Pith. sign in

REVIEW 5 major objections 5 minor 32 references

FreqU-FNet: Frequency-Aware U-Net for Imbalanced Medical Image Segmentation

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

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

Pith's one-line read A U-Net that filters and downsamples in the frequency domain claims to lift minority-class accuracy in medical image segmentation.

desk verdict FreqU-FNet's empirical core collapses under inspection—Table 3 duplicates Table 2, and Eq. (7) reduces to known operations; the motivation is fine but the paper is not ready for review. read the letter →

arxiv 2505.17544 v1 pith:OXAKDRQA submitted 2025-05-23 eess.IV cs.CV

Ruiqi Xing This is my paper
classification eess.IVcs.CV
keywords medicalimagesegmentationclassimbalancefrequencydomainlearningDaubechieswaveletFourierlow-passfilteringU-Netminorityadaptiveupsampling
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 argues that moving the encoder's downsampling and filtering into the frequency domain—using Daubechies wavelets plus a Fourier-domain low-pass box mask—lets a U-shaped network recover small, low-contrast structures that spatial pooling and self-attention tend to blur. It backs this with a decoder whose upsampling is a learned fusion of two sampling pathways, and a loss that adds an L1 term on wavelet high-frequency subbands plus Top-K cross-entropy to Dice. On MSD Prostate, Pancreas, and Lung, it reports the largest gains on the minority class: tumor Dice in pancreas rises to 63.38% versus 54.51% for Swin-UNet, and lung Dice to 80.77%. The intended payoff is a segmentation network that does not need Transformer-scale pretraining yet narrows the majority–minority accuracy gap.

What carries the argument

The load-bearing object is the frequency encoder's composite mapping $\Phi$, a parameter-free sequence of wavelet and Fourier transforms terminating in a low-pass box mask $M(\tau)$ that keeps only a central rectangle of the frequency plane. It is what the paper credits with suppressing aliasing and preserving minority-class frequency bands. Two companion components carry the rest: the Spatial Learnable Decoder (SLD), whose adaptive hierarchical multi-branch upsampling fuses a native-space dynamic-sampling path with a space-channel exchange path, and the frequency-aware loss (FAL), an L1 loss over Daubechies high-frequency subbands combined with multi-class Dice and average Top-K cross-entropy.

What would settle it

An experiment that computes per-class Fourier power spectra of the ground-truth masks on MSD Pancreas and MSD Lung: if the tumor-class energy concentrates in bands that the mask $M(\tau)$ zeroes out, or if sweeping $\tau$ over a grid never beats the reported Dice, the frequency-localization premise fails. A paired re-run of the pancreas comparison with matched training budgets and multiple seeds that shows no significant minority-class increase would also settle it.

Watch

Extended reading notes

Core claim

FreqU-FNet is presented as a U-shaped segmentation network whose encoder works in the frequency domain. Its central claim is that spatial pooling is the main source of aliasing and minority-class signal loss, and that replacing it with the composite mapping $\Phi(X) = \mathrm{WT}_{db}(\mathcal{F}^{-1}(\mathcal{F}(\mathrm{IWT}_{db}(\mathrm{WT}_{db}(X))) \odot M))$—Daubechies wavelet decomposition, inverse-wavelet reconstruction, Fourier low-pass filtering through a box mask $M(\tau)$, and a final wavelet downsampling—recovers the discriminative frequency bands of under-represented structures. The decoder then reconstructs fine detail through two adaptively fused upsampling pathways, and the loss adds an L1 penalty on wavelet high-frequency subbands to Dice and Top-K cross-entropy. On MSD Prostate, Pancreas, and Lung, the paper reports that this design raises tumor Dice in pancreas to 63.38% versus 54.51% for Swin-UNet, cuts the pancreas Dice gap to 15.64% against a 27–39% range for nnU-Net variants, and reaches 80.77% lung Dice.

Load-bearing premise

The minority-class structures in all three datasets are assumed to be separable in the low-frequency Fourier band preserved by the single fixed box mask $M(\tau)$, and the paper offers no spectral measurement showing that this holds.

Editorial extensions

If this is right

  • Spatial pooling can be replaced by a wavelet–Fourier composite downsampling in U-shaped encoders without losing resolution.
  • Minority-class Dice can be improved by roughly nine points over Swin-UNet on pancreas tumor without external pretraining.
  • A loss that supervises wavelet high-frequency subbands can complement Dice and Top-K cross-entropy for boundary fidelity.
  • The class-gap metric (max-minus-min Dice) becomes a practical target for evaluating imbalanced segmentation.

Reading between the lines

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

  • A cheap test of the core premise is to compute per-class power spectra of the ground-truth masks: if tumor-class energy falls inside the masked band, the fixed $\tau$ can be tuned per dataset rather than shared globally.
  • Because the frequency-aware loss is defined independently of the encoder, it could be grafted onto existing CNN or Transformer segmenters to isolate how much of the gain comes from supervision versus architecture.
  • The same composite mapping should adapt to 3D volumes with 3D wavelets and 3D box masks, which the paper does not test; that is the most direct path to clinical CT and MRI volumes.
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 / 5 minor

Summary. The paper proposes FreqU-FNet, a U-Net-like segmentation architecture that replaces conventional spatial downsampling with a frequency-domain encoder combining Fourier low-pass filtering and Daubechies wavelet downsampling, adds a Spatial Learnable Decoder with adaptive multi-branch upsampling, and uses a composite loss (Dice, Top-K cross-entropy, and wavelet-domain frequency loss). Experiments on MSD Prostate, MSD Pancreas, and MSD Lung are reported to support the claim that FreqU-FNet consistently outperforms CNN and Transformer baselines, especially on minority classes. The central empirical support is Tables 1-4, with Table 4 providing an ablation study on the MSD Pancreas dataset.

Significance. If the empirical results were reliable, the paper would address a genuine problem: class imbalance and frequency aliasing in medical image segmentation. The use of three public MSD benchmarks, the explicit DICE and Dice-Gap metrics, and the composite frequency-aware loss are sensible ingredients. However, the current experimental evidence is internally inconsistent and statistically unsupported. The duplicated baseline numbers between Tables 2 and 3, the contradiction between the abstract and Table 1, and the absence of error bars or code make it impossible to credit the claimed consistent improvement. The architectural novelty is also weakened by the algebraic simplification of Eq. (7). The significance of the contribution therefore cannot be assessed until the evidence is corrected and the claims are recalibrated.

major comments (5)
  1. [Section 4.4, Tables 2 and 3] Table 3 reports DICE(1) values for 2D nnU-Net (74.70), 3D nnU-Net (77.69), Ensemble (79.30), AttnUNet (78.65), and Swin-UNet (80.26) that are identical to the corresponding values in Table 2, even though these tables describe different datasets (MSD Pancreas and MSD Lung). Identical scores across different anatomies and class distributions are effectively impossible unless the numbers were reused, so the Lung comparison is not a trustworthy baseline and the claim of consistent superiority is unsupported.
  2. [Abstract and Section 4.4, Table 1] The abstract's claim that FreqU-FNet consistently outperforms both CNN and Transformer baselines is contradicted by Table 1: on MSD Prostate, FreqU-FNet's DICE1 of 73.21 is lower than Swin-UNet (75.11), UNETR (73.37), and TransUNet (73.42), and its Dice Diff of 18.49 is the largest among the transformer-based methods listed (Swin-UNet 14.96, UNETR 15.56, TransUNet 14.35). The paper must either provide a corrected empirical picture or substantially qualify the 'consistently outperforms' claim.
  3. [Section 3.2, Eq. (7)] The mapping Φ(X) simplifies because IWTdb(WTdb(X)) equals X for an orthogonal Daubechies transform; Eq. (7) therefore reduces to WTdb(F^{-1}(F(X) ⊙ M)). The proposed encoder is, after cancellation, a Fourier low-pass filter followed by wavelet downsampling, so the claimed novelty of combining wavelet multi-scale analysis with Fourier filtering is overstated and the design rationale should be restated accordingly.
  4. [Section 3.2, Eq. (4)] The core frequency encoder assumes that minority-class structures are separable by a single fixed box mask M(τ) with one ratio parameter τ across three heterogeneous datasets, yet no spectral analysis, sensitivity study, or per-dataset τ selection is provided. Without such evidence, the mechanism by which the frequency encoder improves minority-class segmentation remains unverified.
  5. [Section 4.3 and Tables 1-4] The paper reports five-fold cross-validation but gives no error bars, standard deviations, or per-fold results, and no code is released; small reported margins (e.g., 80.77 vs 80.26 on Lung) therefore cannot be assessed for statistical significance. This is a load-bearing omission for the 'consistently outperforms' claim.
minor comments (5)
  1. [Section 4.5, Table 4] The ablation discussion uses inconsistent names: FAL is called 'Frequency Attention Layer' here but defined as 'Frequency-Aware Loss' in Section 3.4, and SLD is called 'Spatial-Local Detail' here but defined as 'Spatial Learnable Decoder' in Section 3.3; also, removing DB downsampling narrows the Gap to 13.48%, better than the full model's 15.64%, so the conclusion that the full model offers the 'best trade-off' needs explicit justification.
  2. [Throughout] There are numerous typographical and grammatical errors that impede readability, including 'Comparsion', 'standardlized', 'an spatial', 'whether loss', 'checkboard', 'concentation', and 'expoential'.
  3. [Figure 1] The figure caption and inline labels are garbled and difficult to parse (e.g., 'II Init grid sample'); the figure should be redrawn with clean, consistent notation for both the encoder and decoder components.
  4. [Eq. (4)] The mask definition uses τh and τw as if h and w are integer dimensions; please clarify whether the mask extent is floor(τh) × floor(τw) and describe how τ is set for each dataset.
  5. [Section 2, References 21-25] Several related-work references (fashion generation, pose-guided generation, talking-face generation) are not well integrated into the segmentation argument and should be either removed or explicitly connected to the proposed method.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is evaluated on external benchmarks and the losses are standard; the Eq. (7) wavelet cancellation is a mathematical redundancy, not a circular derivation.

full rationale

I walked the derivation chain in Sections 3.2-3.4 and the experimental comparisons. The proposed encoder, decoder, and loss are not defined in terms of the reported Dice scores or class gaps, and the experiments use external public datasets (MSD Prostate, MSD Pancreas, MSD Lung) rather than self-generated predictions. I found no fitted parameter being renamed as a prediction, no load-bearing self-citation chain, and no uniqueness theorem imported from the authors. The one notable mathematical reduction is in Eq. (7): because IWTdb(WTdb(X)) = X by perfect reconstruction, the formal encoder mapping simplifies to WTdb(F^{-1}(F(X) ⊙ M)), meaning the initial wavelet decomposition and immediate inverse reconstruction cancel identically. This weakens the novelty claim of a 'wavelet-Fourier hybrid' but is not circularity: the resulting operator is still a nontrivial Fourier low-pass filter followed by a wavelet transform, and the reported empirical gains, if trustworthy, are not forced by construction. Separately, the duplication of baseline Dice values between Table 2 and Table 3 raises serious data-integrity concerns, but that is a reliability issue rather than a circular-derivation issue. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim depends on three unproven premises: perfect wavelet reconstruction (which makes Eq. (7) collapse), the existence of frequency-separable minority-class signals, and the safety of a fixed low-pass mask. Five hyperparameters (τ, k, and three loss weights) are introduced without specified values. No new entities are invented.

free parameters (5)
  • τ (low-pass filter ratio)
    Controls the size of the preserved low-frequency region in Eq. (4); chosen by hand, no value or sensitivity analysis given.
  • k (Top-K percentage)
    Percentage of hardest pixels used in the Top-K cross-entropy loss, Eq. (13); hyperparameter, value not reported.
  • wDice
    Weight for multi-class Dice loss in Eq. (14); not specified.
  • wTopK
    Weight for Top-K loss in Eq. (14); not specified.
  • wFreq
    Weight for frequency-aware loss in Eq. (14); not specified.
assumptions (3)
  • standard math Daubechies wavelet transform and its inverse are perfect reconstruction operators, so IWTdb(WTdb(X)) = X.
    Relied on implicitly in Eq. (7); if boundary effects or coefficient truncation break invertibility, the composite mapping changes.
  • ad hoc to paper Minority-class structures (e.g., small tumors) occupy distinguishable low-frequency bands separable by a fixed box mask.
    Stated in Section 3.2 as motivation but no spectral analysis or dataset-specific evidence is provided.
  • domain assumption Fourier low-pass filtering removes aliasing without removing clinically relevant high-frequency boundary information.
    The design replaces pooling with this filter; the loss adds a high-frequency term presumably to compensate, but the filter's effect on small structures is not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FreqU-FNet: Frequency-Aware U-Net for Imbalanced Medical Image Segmentation." pith.science (2026). https://pith.science/paper/OXAKDRQA

@misc{pith2026250517544,
  author       = {Pith},
  title        = {Pith review of: FreqU-FNet: Frequency-Aware U-Net for Imbalanced Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OXAKDRQA}},
  note         = {Machine review of arXiv:2505.17544}
}
read the original abstract

Medical image segmentation faces persistent challenges due to severe class imbalance and the frequency-specific distribution of anatomical structures. Most conventional CNN-based methods operate in the spatial domain and struggle to capture minority class signals, often affected by frequency aliasing and limited spectral selectivity. Transformer-based models, while powerful in modeling global dependencies, tend to overlook critical local details necessary for fine-grained segmentation. To overcome these limitations, we propose FreqU-FNet, a novel U-shaped segmentation architecture operating in the frequency domain. Our framework incorporates a Frequency Encoder that leverages Low-Pass Frequency Convolution and Daubechies wavelet-based downsampling to extract multi-scale spectral features. To reconstruct fine spatial details, we introduce a Spatial Learnable Decoder (SLD) equipped with an adaptive multi-branch upsampling strategy. Furthermore, we design a frequency-aware loss (FAL) function to enhance minority class learning. Extensive experiments on multiple medical segmentation benchmarks demonstrate that FreqU-FNet consistently outperforms both CNN and Transformer baselines, particularly in handling under-represented classes, by effectively exploiting discriminative frequency bands.

Figures

Figures reproduced from arXiv: 2505.17544 by the authors.

Figure 1
Figure 1. The Structure of Component of Frequency domain-based Encoder and Spatial Learnable Decoder. The Encoder component shows both the convolution and downsampling process. Struc￾ture of Decoder mainly shows the upsampling component. 2 Related Work 2.1 Medical Image Segmentation Medical image segmentation is a foundational task in computer-aided diagnosis and sur￾gical planning. Classical architectures like U-Net [20] and… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

  1. [1]

    arXiv preprint arXiv:1707.02937 (2017)

    Aitken, A., Ledig, C., Theis, L., Caballero, J., Wang, Z., Shi, W.: Checkerboard artifact free sub-pixel convolution: A note on sub-pixel convolution, resize convolution and convolution resize. arXiv preprint arXiv:1707.02937 (2017)

    arXiv 2017

  2. [2]

    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)

    work page 2022

  3. [3]

    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 .: Tran- sunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)

    arXiv 2021

  4. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Chen, L., Fu, Y ., Gu, L., Yan, C., Harada, T., Huang, G.: Frequency-aware feature fusion for dense image prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    work page 2024

  5. [5]

    arXiv preprint arXiv:2403.09065 (2024)

    Chen, L., Gu, L., Fu, Y .: When semantic segmentation meets frequency aliasing. arXiv preprint arXiv:2403.09065 (2024)

    arXiv 2024

  6. [6]

    In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV)

    Cheng, Z., Yuan, D., Zhang, W., Lukasiewicz, T.: Effective and efficient medical image seg- mentation with hierarchical context interaction. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). pp. 9396–9405. IEEE (2025)

    work page 2025

  7. [7]

    In: International Workshop on Deep Learning in Medical Image Analysis

    Drozdzal, M., V orontsov, E., Chartrand, G., Kadoury, S., Pal, C.: The importance of skip con- nections in biomedical image segmentation. In: International Workshop on Deep Learning in Medical Image Analysis. pp. 179–187. Springer (2016)

    work page 2016

  8. [8]

    Pattern Recognition 64, 255–267 (2017)

    Duan, Y ., Liu, F., Jiao, L., Zhao, P., Zhang, L.: Sar image segmentation based on convolutional-wavelet neural network and markov random field. Pattern Recognition 64, 255–267 (2017)

    work page 2017

Show all 32 references
  1. [9]

    Advances in neural information processing systems 30 (2017)

    Fan, Y ., Lyu, S., Ying, Y ., Hu, B.: Learning with average top-k loss. Advances in neural information processing systems 30 (2017)

    2017

  2. [10]

    In: European Conference on Computer Vision

    Finder, S.E., Amoyal, R., Treister, E., Freifeld, O.: Wavelet convolutions for large receptive fields. In: European Conference on Computer Vision. pp. 363–380. Springer (2024)

    2024

  3. [11]

    Advances in Neural Information Processing Systems35, 20592–20606 (2022)

    Finder, S.E., Zohav, Y ., Ashkenazi, M., Treister, E.: Wavelet feature maps compression for image-to-image cnns. Advances in Neural Information Processing Systems35, 20592–20606 (2022)

    2022

  4. [12]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Hatamizadeh, A., Tang, Y ., Nath, V ., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 574–584 (2022) 14 Ruiqi Xing

    2022

  5. [13]

    In: ICASSP 2020- 2020 IEEE international conference on acoustics, speech and signal processing (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 2020- 2020 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 1055–1...

    2020

  6. [14]

    Nature meth- ods 18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self- configuring method for deep learning-based biomedical image segmentation. Nature meth- ods 18(2), 203–211 (2021)

    2021

  7. [15]

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

    Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. In: Inter- national Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 488–...

    2024

  8. [16]

    IEEE transactions on medical imaging 37(12), 2663–2674 (2018)

    Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.W., Heng, P.A.: H-denseunet: hybrid densely con- nected unet for liver and tumor segmentation from ct volumes. IEEE transactions on medical imaging 37(12), 2663–2674 (2018)

    2018

  9. [17]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops

    Liu, P., Zhang, H., Zhang, K., Lin, L., Zuo, W.: Multi-level wavelet-cnn for image restora- tion. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 773–782 (2018)

    2018

  10. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Liu, W., Lu, H., Fu, H., Cao, Z.: Learning to upsample by learning to sample. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6027–6037 (2023)

    2023

  11. [19]

    In: European Conference on Computer Vision

    Lu, H., Liu, W., Fu, H., Cao, Z.: Fade: Fusing the assets of decoder and encoder for task- agnostic upsampling. In: European Conference on Computer Vision. pp. 231–247. Springer (2022)

    2022

  12. [20]

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

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

    2015

  13. [21]

    In: Proceedings of the AAAI Conference on Artificial Intel- ligence

    Shen, F., Jiang, X., He, X., Ye, H., Wang, C., Du, X., Li, Z., Tang, J.: Imagdressing-v1: Customizable virtual dressing. In: Proceedings of the AAAI Conference on Artificial Intel- ligence. vol. 39, pp. 6795–6804 (2025)

    2025

  14. [22]

    Advances in neural information processing systems 37, 6246–6266 (2024)

    Shen, F., Tang, J.: Imagpose: A unified conditional framework for pose-guided person gen- eration. Advances in neural information processing systems 37, 6246–6266 (2024)

    2024

  15. [23]

    arXiv preprint arXiv:2502.09533 (2025)

    Shen, F., Wang, C., Gao, J., Guo, Q., Dang, J., Tang, J., Chua, T.S.: Long-term talkingface generation via motion-prior conditional diffusion model. arXiv preprint arXiv:2502.09533 (2025)

    2025 arXiv

  16. [24]

    arXiv preprint arXiv:2310.06313 (2023)

    Shen, F., Ye, H., Zhang, J., Wang, C., Han, X., Yang, W.: Advancing pose-guided image synthesis with progressive conditional diffusion models. arXiv preprint arXiv:2310.06313 (2023)

    2023 arXiv

  17. [25]

    arXiv preprint arXiv:2504.13176 (2025)

    Shen, F., Yu, J., Wang, C., Jiang, X., Du, X., Tang, J.: Imaggarment-1: Fine-grained garment generation for controllable fashion design. arXiv preprint arXiv:2504.13176 (2025)

    2025 arXiv

  18. [26]

    Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice over- lap as a deep learning loss function for highly unbalanced segmentations. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third Inte...

    2017

  19. [27]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, J., Chen, K., Xu, R., Liu, Z., Loy, C.C., Lin, D.: Carafe: Content-aware reassembly of features. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3007–3016 (2019)

    2019

  20. [28]

    arXiv preprint arXiv:1505.02496 (2015) FreqU-FNet: Frequency-Aware U-Net for Imbalanced Medical Image Segmentation 15

    Wang, L., Lee, C.Y ., Tu, Z., Lazebnik, S.: Training deeper convolutional networks with deep supervision. arXiv preprint arXiv:1505.02496 (2015) FreqU-FNet: Frequency-Aware U-Net for Imbalanced Medical Image Segmentation 15

    2015 arXiv

  21. [29]

    Pattern recognition 143, 109819 (2023)

    Xu, G., Liao, W., Zhang, X., Li, C., He, X., Wu, X.: Haar wavelet downsampling: A sim- ple but effective downsampling module for semantic segmentation. Pattern recognition 143, 109819 (2023)

    2023

  22. [30]

    IEEE Geoscience and Remote Sensing Letters 15(5), 749–753 (2018)

    Zhang, Z., Liu, Q., Wang, Y .: Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters 15(5), 749–753 (2018)

    2018

  23. [31]

    arXiv preprint arXiv:2109.03201 (2021)

    Zhou, H.Y ., Guo, J., Zhang, Y ., Yu, L., Wang, L., Yu, Y .: nnformer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201 (2021)

    2021 arXiv

  24. [32]

    IEEE Transactions on Medical Imaging 39(6), 1856–1867 (2020)

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: Redesigning skip connec- tions to exploit multiscale features in image segmentation. IEEE Transactions on Medical Imaging 39(6), 1856–1867 (2020). https://doi.org/10.1109/TMI.2019.2959609

    2020

Pith tools

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