Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Med-URWKV{\dag}: Toward Enhanced Pretrained Pure VRWKV Models for Medical Image Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A pure pretrained VRWKV encoder–decoder, augmented with wavelet-frequency attention and multi-scale channel fusion, reports an 88.00% average Dice score across five medical segmentation benchmarks with 24.41M parameters.

desk verdict Interesting pure-VRWKV segmentation architecture, but the 88.00% DSC headline is selected on the test set, so the SOTA claim does not stand as reported. read the letter →

arxiv 2506.10858 v2 pith:CBLSVB7S submitted 2025-06-12 eess.IV cs.CV

classification eess.IVcs.CV
keywords medicalimagesegmentationVisionRWKVpurearchitectureNetpretrainingwaveletattentionfrequency-domainmodelingmulti-scalefeaturefusionparameterefficiency
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

Med-URWKV† is built to show that a pure Vision-RWKV (VRWKV) network — an RWKV-style linear-attention architecture adapted for vision — can be a competitive medical image segmenter when its encoder starts from large-scale pretrained weights, rather than being confined to hybrid CNN-RWKV designs trained from scratch. The authors construct Med-URWKV-T and Med-URWKV-S by pairing pretrained VRWKV encoders with pure VRWKV decoders, then add a Frequency-Aware Wavelet Attention module that lets high-frequency edge detail be guided by low-frequency global structure, and a Multi-Scale Channel Fusion module that merges decoder features across scales. The enhanced model reports the highest average Dice similarity coefficient among the compared methods — 88.00% across five public datasets — while using 24.41M parameters, less than half of Med-URWKV-S. The significance, if the comparison holds, is that pretrained pure RWKV backbones can match or exceed CNN, Transformer, and Mamba segmenters, and that targeted efficiency modules can outperform simply scaling up parameters.

What carries the argument

The mechanism that carries the argument is the VRWKV block's Bi-WKV attention, a linear-time token-mixing operation in which each output is a decayed, position-dependent weighted sum of values, gated by a Sigmoid-activated receptance and stabilized by a current-token gain. The paper's two modules are built on that same block: FAWA applies a Haar wavelet transform to encoder features, feeds the low-frequency sub-band as the key and value while all four sub-bands serve as the receive vector, and reconstructs the result with an inverse wavelet transform; MSCF upsamples decoder features, applies VRWKV Channel Mix with the deepest global features as the key, and adds the result back to the concatenated multi-scale features. The pretrained VRWKV encoder weights, taken from a large-scale visual recognition checkpoint, are the component the paper credits for the transfer.

What would settle it

Run the same five-dataset protocol with every method initialized from its official pretrained weights (including the three baselines that already use them); if Med-URWKV†'s average-DSC lead over Swin-UMamba shrinks to within the reported run-to-run variance, the pretraining asymmetry is the decisive factor.

Watch

Extended reading notes

Core claim

On the paper's terms, the central discovery is that large-scale pretrained pure VRWKV encoders transfer well to medical image segmentation and that a pure VRWKV decoder can exploit them without any CNN or Transformer mixing. The evidence is the reported 88.00% average DSC, best on four of five datasets, and the 1.80-point average DSC margin over the strongest non-RWKV competitor while using less than half that model's parameters. A second finding is that scale is not the bottleneck: Med-URWKV-T (14.33M parameters) and Med-URWKV-S (53.07M) achieve nearly identical average DSC (86.74% and 86.70%), whereas the FAWA and MSCF modules lift Med-URWKV-T to 88.00% at 24.41M parameters. The paper takes this as evidence that pretrained pure VRWKV models are a viable foundation for medical segmentation and that frequency-domain and multi-scale enhancements are what unlock their performance.

Load-bearing premise

The comparison assumes equal footing, but most competing baselines are randomly initialized while Med-URWKV starts from large-scale pretrained VRWKV weights, so the claimed gains could come from pretraining rather than from the pure architecture.

Editorial extensions

If this is right

  • If the results hold, a pure pretrained VRWKV network can serve as the entire backbone for medical segmentation, without CNN, Transformer, or Mamba components.
  • Pretraining is essential to the reported performance: on BUSI, Kvasir-SEG, and GLAS, the pretrained encoder raises best DSC by up to 26.39 points over random initialization.
  • Increasing encoder scale from Tiny to Small yields almost no average-DSC gain (86.74% vs 86.70%), so the paper's modules rather than parameter count carry the improvement.
  • FAWA and MSCF are complementary: the full Med-URWKV† reaches its best BUSI DSC of 80.90% and Kvasir-SEG DSC of 91.41% only when both are present.
  • Med-URWKV† reports the highest average DSC (88.00%) of the compared models, surpassing all 17 competitors and beating the best alternative by 1.80 points with less than half its parameters.

Reading between the lines

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

  • I read the reported advantage as partly confounded by initialization: because most baselines are trained from scratch while Med-URWKV starts from pretrained VRWKV weights, the cleanest test would rerun the comparison with every encoder initialized from official pretrained weights.
  • The FAWA design — low-frequency sub-band as key/value and all sub-bands as receive vector — is not segmentation-specific and could transfer to other RWKV-based vision tasks such as detection or restoration.
  • The near-identical results of the Tiny and Small pretrained encoders suggest that future gains may come from larger VRWKV backbones with stronger pretraining, rather than from concluding that scale saturates.
  • The fixed 10-epoch encoder freeze is untested; ablating the freeze schedule would clarify whether the transfer is robust or sensitive to optimization details.
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. Med-URWKV† introduces a pure VRWKV-based encoder-decoder for medical image segmentation. Tiny and Small variants reuse pretrained VRWKV encoders, and the proposed FAWA (wavelet-based frequency attention) and MSCF (multi-scale channel fusion) modules are added to form Med-URWKV†. The paper reports an average Dice of 88.00% across five public datasets, outperforming 17 CNN, ViT, Mamba, and RWKV baselines with 24.41M parameters, and presents ablations showing large gains from pretraining and from the two modules.

Significance. The manuscript is a thorough empirical study with a clear research question: whether pretrained pure VRWKV models can compete in medical segmentation. Strengths include the systematic design of Med-URWKV-T/S, the clean isolation of FAWA and MSCF contributions in Table 3, and the breadth of the comparison with triplicate runs. The central numerical claim, however, is undermined by the use of the test set for early stopping (Sec 4.2) and by the asymmetric initialization of baselines. As reported, the SOTA claim is not reliable, though the underlying architecture may still be valuable once these methodological problems are corrected.

major comments (3)
  1. [4.2] The early stopping rule in Sec 4.2 states that 'training was terminated if the performance on the test set did not improve for 10 consecutive epochs.' This uses the test set for model selection, so the DSC values in Table 1, including the headline 88.00% average, are selected maxima over training trajectories and are optimistically biased estimates of generalization. The comparison also favors whichever model overfits the test set most aggressively, so the central claim of state-of-the-art performance is not established. Please retrain (or reselect epochs) using a held-out validation split and report the test numbers selected by validation; if that is infeasible, present both validation-selected and test-selected results so the bias can be quantified.
  2. [4.2 / Table 1] The evaluation is not a level playing field: the proposed models use pretrained VRWKV encoders, while Sec 4.2 states that all competing methods except PVT-EMCAD-b1, Swin-UMamba, and RWKV-UNet were 'trained from scratch with random initialization.' Because pretraining itself can produce large gains (as the paper's own Table 2 shows), the reported superiority over most of the 17 baselines conflates pretraining with architectural merit. Please include at least one randomly initialized Med-URWKV baseline in Table 1 and, ideally, pretrained variants of the strongest baselines, and soften the claim of 'consistently outperforming 17 representative methods' accordingly.
  3. [Section 4.2, Tables 1-3] Although all experiments were repeated three times, only the means are reported; no standard deviations or confidence intervals appear in Tables 1-3. Several key comparisons are very close (e.g., Med-URWKV-T at 86.74% vs. Med-URWKV-S at 86.70% average DSC), so without variance information the reader cannot determine whether the differences are meaningful. Please report per-trial standard deviations or a significance test for the main comparisons, particularly for the SOTA claim.
minor comments (5)
  1. [4.3] The text says 'Med-URWKV† and Med-URWKV-S rank first and second, achieving average DSC scores of 88.00% and 86.74%, respectively,' but Table 1 gives Med-URWKV-T the 86.74% average and Med-URWKV-S 86.70%. Please correct the text or the table.
  2. [1 vs 4.2] The introduction refers to 'ImageNet-pretrained VRWKV encoders,' but Sec 4.2 cites the checkpoints 'upernet_vrwkv_adapter_tiny_512_160k_ade20k' and 'upernet_vrwkv_adapter_small_512_160k_ade20k,' which appear to be ADE20k-trained. Please clarify the exact pretraining data and whether the ImageNet claim is accurate.
  3. [Abstract] The abstract and Section 5 promise that 'the codes will be released,' but no repository or link is provided. Please either provide a link or remove the promise.
  4. [Table 2] Table 2's 'best' row is not explicitly defined; please state whether it corresponds to the best test-set epoch or to a validation-based selection, and add the epoch index at which 'best' is reached.
  5. [Throughout] There are numerous typographical and notation errors, e.g., 'defferent' in Sec 3.3, 'Spacial Mix' and 'Squsred ReLU' in Fig. 4, 'Channal Mix' in Fig. 4, and 'Tianjn Eye Hospital' in the affiliations. Please proofread carefully.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: benchmark results are measured outputs and pretrained weights are external support; self-citations are not load-bearing.

full rationale

Med-URWKV and Med-URWKV† are presented as empirical architecture studies, not as a formal derivation from first principles. The headline 88.00% average DSC is a measured result on held-out test partitions of five public benchmarks (Table 1), and no model parameter is fitted to a target and then reported as a prediction. The pretrained VRWKV encoders come from the external official VRWKV checkpoints named in Section 4.2 (upernet_vrwkv_adapter_tiny_512_160k_ade20k and upernet_vrwkv_adapter_small_512_160k_ade20k), which provide independent support. The central transfer claim is tested directly in Table 2 by comparing the same architecture with and without pretrained initialization, so it is not assumed by construction. The FAWA and MSCF modules are new architectural components whose contributions are measured by ablations in Table 3 rather than being defined in terms of the outcome. Some references are to the authors' prior work, notably [59] for frequency-domain importance and [18] for dataset partitioning protocols, but these citations are not load-bearing: [58] independently supports frequency-domain learning, and the module ablations stand on their own. The test-set early-stopping protocol described in Section 4.2 may raise a concern about benchmark validity, but that is an experimental-design issue, not a circular-equivalence issue: it does not make any reported quantity equal to an input by definition or by fitting. Overall, the paper's empirical claims are self-contained measurements with external pretrained checkpoints and independent baselines.

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

The paper does not introduce ontologically new entities; FAWA and MSCF are architectural modules built from wavelet transforms and existing VRWKV mixing mechanisms. The central claim rests on pretrained weights from the official VRWKV release, which are external. The main free parameters are experimental protocol choices, plus module hyperparameters that are not individually tuned, so I do not list them as major fitted values. The lack of explicit architectural details for the decoder is the largest unstated assumption.

free parameters (4)
  • Freeze epoch duration = 10 epochs
    The encoder is frozen for the first 10 epochs, then unfrozen. This number is chosen by hand and directly affects the measured ablation gains, but is not tuned or justified.
  • Early stopping threshold = 10 consecutive epochs without improvement
    Training stops if performance does not improve for 10 epochs. The paper says "test set", which means the threshold indirectly tunes reported numbers. This is a hand-set protocol parameter.
  • Learning rate and weight decay = 3e-4, 1e-4
    Standard optimizer settings chosen without reported tuning. They affect all models equally, but are still free parameters of the experimental protocol.
  • Image resolution = 512x512
    All images are resized to [512,512]. Resolution choice affects all models equally but is a hand-set protocol parameter.
assumptions (3)
  • domain assumption Pretrained VRWKV encoder weights from "upernet_vrwkv_adapter_tiny_512_160k_ade20k" and "upernet_vrwkv_adapter_small_512_160k_ade20k" are compatible with the medical decoder design.
    The entire Med-URWKV series relies on reusing the official VRWKV encoder as-is, with only the pretraining components discarded. No analysis of architectural mismatch or domain shift is provided.
  • standard math Wavelet transform and inverse wavelet transform are implemented as standard Haar wavelet operations with perfect reconstruction.
    The FAWA module uses WT and IWT. The precise implementation is not specified, and any edge effects or approximation could alter results.
  • domain assumption The "pure VRWKV decoder" faithfully maintains the VRWKV block structure from the pretrained encoder.
    The decoder uses VRWKV blocks, but the exact configuration (depth, dims, patch expand design) is only described via the figure, with no pseudocode or layer counts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Med-URWKV{\dag}: Toward Enhanced Pretrained Pure VRWKV Models for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/CBLSVB7S

@misc{pith2026250610858,
  author       = {Pith},
  title        = {Pith review of: Med-URWKV\dag: Toward Enhanced Pretrained Pure VRWKV Models for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CBLSVB7S}},
  note         = {Machine review of arXiv:2506.10858}
}
read the original abstract

Medical image segmentation is a fundamental task in computer-aided diagnosis and treatment. Existing approaches based on CNNs, ViTs, Mamba, and hybrid models still suffer from limitations such as restricted receptive fields, high computational cost, or insufficient accuracy. Recently, Vision Receptive-field Weighted Key-Value (VRWKV) models have emerged as a promising alternative,delivering strong long-range dependency modeling for visual tasks. However, current studies on VRWKV-based medical image segmentation mainly focus on hybrid architectures trained from scratch, while the potential of large-scale pretrained pure VRWKV models remains unexplored. In this work, we systematically investigate the effectiveness of pure VRWKV architectures for medical image segmentation. We construct Med-URWKV-T and Med-URWKV-S by reusing pretrained VRWKV encoders at different scales and pairing them with pure VRWKV decoders, enabling a comprehensive evaluation of pretrained pure VRWKV models in this domain. To further enhance performance, we propose two VRWKV-compatible modules: a Frequency-Aware Wavelet Attention (FAWA) module, which exploits wavelet transforms to capture edge details and structural characteristics, and a Multi-Scale Channel Fusion (MSCF) module, which integrates multi-scale features to strengthen informative channel representations. By incorporating them into Med-URWKV-T, we obtain the enhanced model Med-URWKV{\dag}. Extensive experiments on five medical image segmentation datasets demonstrate that Med-URWKV achieves performance comparable to or superior to state-of-the-art methods and carefully designed hybrid VRWKV architectures. Moreover, Med-URWKV{\dag} further improves segmentation accuracy, surpassing Med-URWKV-S while using only half of its parameter count, and achieves the highest average Dice similarity coefficient of 88.00%. The codes will be released.

Figures

Figures reproduced from arXiv: 2506.10858 by the authors.

Figure 1
Figure 1. Comparison of Med-URWKV† and other methods in terms of average DSC versus model parameter count. The area of each circle is proportional to the number of model parameters. As shown, the improved Med-URWKV† achieves a better balance between average segmentation performance and model complexity. of self-attention [43] and Vision Transformers (ViT) [11] alleviates this limitation, but their quadratic computational comp… view at source ↗
Figure 2
Figure 2. Conceptual illustrations of representative medical image segmentation architectures: (a) pure CNN; (b) pure ViT; (c) ViT encoder with CNN decoder; (d) hybrid CNN–ViT encoder with CNN decoder; (e) pure Mamba; (f) hybrid CNN–Mamba encoder with CNN decoder; (g) RWKV-based encoder with CNN decoder; and (h) pure RWKV (ours). and H2Former [20], which jointly leverage local feature extraction and global dependency modeling… view at source ↗
Figure 3
Figure 3. (a) The internal structure of VRWKV Block. (b) The process of Spatial Mix. (c) The process of Channel Mix. the receptive field of individual tokens and incorporates local inductive biases for subsequent attention mechanisms. Specifically, it executes four-directional spatial shifts on input tokens 𝑋, concatenating feature slices from adjacent tokens with the original representation: Q-Shift(∗)(𝑋) = 𝑋 + ( 1 − 𝜇(∗)) 𝑋… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Detailed components of the proposed framework: (a) Overview of the Med-URWKV† architecture. (b) The Med￾URWKV, comprising a pretrained VRWKV encoder, a Bottleneck block, and a pure VRWKV decoder. (c) Structure of the Multi-Scale Channel Fusion (MSCF) module. (d) Detail…
Figure 5
Figure 5. Figure 5: The process of first-order wavelet transform and inverse wavelet transform. Image taken from GLAS [41] image features in both the spatial and frequency domains simultaneously. Inspired by the concept of the Frequency Attention Triplet [22], we note that the WKV attenti…
Figure 6
Figure 6. Figure 6: Qualitative visual comparison of Med-URWKV-T and Med-URWKV† with other competing models. GT denotes the ground truth. White pixels inside the light red region and black pixels outside the region indicate correct predictions, while all other cases represent incorrect pr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 41 canonical work pages

  1. [1]

    Dataset of breast ultrasound images

    Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A., 2020. Dataset of breast ultrasound images. Data in brief 28, 104863

  2. [2]

    3D Densely Convolutional Networks for Volumetric Segmentation

    Bui,T.D.,Shin,J.,Moon,T.,2017.3ddenselyconvolutionalnetworks for volumetric segmentation. arXiv preprint arXiv:1709.03199

  3. [3]

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

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M., 2022. Swin-unet: Unet-like pure transformer for medical image segmentation,in:Europeanconferenceoncomputervision,Springer. pp. 205–218

  4. [5]

    Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images

    Chen, G., Li, L., Dai, Y., Zhang, J., Yap, M.H., 2022. Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Transactions on Medical Imaging 42, 1289–1300

  5. [6]

    Transunet:Transformersmakestrongencoders for medical image segmentation

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L.,Zhou,Y.,2021. Transunet:Transformersmakestrongencoders for medical image segmentation. arXiv preprint arXiv:2102.04306

  6. [7]

    Zig-rir: Zigzag rwkv-in-rwkv for efficient medicalimagesegmentation

    Chen,T.,Zhou,X.,Tan,Z.,Wu,Y.,Wang,Z.,Ye,Z.,Gong,T.,Chu, Q., Yu, N., Lu, L., 2025. Zig-rir: Zigzag rwkv-in-rwkv for efficient medicalimagesegmentation. IEEETransactionsonMedicalImaging

  7. [8]

    Çiçek,Ö.,Abdulkadir,A.,Lienkamp,S.S.,Brox,T.,Ronneberger,O.,

  8. [9]

    Codella,N.C.,Gutman,D.,Celebi,M.E.,Helba,B.,Marchetti,M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., et al.,

Show all 68 references
  1. [10]

    Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conferenceoncomputervisionandpatternrecognition,Ieee.pp.248– 255

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L., 2009. Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conferenceoncomputervisionandpatternrecognition,Ieee.pp.248– 255

  2. [11]

    Animageisworth16x16words:Transformersforimage recognition at scale

    Dosovitskiy,A.,Beyer,L.,Kolesnikov,A.,Weissenborn,D.,Zhai,X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., etal.,2020. Animageisworth16x16words:Transformersforimage recognition at scale. arXiv preprint arXiv:2010.11929

  3. [12]

    Vision-rwkv: Efficient and scal- able visual perception with rwkv-like architectures

    Duan, Y., Wang, W., Chen, Z., Zhu, X., Lu, L., Lu, T., Qiao, Y., Li, H., Dai, J., Wang, W., 2024. Vision-rwkv: Efficient and scal- able visual perception with rwkv-like architectures. arXiv preprint arXiv:2403.02308

  4. [13]

    Y-net: A spatiospectral dual-encoder network for medical image segmenta- tion, in: International conference on medical image computing and computer-assisted intervention, Springer

    Farshad, A., Yeganeh, Y., Gehlbach, P., Navab, N., 2022. Y-net: A spatiospectral dual-encoder network for medical image segmenta- tion, in: International conference on medical image computing and computer-assisted intervention, Springer. pp. 582–592

  5. [14]

    Feature extraction technique using discrete wavelet transform for image classification, in: 2007 5th Student Conference on Research and Development, IEEE

    Ghazali, K.H., Mansor, M.F., Mustafa, M.M., Hussain, A., 2007. Feature extraction technique using discrete wavelet transform for image classification, in: 2007 5th Student Conference on Research and Development, IEEE. pp. 1–4

  6. [15]

    Gong, H., Kang, L., Wang, Y., Wang, Y., Wan, X., Wu, X., Li, H.,

  7. [16]

    Mamba: Linear-time sequence modeling with selective state spaces

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

  8. [17]

    Unetr: Transformers for 3d medicalimagesegmentation,in:ProceedingsoftheIEEE/CVFwinter conference on applications of computer vision, pp

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D., 2022. Unetr: Transformers for 3d medicalimagesegmentation,in:ProceedingsoftheIEEE/CVFwinter conference on applications of computer vision, pp. 574–584

  9. [18]

    He, A., Li, T., Wu, Y., Zou, K., Fu, H., 2024. Frcnet: Frequency andregionconsistencyforsemi-supervisedmedicalimagesegmenta- tion, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 305–315

  10. [19]

    Progressive multiscale consistent network for multiclass fundus lesion segmenta- tion

    He,A.,Wang,K.,Li,T.,Bo,W.,Kang,H.,Fu,H.,2022. Progressive multiscale consistent network for multiclass fundus lesion segmenta- tion. IEEE transactions on medical imaging 41, 3146–3157

  11. [20]

    H2former:An efficienthierarchicalhybridtransformerformedicalimagesegmenta- tion

    He,A.,Wang,K.,Li,T.,Du,C.,Xia,S.,Fu,H.,2023. H2former:An efficienthierarchicalhybridtransformerformedicalimagesegmenta- tion. IEEE Transactions on Medical Imaging 42, 2763–2775

  12. [21]

    He, Z., Tang, J., Zhao, Z., Gong, Z., 2025. Rwkvmatch: Vision rwkv-based multi-scale feature matching network for unsupervised deformablemedicalimageregistration,in:ICASSP2025-2025IEEE InternationalConferenceonAcoustics,SpeechandSignalProcessing (ICASSP), IEEE. pp. 1–5

  13. [22]

    Wavelet-assisted multi-frequency attention network for pansharpen- ing, in: Proceedings of the AAAI Conference on Artificial Intelli- gence, pp

    Huang, J., Huang, R., Xu, J., Peng, S., Duan, Y., Deng, L.J., 2025. Wavelet-assisted multi-frequency attention network for pansharpen- ing, in: Proceedings of the AAAI Conference on Artificial Intelli- gence, pp. 3662–3670

  14. [23]

    Missformer: An effective medical image segmentation transformer

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

  15. [24]

    Polyp2former: Boundary guided network based on transformer for polyp segmentation, in: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE

    Huang, X., Huang, J., Wang, S., Wei, Y., An, D., Liu, J., 2023. Polyp2former: Boundary guided network based on transformer for polyp segmentation, in: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE. pp. 1971–1976

  16. [25]

    Acc-unet: A completely convolutional unet model for the 2020s, in: International conference on medical image computing and computer-assisted intervention, Springer

    Ibtehaz, N., Kihara, D., 2023. Acc-unet: A completely convolutional unet model for the 2020s, in: International conference on medical image computing and computer-assisted intervention, Springer. pp. 692–702

  17. [26]

    Kvasir-seg: A segmented polyp dataset, in: International conference on multimedia modeling, Springer

    Jha, D., Smedsrud, P.H., Riegler, M.A., Halvorsen, P., De Lange, T., Johansen, D., Johansen, H.D., 2019. Kvasir-seg: A segmented polyp dataset, in: International conference on multimedia modeling, Springer. pp. 451–462

  18. [27]

    Rwkv-unet: Improving unet with long-range cooperationforeffectivemedicalimagesegmentation

    Jiang, J., Zhang, J., Liu, W., Gao, M., Hu, X., Yan, X., Huang, F., Liu, Y., 2025. Rwkv-unet: Improving unet with long-range cooperationforeffectivemedicalimagesegmentation. arXivpreprint arXiv:2501.08458

  19. [28]

    Mixunet: Mix the 2d and 3d models for robust medical image segmentation, in: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE

    Li, J., Chen, S., Ma, S., Guo, F., Tang, J., 2023a. Mixunet: Mix the 2d and 3d models for robust medical image segmentation, in: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE. pp. 1242–1247

  20. [29]

    H-denseunet: hybrid densely connected unet for liver and tumor segmentationfromctvolumes.IEEEtransactionsonmedicalimaging 37, 2663–2674

    Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.W., Heng, P.A., 2018. H-denseunet: hybrid densely connected unet for liver and tumor segmentationfromctvolumes.IEEEtransactionsonmedicalimaging 37, 2663–2674

  21. [30]

    Can: Context-assisted full attention network for brain tissue segmentation

    Li, Z., Zhang, C., Zhang, Y., Wang, X., Ma, X., Zhang, H., Wu, S., 2023b. Can: Context-assisted full attention network for brain tissue segmentation. Medical Image Analysis 85, 102710

  22. [31]

    Lightm-unet: Mamba assists in lightweight unet for medical image segmentation

    Liao, W., Zhu, Y., Wang, X., Pan, C., Wang, Y., Ma, L., 2024. Lightm-unet: Mamba assists in lightweight unet for medical image segmentation. arXiv preprint arXiv:2403.05246

  23. [32]

    IEEE Transactions on Medical Imaging

    Liu, J., Yang, H., Zhou, H.Y., Yu, L., Liang, Y., Yu, Y., Zhang, S., Zheng,H.,Wang,S.,2024a.Swin-umamba†:Adaptingmamba-based vision foundation models for medical image segmentation. IEEE Transactions on Medical Imaging

  24. [33]

    Vmamba: Visual state space model

    Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Jiao, J., Liu, Y., 2024b. Vmamba: Visual state space model. Advances in neural information processing systems 37, 103031–103063

  25. [34]

    V-net: Fully convolu- tional neural networks for volumetric medical image segmentation, in: 2016 fourth international conference on 3D vision (3DV), Ieee

    Milletari, F., Navab, N., Ahmadi, S.A., 2016. V-net: Fully convolu- tional neural networks for volumetric medical image segmentation, in: 2016 fourth international conference on 3D vision (3DV), Ieee. pp. 565–571

  26. [35]

    Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Mis- awa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.,

  27. [36]

    Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M., Grella, M., et al.,

  28. [37]

    Qiao, S., Zhao, Z., Xie, P., Yin, W., Wang, M., Zhang, Y., Pang, S.,

  29. [38]

    Rahman, M.M., Munir, M., Marculescu, R., 2024. Emcad: Efficient multi-scale convolutional attention decoding for medical image seg- mentation, in: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pp. 11769–11779

  30. [39]

    arXiv Zhou et al.:Preprint submitted to ElsevierPage 10 of 11 Med-URWKV†: Toward Enhanced Pretrained Pure VRWKV Models for Medical Image Segmentation preprint arXiv:1804.03999

    Attentionu-net:Learningwheretolookforthepancreas. arXiv Zhou et al.:Preprint submitted to ElsevierPage 10 of 11 Med-URWKV†: Toward Enhanced Pretrained Pure VRWKV Models for Medical Image Segmentation preprint arXiv:1804.03999

  31. [40]

    Vm-unet: Vision mamba unet for medical image segmentation

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

  32. [41]

    Sirinukunwattana,K., Pluim,J.P.,Chen,H., Qi,X.,Heng, P.A.,Guo, Y.B., Wang, L.Y., Matuszewski, B.J., Bruni, E., Sanchez, U., et al.,

  33. [42]

    Unext: Mlp-based rapid medical image segmentation network, in: International conference on medical image computing and computer-assisted intervention, Springer

    Valanarasu, J.M.J., Patel, V.M., 2022. Unext: Mlp-based rapid medical image segmentation network, in: International conference on medical image computing and computer-assisted intervention, Springer. pp. 23–33

  34. [43]

    Attention is all you need

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. Advances in neural information processing systems 30

  35. [44]

    Uctransnet:rethink- ing the skip connections in u-net from a channel-wise perspective withtransformer,in:ProceedingsoftheAAAIconferenceonartificial intelligence, pp

    Wang,H.,Cao,P.,Wang,J.,Zaiane,O.R.,2022. Uctransnet:rethink- ing the skip connections in u-net from a channel-wise perspective withtransformer,in:ProceedingsoftheAAAIconferenceonartificial intelligence, pp. 2441–2449

  36. [45]

    Ronneberger, O., Fischer, P., Brox, T., 2015. 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, pro- ceedings, part III...

  37. [46]

    Large window-based mamba unet for medical image segmentation: Beyond convolution and self-attention

    Wang, J., Chen, J., Chen, D., Wu, J., 2024. Large window-based mamba unet for medical image segmentation: Beyond convolution and self-attention. CoRR

  38. [47]

    Set: Superpixel embedded transformer for skin lesion segmentation

    Wang, Z., Lyu, J., Tang, X., 2025b. Set: Superpixel embedded transformer for skin lesion segmentation. Medical Image Analysis , 103738

  39. [48]

    H-vmunet: High-order vision mamba unet for medical image segmentation

    Wu, R., Liu, Y., Liang, P., Chang, Q., . H-vmunet: High-order vision mamba unet for medical image segmentation. arxiv 2024. arXiv preprint arXiv:2403.13642

  40. [49]

    Mgfuseseg: Attention-guidedmulti-granularityfusionformedicalimagesegmen- tation,in:2023IEEEInternationalConferenceonBioinformaticsand Biomedicine (BIBM), IEEE

    Xu, G., Leng, X., Li, C., He, X., Wu, X., 2023. Mgfuseseg: Attention-guidedmulti-granularityfusionformedicalimagesegmen- tation,in:2023IEEEInternationalConferenceonBioinformaticsand Biomedicine (BIBM), IEEE. pp. 3587–3594

  41. [50]

    Smnet: A semantic guided mamba network for remote sensing change detec- tion

    Xu, G., Liu, Y., Deng, L., Wang, X., Zhu, H., 2025a. Smnet: A semantic guided mamba network for remote sensing change detec- tion. IEEE Transactions on Aerospace and Electronic Systems , 1– 12doi:10.1109/TAES.2025.3580691

  42. [51]

    Urwkv: Unified rwkv model with multi-state perspective for low-light image restoration, in: Proceedings of the Computer Vision and Pattern Recognition Conference, pp

    Xu, R., Niu, Y., Li, Y., Xu, H., Liu, W., Chen, Y., 2025b. Urwkv: Unified rwkv model with multi-state perspective for low-light image restoration, in: Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 21267–21276

  43. [52]

    arXiv preprint arXiv:2506.04956

    Wang,H.,Yang,Z.,Zhang,H.,Zhao,D.,Wei,B.,Xu,Y.,2025a.Feat: Full-dimensional efficient attention transformer for medical video generation. arXiv preprint arXiv:2506.04956

  44. [53]

    Yuan, H., Li, X., Qi, L., Zhang, T., Yang, M.H., Yan, S., Loy, C.C.,

  45. [54]

    An attention free transformer

    Zhai, S., Talbott, W., Srivastava, N., Huang, C., Goh, H., Zhang, R., Susskind, J., 2021. An attention free transformer. arXiv preprint arXiv:2105.14103

  46. [55]

    Out-of-distribution semantic occupancy prediction

    Zhang, Y., Duan, M., Peng, K., Wang, Y., Liu, R., Teng, F., Luo, K., Li, Z., Yang, K., 2025. Out-of-distribution semantic occupancy prediction. arXiv preprint arXiv:2506.21185

  47. [56]

    Smaformer: Synergistic multi-attention transformer for medicalimagesegmentation,in:2024IEEEInternationalConference onBioinformaticsandBiomedicine(BIBM),pp.4048–4053

    Zheng,F.,Chen,X.,Liu,W.,Li,H.,Lei,Y.,He,J.,Pun,C.M.,Zhou, S., 2024. Smaformer: Synergistic multi-attention transformer for medicalimagesegmentation,in:2024IEEEInternationalConference onBioinformaticsandBiomedicine(BIBM),pp.4048–4053. doi:10. 1109/BIBM62325.2024.10822736

  48. [57]

    Bsbp-rwkv: Background suppression with boundary preservation for efficient medical image segmentation, in: Proceedingsofthe32ndACMInternationalConferenceonMultime- dia, pp

    Zhou, X., Chen, T., 2024. Bsbp-rwkv: Background suppression with boundary preservation for efficient medical image segmentation, in: Proceedingsofthe32ndACMInternationalConferenceonMultime- dia, pp. 4938–4946

  49. [58]

    Zhou, Y., Huang, J., Wang, C., Song, L., Yang, G., 2023. Xnet: Wavelet-based low and high frequency fusion networks for fully- andsemi-supervisedsemanticsegmentationofbiomedicalimages,in: Proceedings of the IEEE/CVF international conference on computer vision, pp. 21085–21096

  50. [59]

    Ye, Z., Chen, T., Wang, Z., Zhang, H., Zhang, L., 2025. Hfe-rwkv: High-frequency enhanced rwkv model for efficient left ventricle seg- mentationinpediatricechocardiograms,in:ICASSP2025-2025IEEE InternationalConferenceonAcoustics,SpeechandSignalProcessing (ICASSP), IEEE. pp. 1–5

  51. [60]

    Unet++: Redesigning skip connections to exploit multiscale features in image segmentation

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J., 2019. Unet++: Redesigning skip connections to exploit multiscale features in image segmentation. IEEE transactions on medical imaging 39, 1856–1867

  52. [61]

    arXiv preprint arXiv:2406.19369

    Mamba or rwkv: Exploring high-quality and high-efficiency segment anything model. arXiv preprint arXiv:2406.19369

  53. [67]

    Spatial- frequency dual domain attention network for medical image segmen- tation,in:2024IEEEInternationalConferenceonBioinformaticsand Biomedicine (BIBM), IEEE

    Zhou, Z., He, A., Wu, Y., Yao, R., Xie, X., Li, T., 2024. Spatial- frequency dual domain attention network for medical image segmen- tation,in:2024IEEEInternationalConferenceonBioinformaticsand Biomedicine (BIBM), IEEE. pp. 4076–4081

  54. [69]

    Personalizable long- context symbolic music infilling with midi-rwkv

    Zhou-Zheng, C., Pasquier, P., 2025. Personalizable long- context symbolic music infilling with midi-rwkv. arXiv preprint arXiv:2506.13001 . Zhou et al.:Preprint submitted to ElsevierPage 11 of 11

  55. [2016]

    3d u-net: learning dense volumetric segmentation from sparse annotation,in:Internationalconferenceonmedicalimagecomputing and computer-assisted intervention, Springer. pp. 424–432

  56. [2017]

    Medical image analysis 35, 489–502

    Gland segmentation in colon histology images: The glas challenge contest. Medical image analysis 35, 489–502

  57. [2018]

    Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hostedbytheinternationalskinimagingcollaboration(isic),in:2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), IEEE. pp. 168–172

  58. [2023]

    arXivpreprint arXiv:2305.13048

    Rwkv:Reinventingrnnsforthetransformerera. arXivpreprint arXiv:2305.13048

  59. [2024]

    4261–4268

    A novel conv-mamba-hybrid network for medical image seg- mentation,in:2024IEEEInternationalConferenceonBioinformatics and Biomedicine (BIBM), pp. 4261–4268. doi:10.1109/BIBM62325. 2024.10821714

  60. [2025]

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

Pith tools

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