Pith. sign in

REVIEW 3 major objections 6 minor 46 references

HNOSeg-XS: Extremely Small Hartley Neural Operator for Efficient and Resolution-Robust 3D Image Segmentation

T0 review · 3 major / 6 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 HNOSeg-XS, a Hartley-based neural operator with fewer than 34.7k parameters, matches or beats much larger CNN and transformer models on 3D medical image segmentation while staying accurate when training and test resolutions differ.

desk verdict A clever efficiency trick worth a look, but the resolution-invariance story needs much more care than the paper gives it. read the letter →

arxiv 2507.08205 v1 pith:N5PGM2ZY submitted 2025-07-10 cs.CV

Ken C. L. Wong , Hongzhi Wang , Tanveer Syeda-Mahmood This is my paper · ORCID
classification cs.CV
keywords HartleyneuraloperatorFourierzero-shotsuper-resolution3Dmedicalimagesegmentationresolutionrobustnessfrequency-domainlearningparameter-efficientdeep
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 aims to show that 3D medical image segmentation can be done by an extremely small neural operator that keeps its accuracy when the input resolution changes. The proposed HNOSeg-XS uses the Hartley transform—a real-valued counterpart of the Fourier transform—inside a Fourier-style neural operator, with a single shared frequency-domain kernel and stacked nonlinearities. The authors report that the model has fewer than 34.7k parameters, segments a full 3D volume in under 0.24 seconds using less than 1.8 GiB of GPU memory, and loses far less accuracy than CNN and transformer baselines when trained at low resolution and tested at the full image size on three public datasets. If correct, medical segmentation models could be trained cheaply at low resolution and then applied directly at clinical resolution, avoiding the accuracy loss that discrete CNN and transformer models suffer when training and test resolutions differ.

What carries the argument

The load-bearing object is the HNO-XS block, a Hartley-domain neural operator block defined by the iteration above. The Hartley transform $\mathcal{H}$ maps real signals to real frequency-domain signals (it is $\operatorname{Re}(\mathcal{F}f)-\operatorname{Im}(\mathcal{F}f)$ for the Fourier transform $\mathcal{F}$), so ordinary real-valued deep learning operations—the SELU nonlinearity, $1\times1\times1$ convolutions, and residual additions—can be applied inside the frequency domain. Cropping to $k_{\max}$ modes in each dimension keeps only the low-frequency content that matters for segmentation, and the shared kernel means the parameter count does not grow with the number of modes. The block's composition rule collapses $n_{\mathrm{XS}}$ frequency-domain updates into one forward transform, one backward transform, and a chain of cheap frequency-domain convolutions and activations; this is what makes the model fast and memory-efficient.

What would settle it

On a new 3D dataset that is not used for tuning, fix $k_{\max}$ and $n_{\mathrm{XS}}$ to the paper's reported values, train HNOSeg-XS at one-quarter resolution, and evaluate at full resolution. If the Dice drop is comparable to a CNN baseline's drop (more than roughly ten points) rather than the few points reported here, the central claim of working across resolutions fails; the same test would also reveal whether the reported $n_{\mathrm{XS}}=2$ to $3$ ceiling is an artifact of the chosen datasets.

Watch

Extended reading notes

Core claim

The paper's central claim is that the Fourier neural operator's useful properties—global receptive field and zero-shot super-resolution—survive replacing the complex Fourier transform with the real Hartley transform and sharing one learned kernel $\hat{R}_t$ over all frequency modes. In the resulting HNO-XS block, multiple learned updates are composed entirely in the frequency domain, $$u_{t+n_{\mathrm{XS}}}(x)=\$\sigma$\!\left(\mathcal{H}^{-1}\!\left[\$\sigma$(\hat{R}_{t+n_{\mathrm{XS}}-1}\cdots\$\sigma$(\hat{R}_t\,\mathcal{H}u_t)\cdots)\right](x)\right),$$ so only one forward and one inverse transform pair is needed per block. With SELU activations, residual connections, and U-Net-style skip connections, the paper reports that this model holds its segmentation accuracy across a four-fold change in training resolution on BraTS'23, KiTS'23, and MVSeg'23, while having fewer parameters, faster inference, and lower memory than the tested CNN and transformer models.

Load-bearing premise

The whole result rests on the premise that a shared-kernel nonlinear stack in the real Hartley domain inherits the Fourier neural operator's property of working across resolutions; the paper supports this on three datasets, but its key hyperparameters ($k_{\max}$ and $n_{\mathrm{XS}}$) were chosen by sweeping over test-set results, so the property has not been shown to hold independently of the data used to tune it.

Editorial extensions

If this is right

  • A HNOSeg-XS model trained at half or quarter resolution can be applied at full resolution with only a small Dice drop, so training compute and memory can be cut by a large factor without retraining.
  • With fewer than 34.7k parameters and under 1.8 GiB of memory, the model fits on GPUs that cannot host a 140M-parameter transformer, and its sub-0.24-second inference is suitable for near-real-time use.
  • Because the operator has a global receptive field and no patch tokenization, the same weights work across input sizes, removing the need for resolution-specific preprocessing or patch stitching.
  • The block can be inserted into U-Net-style architectures and other designs, so the efficiency gains are not limited to the specific network shown in the paper.

Reading between the lines

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

  • A natural next step is to test the same HNO-XS block on 2D segmentation, registration, or image reconstruction; the block itself is architecture-agnostic, and the paper notes it can be incorporated into other designs, but does not evaluate those tasks.
  • The current implementation computes the Hartley transform via the standard FFT with real-imaginary subtraction, which the paper says is slower than a native real-input FFT; a dedicated Hartley FFT would likely make the already fast inference even faster.
  • The reported numbers were obtained with per-dataset test-set tuning of $k_{\max}$ and $n_{\mathrm{XS}}$, so a user applying the architecture to a new dataset with default hyperparameters should expect somewhat weaker resistance to resolution change than the reported figures.
  • The comparison is limited to downsampling factors of 1 to 4; extrapolating to larger factors or to variable voxel spacings across scanners is an untested extension, so the model should not be expected to be resolution-agnostic in the strict mathematical sense.
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 / 6 minor

Summary. The manuscript proposes HNOSeg-XS, a 3D medical image segmentation architecture built on a Hartley neural operator. The key modifications are replacing the complex Fourier transform with the real Hartley transform, sharing a single frequency-domain kernel across all modes (Eq. 4), applying consecutive SELU nonlinearities inside the frequency domain (Eqs. 5-7), removing normalization layers via self-normalizing network principles, and adding U-Net-style skip connections. The authors claim fewer than 34.7k parameters, resolution robustness inherited from FNO, and the fastest and most memory-efficient inference among tested CNN and transformer baselines on BraTS'23, KiTS'23, and MVSeg'23.

Significance. If the claims are sustained, this is a useful efficiency contribution: it shows that a sub-35k-parameter neural operator can segment 3D medical images with a global receptive field and retain reasonable accuracy after training at reduced resolution. The algebraic reformulation in Eqs. (6)-(7) is transparent, the code repository is referenced, and the experiments use three public challenges. The significance is tempered by two unaddressed issues: the frequency-domain SELU in Eq. (7) makes the inherited 'resolution-agnostic' claim non-automatic, and the empirical evaluation selects key hyperparameters on the test set.

major comments (3)
  1. [Section III-C, Eq. (7)] The claimed inheritance of FNO's zero-shot super-resolution property is not established for the modified block. With the standard unnormalized FFT/Hartley transform used in Section IV-F, forward Hartley coefficients scale with the number of voxels N and the inverse transform carries a 1/N factor. For the linear spectral multiplication in Eq. (4) these factors cancel; with the pointwise SELU inside the inverse transform in Eqs. (5)-(7) they do not, because SELU is not homogeneous and treats negative pre-activations nonlinearly. The operator in Eq. (7) is therefore, in general, resolution-dependent unless a specific transform normalization is supplied and a scale-invariance argument is given. The paper provides neither. The robustness observed in Figs. 6-9 may be an empirical property of the trained weights and this particular scaling rather than a guaranteed architectural property.
  2. [Section IV-C and IV-D, Figs. 4-5] Hyperparameters nXS and kmax are selected from test-set results. The captions of Figs. 4 and 5 state that each point is the average over the testing data, and Section IV-D explicitly says 'Fig. 5 was used to set kmax for each model and dataset.' The Dice values reported in Table I and Fig. 6 for HNOSeg-XS and the other neural-operator models are therefore partly fitted to the test data. This makes the claimed superiority over baselines, whose hyperparameters are not tuned on this test set, an optimistic comparison. The authors should select these hyperparameters on the validation split before evaluating on the test set, or report both validation-selected performance and final test performance for the chosen configuration.
  3. [Section IV-E, Table I and Fig. 6] The main resolution-robustness comparison is presented without any measure of variability. Each cell is a single Dice or HD95 value with no number of runs, standard deviation, or significance test. Some of the claimed advantages at low training resolutions are small, for example HNOSeg-XS versus HNOSeg in several rows of Table I, so it is not possible to tell whether the observed ordering is reproducible. At least three runs with mean and standard deviation, or a paired statistical test across test subjects, should be reported for the headline robustness comparisons.
minor comments (6)
  1. [Section V] The text 'HNOSeg and NHOSeg-XS' contains a typo; it should read 'HNOSeg and HNOSeg-XS'.
  2. [Section IV-B] The dataset name 'KiT'23' is a typo and should be 'KiTS'23'.
  3. [Section IV-E] The phrase 'could not probably segment' should read 'could not properly segment'.
  4. [Section IV-F] The statement that HNOSeg-XS parameters are 'fewer than 0.2% of TransUNet' is slightly inaccurate: 34.7k / 16.7M is about 0.21%.
  5. [Footnote 1] The code repository URL contains a line-break space in 'mul timodal'; the URL should be typeset as a single token.
  6. [Figs. 4-6] The figure captions should state explicitly whether the plotted values are from a single run or a single fold; if so, error bars are needed in light of major comment 3.

Circularity Check

1 steps flagged · score 4.0 of 10

Test-set hyperparameter selection makes the reported accuracy comparisons partly in-sample, but no equation-level derivation reduces to a fitted value.

  1. fitted input called prediction [Section III-D; Section IV-C/IV-D, Figs. 4-5; Table I]
    "The hyperparameters nXS and kmax can be obtained empirically (Section IV-C and IV-D). ... Fig. 5 was used to set kmax for each model and dataset. ... Each point represents the average value of different regions of the testing data."

    The model hyperparameters kmax and nXS are chosen by reading Dice coefficients computed on the testing splits, as stated in the Fig. 4 and Fig. 5 captions: “Each point represents the average value of different regions of the testing data.” The same testing splits are then used in Table I and Section IV-E to substantiate the “superior resolution robustness” claim. The reported test results are therefore not out-of-sample predictions for the selected configurations; they are in-sample evaluations after hyperparameter selection on the same labels. This does not make Eq. (7) a tautology, but it makes the numerical headline statistically forced by the selection procedure rather than independently verified.

full rationale

The architecture is defined by Eqs. (1)-(7) and is not, by construction, equivalent to the experimental outcomes; the zero-shot super-resolution property is inherited from the externally cited FNO framework [15], and the Hartley reformulation plus frequency-domain nonlinearity is an architectural hypothesis rather than a tautology. Self-citations [14], [39], [45] are used as prior work, baselines, or loss references, not as uniqueness theorems or unverified load-bearing justifications. The main circularity-adjacent issue is validation: nXS and kmax were tuned on the same testing data that produce the headline Dice and robustness comparisons, so the central empirical claim is partially fitted. The separate concern that SELU in Eq. (7) is not scale-invariant and may break FNO's zero-shot property is a correctness or foundation gap rather than a circular reduction, so it does not raise the circularity score further.

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

The central claims rest primarily on empirical validation. The only mathematical dependencies are standard transform identities and the SNN self-normalization result. The key domain assumptions (few modes suffice, resolution-agnosticism carries over) are plausible but only indirectly evidenced.

free parameters (4)
  • kmax = (14,14,10) BraTS; (16,16,8) KiTS; (14,10,12) MVSeg
    Number of frequency modes per dimension; selected empirically from test data (Fig. 5).
  • nXS = 3 (BraTS, MVSeg); 2 (KiTS)
    Number of frequency-domain convolutions per block; selected from test data (Fig. 4).
  • d = 24
    Channel width for all neural operator models; chosen by hand.
  • nB or nXS*nB = 24 total
    Depth of network; chosen by hand.
assumptions (4)
  • standard math The discrete Hartley convolution theorem (Eq. 3) with a shared kernel simplifies to Eq. (4).
    A direct consequence of Bracewell's theorem when Rhat_t is constant over k.
  • domain assumption A small number of low-frequency modes captures sufficient information for image segmentation.
    Motivated by Fig. 2, but figure shows image reconstruction, not segmentation; no proof for the segmentation objective.
  • standard math SELU activation with the stated initialization self-normalizes the network, removing the need for normalization layers.
    Taken from SNN [16] and applied here without re-derivation.
  • domain assumption Zero-shot super-resolution of the continuous neural operator is preserved under the Hartley reformulation with mode cropping and zero-padding.
    Assumed by design; experiments support it but there is no formal analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HNOSeg-XS: Extremely Small Hartley Neural Operator for Efficient and Resolution-Robust 3D Image Segmentation." pith.science (2026). https://pith.science/paper/N5PGM2ZY

@misc{pith2026250708205,
  author       = {Pith},
  title        = {Pith review of: HNOSeg-XS: Extremely Small Hartley Neural Operator for Efficient and Resolution-Robust 3D Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N5PGM2ZY}},
  note         = {Machine review of arXiv:2507.08205}
}
read the original abstract

In medical image segmentation, convolutional neural networks (CNNs) and transformers are dominant. For CNNs, given the local receptive fields of convolutional layers, long-range spatial correlations are captured through consecutive convolutions and pooling. However, as the computational cost and memory footprint can be prohibitively large, 3D models can only afford fewer layers than 2D models with reduced receptive fields and abstract levels. For transformers, although long-range correlations can be captured by multi-head attention, its quadratic complexity with respect to input size is computationally demanding. Therefore, either model may require input size reduction to allow more filters and layers for better segmentation. Nevertheless, given their discrete nature, models trained with patch-wise training or image downsampling may produce suboptimal results when applied on higher resolutions. To address this issue, here we propose the resolution-robust HNOSeg-XS architecture. We model image segmentation by learnable partial differential equations through the Fourier neural operator which has the zero-shot super-resolution property. By replacing the Fourier transform by the Hartley transform and reformulating the problem in the frequency domain, we created the HNOSeg-XS model, which is resolution robust, fast, memory efficient, and extremely parameter efficient. When tested on the BraTS'23, KiTS'23, and MVSeg'23 datasets with a Tesla V100 GPU, HNOSeg-XS showed its superior resolution robustness with fewer than 34.7k model parameters. It also achieved the overall best inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) compared to the tested CNN and transformer models.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages

  1. [1]

    Deep learning tech- niques for medical image segmentation: achievements and challenges,

    M. H. Hesamian, W. Jia, X. He, and P. Kennedy, “Deep learning tech- niques for medical image segmentation: achievements and challenges,” Journal of Digital Imaging , vol. 32, no. 4, pp. 582–596, 2019. KEN C. L. WONG et al. : HNOSEG-XS: EXTREMEL Y SMALL HARTLEY NEURAL OPERATOR FOR EFFICIENT AND RESOLUTION-ROBUST 3D IMAGE SEGMENTATION 11

    work page 2019

  2. [2]

    A review of deep-learning-based medical image segmentation methods,

    X. Liu, L. Song, S. Liu, and Y . Zhang, “A review of deep-learning-based medical image segmentation methods,” Sustainability, vol. 13, no. 3, p. 1224, 2021

    work page 2021

  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

    2022

  4. [4]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , 2017

    work page 2017

  5. [5]

    An image is worth 16x16 words: Trans- formers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Trans- formers for image recognition at scale,” in International Conference on Learning Representations, 2021

    2021

  6. [6]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2021, pp. 10 012–10 022

    work page 2021

  7. [7]

    Transformers in medical image segmentation: A review,

    H. Xiao, L. Li, Q. Liu, X. Zhu, and Q. Zhang, “Transformers in medical image segmentation: A review,” Biomedical Signal Processing and Control, vol. 84, p. 104791, 2023

    work page 2023

  8. [8]

    UNETR: Transformers for 3D medical image segmentation,

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Land- man, H. R. Roth, and D. Xu, “UNETR: Transformers for 3D medical image segmentation,” in Proceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision , 2022, pp. 574–584

    work page 2022

Show all 46 references
  1. [9]

    Swin-Unet: Unet-like pure transformer for medical image segmenta- tion,

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-Unet: Unet-like pure transformer for medical image segmenta- tion,” in Computer Vision – ECCV 2022 Workshops, 2023, pp. 205–218

    2022

  2. [10]

    nn- Former: V olumetric medical image segmentation via a 3D transformer,

    H.-Y . Zhou, J. Guo, Y . Zhang, X. Han, L. Yu, L. Wang, and Y . Yu, “nn- Former: V olumetric medical image segmentation via a 3D transformer,” IEEE Transactions on Image Processing, vol. 32, pp. 4036–4045, 2023

    2023

  3. [11]

    UTNet: a hybrid transformer architecture for medical image segmentation,

    Y . Gao, M. Zhou, and D. N. Metaxas, “UTNet: a hybrid transformer architecture for medical image segmentation,” in International Confer- ence on Medical Image Computing and Computer-Assisted Intervention , 2021, pp. 61–71

    2021

  4. [12]

    CoTr: Efficiently bridging CNN and transformer for 3D medical image segmentation,

    Y . Xie, J. Zhang, C. Shen, and Y . Xia, “CoTr: Efficiently bridging CNN and transformer for 3D medical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Inter- vention, 2021, pp. 171–180

    2021

  5. [13]

    TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,

    J. Chen, J. Mei, X. Li, Y . Lu, Q. Yu, Q. Wei, X. Luo, Y . Xie, E. Adeli, Y . Wanget al. , “TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers,” Medical Image Analysis , vol. 97, p. 103280, 2024

    2024

  6. [14]

    FNOSeg3D: Resolution-robust 3D image segmentation with Fourier neural operator,

    K. C. L. Wong, H. Wang, and T. Syeda-Mahmood, “FNOSeg3D: Resolution-robust 3D image segmentation with Fourier neural operator,” in IEEE International Symposium on Biomedical Imaging , 2023, pp. 1– 5

    2023

  7. [15]

    Fourier neural operator for parametric partial differential equations,

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, “Fourier neural operator for parametric partial differential equations,” in International Conference on Learning Representations, 2021

    2021

  8. [16]

    Self- normalizing neural networks,

    G. Klambauer, T. Unterthiner, A. Mayr, and S. Hochreiter, “Self- normalizing neural networks,” in Advances in Neural Information Pro- cessing Systems, 2017, pp. 972–981

    2017

  9. [17]

    The multimodal brain tumor image segmentation benchmark (BRATS),

    B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby, Y . Burren, N. Porz, J. Slotboom, R. Wiest et al. , “The multimodal brain tumor image segmentation benchmark (BRATS),” IEEE Transactions on Medical Imaging, vol. 34, no. 10, pp. 1993–2024, 2015

    1993

  10. [18]

    Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features,

    S. Bakas, H. Akbari, A. Sotiras, M. Bilello, M. Rozycki, J. S. Kirby, J. B. Freymann, K. Farahani, and C. Davatzikos, “Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features,” Scientific Data, vol. 4, no. 170117, 2017

    2017

  11. [19]

    The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor seg- mentation and radiogenomic classification,

    U. Baid, S. Ghodasara, S. Mohan, M. Bilello, E. Calabrese, E. Colak, K. Farahani, J. Kalpathy-Cramer, F. C. Kitamura, S. Pati et al. , “The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor seg- mentation and radiogenomic classification,” arXiv:2107.02314 [cs.CV], 2021

    2021 arXiv

  12. [20]

    The KiTS21 Challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase CT,

    N. Heller, F. Isensee, D. Trofimova, R. Tejpaul, Z. Zhao, H. Chen, L. Wang, A. Golts, D. Khapun, D. Shats et al., “The KiTS21 Challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase CT,” arXiv:2307.01984 [cs.CV], 2023

    2023 arXiv

  13. [21]

    DeepMitral: Fully automatic 3D echocardiography segmentation for patient specific mitral valve modelling,

    P. Carnahan, J. Moore, D. Bainbridge, M. Eskandari, E. C. S. Chen, and T. M. Peters, “DeepMitral: Fully automatic 3D echocardiography segmentation for patient specific mitral valve modelling,” in Interna- tional Conference on Medical Image Computing and Computer-Assisted Inter...

    2021

  14. [22]

    U-Net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention , ser. LNCS, vol. 9351, 2015, pp. 234–241

    2015

  15. [23]

    V-Net: Fully convolutional neural networks for volumetric medical image segmentation,

    F. Milletari, N. Navab, and S.-A. Ahmadi, “V-Net: Fully convolutional neural networks for volumetric medical image segmentation,” in IEEE International Conference on 3D Vision , 2016, pp. 565–571

    2016

  16. [24]

    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 Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support , 2018, pp. 3–11

    2018

  17. [25]

    nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. 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

  18. [26]

    FDA: Fourier domain adaptation for semantic segmentation,

    Y . Yang and S. Soatto, “FDA: Fourier domain adaptation for semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , 2020

    2020

  19. [27]

    Global filter networks for image classification,

    Y . Rao, W. Zhao, Z. Zhu, J. Lu, and J. Zhou, “Global filter networks for image classification,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 980–993

    2021

  20. [28]

    FreMIM: Fourier transform meets masked image modeling for medical image segmentation,

    W. Wang, J. Wang, C. Chen, J. Jiao, Y . Cai, S. Song, and J. Li, “FreMIM: Fourier transform meets masked image modeling for medical image segmentation,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 7860–7870

    2024

  21. [29]

    Implicit neural representation in medical imaging: A comparative survey,

    A. Molaei, A. Aminimehr, A. Tavakoli, A. Kazerouni, B. Azad, R. Azad, and D. Merhof, “Implicit neural representation in medical imaging: A comparative survey,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , 2023, pp. 2381– 2391

    2023

  22. [30]

    Implicit neural representations for medical imaging segmentation,

    M. O. Khan and Y . Fang, “Implicit neural representations for medical imaging segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention , 2022, pp. 433–443

    2022

  23. [31]

    Implicit anatomical rendering for medical image segmentation with stochastic experts,

    C. You, W. Dai, Y . Min, L. Staib, and J. S. Duncan, “Implicit anatomical rendering for medical image segmentation with stochastic experts,” in International Conference on Medical Image Computing and Computer- Assisted Intervention, 2023, pp. 561–571

    2023

  24. [32]

    NISF: Neural implicit segmentation functions,

    N. Stolt-Ans ´o, J. McGinnis, J. Pan, K. Hammernik, and D. Rueckert, “NISF: Neural implicit segmentation functions,” inInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention , 2023, pp. 734–744

    2023

  25. [33]

    Aubert and P

    G. Aubert and P. Kornprobst, Mathematical problems in image pro- cessing: partial differential equations and the calculus of variations . Springer, 2006, vol. 147

    2006

  26. [34]

    Sapiro, Geometric partial differential equations and image analysis

    G. Sapiro, Geometric partial differential equations and image analysis . Cambridge university press, 2006

    2006

  27. [35]

    Neural operator: Graph kernel network for partial differential equations,

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, “Neural operator: Graph kernel network for partial differential equations,” arXiv:2003.03485 [cs.LG], 2020

    2003 arXiv

  28. [36]

    A more symmetrical Fourier analysis applied to transmission problems,

    R. V . L. Hartley, “A more symmetrical Fourier analysis applied to transmission problems,” Proceedings of the IRE, vol. 30, no. 3, pp. 144– 150, 1942

    1942

  29. [37]

    Discrete Hartley transform,

    R. N. Bracewell, “Discrete Hartley transform,” Journal of the Optical Society of America , vol. 73, no. 12, pp. 1832–1835, 1983

    1983

  30. [38]

    The importance of phase in signals,

    A. V . Oppenheim and J. S. Lim, “The importance of phase in signals,” Proceedings of the IEEE , vol. 69, no. 5, pp. 529–541, 1981

    1981

  31. [39]

    3D segmentation with fully trainable Gabor kernels and Pearson’s correlation coefficient,

    K. C. L. Wong and M. Moradi, “3D segmentation with fully trainable Gabor kernels and Pearson’s correlation coefficient,” in Machine Learn- ing in Medical Imaging , 2022, pp. 53–61

    2022

  32. [40]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. L. Ba, “Adam: A method for stochastic optimiza- tion,” arXiv:1412.6980 [cs.LG], 2014

    2014 arXiv

  33. [41]

    SGDR: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” in International Conference on Learning Representa- tions, 2017

    2017

  34. [42]

    BraTS 2023 Challenge,

    “BraTS 2023 Challenge,” https://www.synapse.org/Synapse:syn51156 910/wiki/, 2023, accessed on February 14, 2025

    2023

  35. [43]

    The 2023 Kidney and Kidney Tumor Segmentation Challenge,

    “The 2023 Kidney and Kidney Tumor Segmentation Challenge,” https: //kits-challenge.org/kits23/, 2023, accessed on July 21, 2024

    2023

  36. [44]

    MVSEG2023,

    “MVSEG2023,” https://www.synapse.org/Synapse:syn51186045/wiki/, 2023, accessed on February 14, 2025

    2023

  37. [45]

    3D segmentation with exponential logarithmic loss for highly unbalanced object sizes,

    K. C. L. Wong, M. Moradi, H. Tang, and T. Syeda-Mahmood, “3D segmentation with exponential logarithmic loss for highly unbalanced object sizes,” in International Conference on Medical Image Computing and Computer-Assisted Intervention , ser. LNCS, vol. 11072, 2018, pp. 612–619

    2018

  38. [46]

    Variability in image quality and radiation dose within and across 97 medical facilities,

    T. B. Smith, S. Zhang, A. Erkanli, D. Frush, and E. Samei, “Variability in image quality and radiation dose within and across 97 medical facilities,” Journal of Medical Imaging , vol. 8, no. 5, p. 052105, 2021

    2021

Pith tools

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