Pith. sign in

REVIEW 3 major objections 4 minor 31 references

U-RWKV: Lightweight medical image segmentation with direction-adaptive RWKV

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 U-RWKV, a 2.97M-parameter U-shaped network with direction-adaptive RWKV scans, reaches an average Dice of 82.27 across five medical segmentation benchmarks at linear computational cost.

desk verdict Architecture is plausible and the ablations are useful, but the paper's own Table 1 contradicts its headline SOTA claim, so the numbers need reconciliation before the claim can stand. read the letter →

arxiv 2507.11415 v1 pith:H4XRKNKE submitted 2025-07-15 eess.IV cs.AIcs.CV

Hongbo Ye , Fenghe Tang , Peiang Zhao , Zhen Huang , Dexin Zhao , Minghao Bian , S.Kevin Zhou This is my paper
classification eess.IVcs.AIcs.CV
keywords U-RWKVmedicalimagesegmentationRWKVeffectivereceptivefieldlightweightdirectionalscanningsqueeze-and-excitationlinear-complexityattention
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 U-RWKV, a lightweight encoder-decoder for medical image segmentation built on the RWKV recurrent architecture instead of quadratic-attention transformers. The central claim is that its two modules, DARM and SASE, together give a global effective receptive field at $O(N)$ computational cost while preserving local detail. On five public benchmarks (BUSI, Kvasir, ClinicDB, ISIC 2017, and ISIC 2018), the paper reports an average Dice of 82.27 with 2.97M parameters and 7.28 GFLOPs, plus 80.64 Dice on Synapse, saying this surpasses most existing methods at a fraction of their size. If the claim holds, resource-limited clinical settings could run global-context segmentation without large transformer models.

What carries the argument

Two interlocking modules carry the argument. DARM (Direction-Adaptive RWKV Module) expands a feature map into four one-dimensional sequences by scanning left-to-right, right-to-left, top-to-bottom, and bottom-to-top (QuadScan). Each sequence passes through a spatial mixer using the RWKV WKV recurrence, then the four outputs are reshaped and averaged pixel-wise, and a channel mixer runs over the flattened result; a Dual-RWKV pass repeats the procedure in forward and reverse order without shared weights. This propagates context from all image edges through a compressed hidden state at $O(N)$ cost. SASE (Stage-Adaptive Squeeze-and-Excitation Module) changes structure by feature stage, using dilated inverted bottlenecks in early high-resolution stages and channel-split, depthwise-separable bottlenecks in deeper low-resolution stages, balancing detail preservation with semantic capture.

What would settle it

Re-run U-Net, TransUNet, CMUNeXt, UNeXt, and the other listed baselines on the same 70/30 splits and 280-epoch schedule used for U-RWKV, then compare average Dice and FLOPs; if U-RWKV no longer holds the top average Dice or its efficiency advantage, the central comparison collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that a U-shaped segmentation network can replace quadratic attention with RWKV-style linear recurrence and still beat established baselines. DARM scans each feature map in four directions, applies dual RWKV spatial mixing in both forward and reverse order, and averages the four reconstructed maps; SASE changes its squeeze-and-excitation structure by stage. The paper reports that this yields a high-contribution effective receptive field ratio of 0.992 at a 0.99 threshold, compared with 0.568 for U-Net and 0.946 for ViT- or Mamba-based variants, and that the full model reaches the highest average Dice of 82.27 on the five datasets it evaluates.

Load-bearing premise

The claim that U-RWKV surpasses most existing methods rests on the baseline scores in Tables 1 and 2 being measured under the same data splits, preprocessing, resolution, and training budget as U-RWKV's runs, since the paper reports those baselines from the literature rather than re-running them.

Editorial extensions

If this is right

  • If the central claim holds, a model with about 3 million parameters can maintain a global effective receptive field, making global-context segmentation feasible on hardware without large GPU memory.
  • The ablation numbers imply that both modules matter: dropping DARM lowers average IoU from 77.62 to 74.97, and dropping SASE lowers it to 75.29.
  • Combining opposite scan directions improves over any single direction, supporting the claim that the four-direction design reduces directional bias.
  • The reported high-contribution ERF ratio of 0.992 at a 0.99 threshold, versus 0.568 for U-Net, indicates predictions depend on inputs across nearly the whole image rather than only local texture.
  • The paper states that extending the design to 3D segmentation is future work; if the $O(N)$ cost carries over, volumetric medical images would be a natural next target.

Reading between the lines

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

  • Inference: DARM is a generic feature-refinement block, so the same direction-adaptive scan could be dropped into other U-shaped encoders for different dense prediction tasks.
  • Inference: the efficiency comparison is based on parameters and FLOPs, not wall-clock latency; the paper itself notes inference is slower than pure CNNs such as UNeXt, so the practical benefit is cost-accuracy balance rather than raw speed.
  • Inference: if the baseline numbers in Tables 1 and 2 were not produced under identical protocols, the reported ranking could shift; re-running the baselines under the paper's schedule would settle the comparison.
  • Inference: a direct test of the $O(N)$ claim would be scaling input resolution to 1024 by 1024 and checking whether memory and runtime grow roughly linearly while Dice is maintained.
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 U-RWKV, a lightweight U-shaped network for medical image segmentation that replaces standard self-attention with RWKV-style linear-complexity token mixing. The two architectural contributions are Direction-Adaptive RWKV (DARM), which combines Dual-RWKV and QuadScan directional scanning, and Stage-Adaptive Squeeze-and-Excitation (SASE), which changes channel-expansion behavior across encoder/decoder stages. The authors report a 2.97M-parameter model with 7.28 GFLOPs and claim state-of-the-art segmentation performance, specifically the highest average Dice of 82.27 on five datasets, plus results on Synapse and ablation studies supporting the direction-scanning and ERF claims.

Significance. If the empirical claims hold, this is a potentially useful contribution: the model is lightweight, the code is released, and the idea of combining RWKV-style linear attention with multiple directional scans is worth studying. The paper also reports experiments across several medical segmentation datasets and includes an ERF analysis. However, the central state-of-the-art claim is currently contradicted by the paper's own tables, and the absence of statistical or protocol details makes the reported rankings difficult to trust. The architectural novelty is plausible but the evidence presented for it is not yet reliable.

major comments (3)
  1. [Section 3.2, Table 1] The Average column in Table 1 is not reproducible from the per-dataset Dice scores and contradicts the central claim. For U-RWKV, the five displayed scores (82.34, 88.17, 90.58, 90.13, 87.26) average to 87.70, not the printed 82.27; for U-Net the printed Average is 81.48 while the displayed values average to 87.01; for CMU-Net the printed Average is 83.06 while the displayed values average to 88.01. No weighting rule, dataset-size adjustment, or per-sample statistic is given, so no stated aggregation reproduces the column. The abstract and Section 3.2 claim that U-RWKV attains the highest average Dice of 82.27, but CMU-Net's printed Average of 83.06 already exceeds it, and if the Average is the unweighted mean of the displayed values then CMU-Net's 88.01 also exceeds U-RWKV's 87.70. The state-of-the-art claim is therefore unsupported by the paper's own data and must be corrected by recomputing all summary statistics and adjusting the claim accordingly.
  2. [Section 3.3, Table 3] Table 3 has the same internal inconsistency in its summary column, and this directly affects the ablative argument. For the first eight rows, the Average column does not equal the arithmetic mean of the three displayed IoU values: Left-to-Right (69.61, 77.91, 81.65) averages to 76.39, not the printed 76.75, and Top-to-Bottom (69.04, 78.87, 82.23) averages to 76.71, not the printed 77.02. In addition, the sentence 'baseline results show that combining multi-directional scans improves IoU' is not supported even by the printed summaries: L-to-R combined with R-to-L (76.64) is below L-to-R alone (76.75), and T-to-B combined with B-to-T (76.98) is below T-to-B alone (77.02). The authors should correct the summary statistics and support the directional-bias claim with standard deviations or paired tests rather than asserting a pattern that the table does not show.
  3. [Section 3.2 and Section 3.1] The comparison against prior methods is not established under a common protocol. The manuscript says training follows the settings in references [22,25] with several modifications, but it does not state whether all baseline models in Tables 1 and 2 were re-run under the same data splits, preprocessing, input resolution, training budget, and evaluation code, or whether the baseline numbers are imported from their original papers. Several of the reported differences between top methods are around 1% or less, and no standard deviations, confidence intervals, or significance tests are reported. Without a shared evaluation protocol or variance estimates, the ranking of U-RWKV against CMU-Net, CMUNeXt, and MedT in Table 1 and against ConvUNeXt and CMUNeXt in Table 2 cannot be assessed.
minor comments (4)
  1. [Section 3.2] In the sentence 'attaining the highest average Dice score of82.27', there is a missing space between 'of' and the number; this and similar spacing issues elsewhere should be fixed.
  2. [Throughout] The model name appears inconsistently as 'U-R WKV' and 'U-RWKV' in the abstract and body; one form should be used consistently.
  3. [Fig. 1] The caption and text state that Fig. 1(a) shows U-RWKV achieving the highest average Dice, but Table 1 does not support this; the figure must be regenerated from corrected statistics.
  4. [Section 4] The conclusion says inference speed is slower than CNNs like UNeXt, but no runtime measurements are reported; reporting inference time would strengthen the efficiency claim made in the abstract.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: U-RWKV is an empirically validated architecture study; self-citations are non-load-bearing.

full rationale

The paper's derivation chain is self-contained and empirical. U-RWKV is constructed by combining the publicly defined RWKV spatial/channel mixing operators (Eq. 1, explicitly credited to Vision-RWKV [9]) with two new modules, DARM (Algorithm 1, Eq. 3) and SASE (Sec. 2.2), both defined by the authors' own equations. There is no step in which a fitted parameter is renamed as a prediction, no uniqueness theorem imported from self-citations, and no ansatz smuggled in via citation: the RWKV operators are borrowed openly, and the architectural choices are evaluated on external benchmarks (BUSI, Kvasir, ClinicDB, ISIC 2017/2018, Synapse) through Dice/IoU and measured ERF. The ablations (Table 3) provide independent evidence for the contributions of DARM and SASE. The authors' self-citations (e.g., [22,25] for training settings and as baselines) are procedural/comparative and not load-bearing for any mathematical derivation. One non-circular correctness concern must be flagged: the printed Avg column in Table 1 is internally inconsistent (U-RWKV's Avg 82.27 is lower than its lowest per-dataset Dice of 82.34, so no weighted average of the displayed scores can produce it; CMU-Net's 83.06 also exceeds it), which undermines the 'highest average Dice' headline; this is a data-integrity issue affecting correctness, not circularity of the method's derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its empirical claims rely on transferability of pretrained RWKV weights, comparability of externally reported baselines, and hand-set architecture hyperparameters such as SASE expansion ratios, channel split count, patch size, and the four scan directions.

free parameters (4)
  • SASE shallow-mode channel expansion ratio = 4
    Set by hand in Section 2.2; the module uses four times more channels in early stages. No ablation over other ratios is shown.
  • SASE deep-mode channel split count = 8
    Set by hand in Section 2.2; channels are split into 8 parts before doubling. No ablation over split counts is shown.
  • DARM patch size = 1
    Set in Section 2.3 so that sequence length equals H x W, which depends on encoder downsampling choices.
  • Number of QuadScan directions = 4
    Chosen in Section 2.3 as left-to-right, right-to-left, top-to-bottom, and bottom-to-top. The paper ablates pairs of directions but not other direction sets.
assumptions (5)
  • standard math RWKV's WKV operation provides linear-complexity long-range mixing as defined in references [18] and [9].
    Section 2.3 relies on the recurrence and complexity claims from RWKV and Vision-RWKV without re-deriving them.
  • domain assumption Pretrained Vision-RWKV weights [9] transfer to medical imaging features.
    Section 3.1 initializes the RWKV model with weights from [9]; no analysis of transfer quality is provided.
  • domain assumption Baseline scores from earlier papers are comparable with the paper's own data splits and preprocessing.
    Section 3.2 and Tables 1 and 2 compare U-RWKV against literature values without stating whether all models were retrained under a shared protocol.
  • ad hoc to paper The high-contribution area ratio at the 0.99 threshold is a valid summary of effective receptive field and segmentation quality.
    Section 3.3 uses Figure 1(b) as evidence, but the map construction and threshold rationale are not described.
  • ad hoc to paper Four fixed scan directions plus forward and backward sequence processing eliminate directional bias sufficiently to improve segmentation.
    Sections 2.3 and 3.3 claim directional-bias mitigation without a formal argument or a systematic search over scanning configurations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of U-RWKV: Lightweight medical image segmentation with direction-adaptive RWKV." pith.science (2026). https://pith.science/paper/H4XRKNKE

@misc{pith2026250711415,
  author       = {Pith},
  title        = {Pith review of: U-RWKV: Lightweight medical image segmentation with direction-adaptive RWKV},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4XRKNKE}},
  note         = {Machine review of arXiv:2507.11415}
}
read the original abstract

Achieving equity in healthcare accessibility requires lightweight yet high-performance solutions for medical image segmentation, particularly in resource-limited settings. Existing methods like U-Net and its variants often suffer from limited global Effective Receptive Fields (ERFs), hindering their ability to capture long-range dependencies. To address this, we propose U-RWKV, a novel framework leveraging the Recurrent Weighted Key-Value(RWKV) architecture, which achieves efficient long-range modeling at O(N) computational cost. The framework introduces two key innovations: the Direction-Adaptive RWKV Module(DARM) and the Stage-Adaptive Squeeze-and-Excitation Module(SASE). DARM employs Dual-RWKV and QuadScan mechanisms to aggregate contextual cues across images, mitigating directional bias while preserving global context and maintaining high computational efficiency. SASE dynamically adapts its architecture to different feature extraction stages, balancing high-resolution detail preservation and semantic relationship capture. Experiments demonstrate that U-RWKV achieves state-of-the-art segmentation performance with high computational efficiency, offering a practical solution for democratizing advanced medical imaging technologies in resource-constrained environments. The code is available at https://github.com/hbyecoding/U-RWKV.

Figures

Figures reproduced from arXiv: 2507.11415 by the authors.

Figure 1
Figure 1. Performance comparison and Effective Receptive Field (ERF). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed U-RWKV; SASE Module and DARM [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of segmentation results on four datasets: CVC-ClinicDB, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 16 canonical work pages

  1. [1]

    Data in brief28, 104863 (2020)

    Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

    2020

  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]

    In: European conference on computer vision

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205–218. Springer (2022)

    2022

  4. [4]

    arXiv preprint arXiv:2102.04306 (2021)

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.:Transunet:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021)

    arXiv 2021

  5. [5]

    In: MICCAI

    Chen, J., Chen, R., Wang, W., Cheng, J., Zhang, L., Chen, L.: Tinyu-net: Lighter yet better u-net with cascaded multi-receptive fields. In: MICCAI. pp. 626–635. Springer (2024)

    work page 2024

  6. [6]

    arXiv preprint arXiv:1902.03368 (2019)

    Codella, N., Rotemberg, V., Tschandl, P., Celebi, M.E., Dusza, S., Gutman, D., Helba,B.,Kalloo,A.,Liopyris,K., Marchetti,M.,et al.:Skinlesionanalysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368 (2019)

    arXiv 2019

  7. [7]

    In: CVPR

    Ding, X., Zhang, X., Han, J., Ding, G.: Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In: CVPR. pp. 11963–11975 (2022)

    work page 2022

  8. [8]

    arXiv preprint arXiv:2010.11929 (2020) 10 H

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 10 H. Ye et al

    arXiv 2020

Show all 31 references
  1. [9]

    arXiv preprint arXiv:2403.02308 (2024)

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

    2024 arXiv

  2. [10]

    Journal of Imaging 6(7), 69 (2020)

    Guo, Y., Bernal, J., J Matuszewski, B.: Polyp Segmentation with Fully Convo- lutional Deep Neural Networks—Extended Evaluation Study. Journal of Imaging 6(7), 69 (2020)

    2020

  3. [11]

    Knowledge-based systems253, 109512 (2022)

    Han, Z., Jian, M., Wang, G.G.: Convunext: An efficient convolution neural network for medical image segmentation. Knowledge-based systems253, 109512 (2022)

    2022

  4. [12]

    In: ICASSP

    Huang,H.,Lin,L.,Tong,R.,Hu,H.,Zhang,Q.,Iwamoto,Y.,Han,X.,Chen,Y.W., Wu, J.: Unet 3+: A full-scale connected unet for medical image segmentation. In: ICASSP. pp. 1055–1059. IEEE (2020)

    2020

  5. [13]

    IJCAS 19(5), 939–950 (2024)

    Huang, Z., Li, H., Shao, S., Zhu, H., Hu, H., Cheng, Z., Wang, J., Kevin Zhou, S.: Pele scores: pelvic x-ray landmark detection with pelvis extraction and enhance- ment. IJCAS 19(5), 939–950 (2024)

    2024

  6. [14]

    In: CBMS (2020)

    Jha, D., Riegler, M., Johansen, D., Halvorsen, P., Johansen, H.: DoubleU-Net: A Deep Convolutional Neural Network for Medical Image Segmentation. In: CBMS (2020)

    2020

  7. [15]

    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 Segmentation. In: ISM. pp. 225–230 (2019)

    2019

  8. [16]

    NeurIPS29 (2016)

    Luo, W., Li, Y., Urtasun, R., Zemel, R.: Understanding the effective receptive field in deep convolutional neural networks. NeurIPS29 (2016)

    2016

  9. [17]

    arXiv preprint arXiv:1804.03999 (2018)

    Oktay, O.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)

    2018 arXiv

  10. [18]

    In: EMNLP (2023)

    Peng, B., Alcaide, E., Anthony, Q.G., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M.N., Derczynski, L., et al.: Rwkv: Reinventing rnns for the transformer era. In: EMNLP (2023)

    2023

  11. [19]

    NPJ digital medicine5(1), 119 (2022)

    Richardson, S., Lawrence, K., Schoenthaler, A.M., Mann, D.: A framework for digital health equity. NPJ digital medicine5(1), 119 (2022)

    2022

  12. [20]

    In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

    2015

  13. [21]

    Ruan, J., Li, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image seg- mentation (2024), https://arxiv.org/abs/2402.02491

    2024 arXiv

  14. [22]

    In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI)

    Tang, F., Ding, J., Quan, Q., Wang, L., Ning, C., Zhou, S.K.: Cmunext: An efficient medical image segmentation network based on large kernel and skip fusion. In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2024)

    2024

  15. [23]

    arXiv preprint arXiv:2312.01740 (2023)

    Tang, F., Nian, B., Ding, J., Quan, Q., Yang, J., Liu, W., Zhou, S.K.: Mobileutr: Revisiting the relationship between light-weight cnn and transformer for efficient medical image segmentation. arXiv preprint arXiv:2312.01740 (2023)

    2023 arXiv

  16. [24]

    Medical Image Analysis p

    Tang, F., Nian, B., Li, Y., Jiang, Z., Yang, J., Liu, W., Zhou, S.K.: Mambamim: Pre-training mamba with state space token interpolation and its application to medical image segmentation. Medical Image Analysis p. 103606 (2025)

    2025

  17. [25]

    In: ISBI

    Tang, F., Wang, L., Ning, C., Xian, M., Ding, J.: Cmu-net: a strong convmixer- based medical ultrasound image segmentation network. In: ISBI. pp. 1–5. IEEE (2023)

    2023

  18. [26]

    In: MICCAI

    Tang, F., Xu, R., Yao, Q., Fu, X., Quan, Q., Zhu, H., Liu, Z., Zhou, S.K.: Hyspark: Hybrid sparse masking for large scale medical image pre-training. In: MICCAI. pp. 330–340. Springer (2024) U-RWKV 11

    2024

  19. [27]

    arXiv preprint arXiv:2502.08347 (2025)

    Tang, F., Yao, Q., Ma, W., Wu, C., Jiang, Z., Zhou, S.K.: Hi-end-mae: Hierarchical encoder-driven masked autoencoders are stronger vision learners for medical image segmentation. arXiv preprint arXiv:2502.08347 (2025)

    2025 arXiv

  20. [28]

    In: MICCAI

    Valanarasu, J.M.J., Oza, P., Hacihaliloglu, I., Patel, V.M.: Medical transformer: Gated axial-attention for medical image segmentation. In: MICCAI. pp. 36–46 (2021)

    2021

  21. [29]

    In: MICCAI

    Valanarasu, J.M.J., Patel, V.M.: Unext: Mlp-based rapid medical image segmen- tation network. In: MICCAI. pp. 23–33. Springer (2022)

    2022

  22. [30]

    arXiv preprint arXiv:2412.06499 (2024)

    Zhou, X., Huang, Z., Zhu, H., Yao, Q., Zhou, S.K.: Hybrid attention net- work: An efficient approach for anatomy-free landmark detection. arXiv preprint arXiv:2412.06499 (2024)

    2024 arXiv

  23. [31]

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

    2018

Pith tools

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