21. [29]

    U-net++ dsm: improved u-net++ for brain tumor segmen- tation with deep supervision mechanism,

    K. Wisaeng, “U-net++ dsm: improved u-net++ for brain tumor segmen- tation with deep supervision mechanism,” IEEE Access, vol. 11, pp. 132 268–132 285, 2023

    2023

  22. [30]

    Unetr: Transformers for 3d medical image segmentation,

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Landman, H. R. Roth, and D. Xu, “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

  23. [31]

    Polyp-sam++: Can a text guided sam perform better for polyp segmentation?

    R. Biswas, “Polyp-sam++: Can a text guided sam perform better for polyp segmentation?” arXiv preprint arXiv:2308.06623, 2023

    2023 arXiv

  24. [32]

    Connecting targets via latent topics and contrastive learning: A unified framework for robust zero-shot and few-shot stance detection,

    R. Liu, Z. Lin, P. Fu, Y . Liu, and W. Wang, “Connecting targets via latent topics and contrastive learning: A unified framework for robust zero-shot and few-shot stance detection,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (I...

    2022

  25. [33]

    Cswin-unet: Transformer unet with cross-shaped windows for medical image segmentation,

    X. Liu, P. Gao, T. Yu, F. Wang, and R.-Y . Yuan, “Cswin-unet: Transformer unet with cross-shaped windows for medical image segmentation,” Infor- mation Fusion, vol. 113, p. 102634, 2025

    2025

  26. [34]

    Pefnet: Position enhancement faster network for object detection in roadside perception system,

    L. Huang, W. Huang, H. Gong, C. Yu, and Z. You, “Pefnet: Position enhancement faster network for object detection in roadside perception system,” IEEE Access, 2023

    2023

  27. [35]

    St-unet: A spatio-temporal u-network for graph-structured time series modeling,

    B. Yu, H. Yin, and Z. Zhu, “St-unet: A spatio-temporal u-network for graph-structured time series modeling,” arXiv preprint arXiv:1903.05631, 2019

    1903 arXiv

  28. [36]

    Transnetr: transformer- based residual network for polyp segmentation with multi-center out-of- distribution testing,

    D. Jha, N. K. Tomar, V . Sharma, and U. Bagci, “Transnetr: transformer- based residual network for polyp segmentation with multi-center out-of- distribution testing,” in Medical Imaging with Deep Learning. PMLR, 2024, pp. 1372–1384

    2024

  29. [37]

    Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,

    L. Wang, R. Li, C. Zhang, S. Fang, C. Duan, X. Meng, and P. M. Atkinson, “Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 190, pp. 196–214, 2022

    2022

  30. [38]

    Resunet++: An advanced architecture for medical image segmentation,

    D. Jha, P. H. Smedsrud, M. A. Riegler, D. Johansen, T. De Lange, P. Halvorsen, and H. D. Johansen, “Resunet++: An advanced architecture for medical image segmentation,” in 2019 IEEE international symposium on multimedia (ISM). IEEE, 2019, pp. 225–2255

    2019

  31. [39]

    Contextual attention network: Transformer meets u-net,

    A. Reza, H. Moein, W. Yuli, and M. Dorit, “Contextual attention network: Transformer meets u-net,” arXiv preprint arXiv:2203.01932, 2022. VI. ACKNOWLEDGMENT Special thanks to Yuanyuan Li for her insightful comments and suggestions that helped improve this manuscript. 12 VOLUME 4, 2016

    2022 arXiv

Pith tools

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