Pith. sign in

REVIEW 5 major objections 4 minor 29 references

Localized FNO for Spatiotemporal Hemodynamic Upsampling in Aneurysm MRI

T0 review · 5 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 The paper claims that the Localized Fourier Neural Operator (LoFNO), which combines a super-resolution network, a domain-localized Fourier operator, and Laplacian eigenvector geometric priors, outperforms interpolation and alternative…

desk verdict A clean engineering contribution with honest synthetic benchmarks, but the temporal upsampling claim is weakened by a shared pressure pulse across all simulations. read the letter →

arxiv 2507.13789 v1 pith:5BHDZU56 submitted 2025-07-18 cs.CV cs.AIphysics.comp-ph

Kyriakos Flouris , Moritz Halter , Yolanne Y. R. Lee , Samuel Castonguay , Luuk Jacobs , Pietro Dirix , Jonathan Nestmann , Sebastian Kozerke
show 1 more author
Ender Konukoglu
This is my paper
classification cs.CVcs.AIphysics.comp-ph
keywords 4DflowMRIFourierneuraloperatorsuper-resolutionhemodynamicswallshearstressintracranialaneurysmLaplacianeigenvectorstemporalupsampling
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 proposes LoFNO, a three-dimensional Fourier neural operator that upsamples noisy, low-resolution 4D flow MRI data into high-resolution velocity fields and wall shear stress maps on patient-specific aneurysm geometries. The authors claim that combining an EDSR super-resolution layer, a domain-localized Fourier operator, and Laplacian eigenvectors as geometric priors lets the network generalize to aneurysm shapes it never saw during training. On a synthetic dataset of 95 CFD simulations built from 80 aneurysm geometries, LoFNO reports lower test error than linear, RBF, and DFI interpolation and than SRCNN, EDSR, and plain FNO for both spatial and temporal upsampling. The temporal experiment includes predicting a full 24-frame pulse sequence from a single initial frame, a task interpolation methods cannot do.

What carries the argument

The load-bearing object is the LoFNO block: an EDSR super-resolution network concatenates noisy low-resolution velocity with the first 32 normalized graph-Laplacian eigenvectors resampled onto a voxel grid, producing a high-resolution feature volume; a domain-agnostic Fourier operator then applies FFT-based convolutions that are masked by a domain characteristic function so computations only interact inside the vessel lumen; implicit-FNO parameter sharing reduces overfitting; and a final MLP projects back to velocity. The eigenvectors supply a coordinate system for the shape, allowing the operator to be evaluated on unseen geometries, while the characteristic function keeps Fourier modes from being wasted on empty voxels, which matters most at the wall where wall shear stress is defined.

What would settle it

Take LoFNO trained on the 95 shared-pulse simulations and evaluate it on CFD runs with different inlet pressure waveforms and heart rates, or on gated 4D flow MRI; if full-sequence temporal prediction error rises to interpolation-level performance, the temporal claim is an artifact of the repeated pulse.

Watch

Extended reading notes

Core claim

The central claim is that hemodynamic fields over irregular vascular domains can be super-resolved by a neural operator that learns a spectral representation of the domain geometry through graph-Laplacian eigenvectors, localizes all Fourier interactions with a characteristic function that masks out non-vessel voxels, and uses a convolutional super-resolution head that also denoises the input. The authors state that this combination achieves the lowest test error on unseen aneurysm geometries for velocity and wall shear stress at spatial upsampling scales of two, three, and four and temporal upsampling scales of two, four, and full-sequence reconstruction. The full architecture is the Localized Fourier Neural Operator (LoFNO), and the gains are attributed to both the domain localization and the geometric prior, with the Laplacian prior giving a small consistent improvement over the localized operator alone.

Load-bearing premise

The temporal upsampling results assume that every test case follows the same inlet pressure pulse as the training cases, so a network that learned that repeated pulse may fail when applied to real patients with varying pulse waveforms and heart rates.

Editorial extensions

If this is right

  • If the reported errors hold, LoFNO could turn routinely acquired low-resolution 4D flow MRI into higher-resolution velocity fields and wall shear stress maps without running CFD, since the model is trained once on simulations and then applied directly to imaging data.
  • The temporal results imply that a single ECG-gated frame may suffice to reconstruct the full pulsatile cycle for aneurysms, potentially cutting scan time for hemodynamic assessment.
  • Because the eigenvectors and characteristic function are computed per geometry at inference time, the geometric prior should transfer to other irregular vascular territories without retraining.
  • Since masking out non-vessel voxels is the main source of gain, the same localization trick should apply to other operator learning tasks on bounded, irregular domains.

Reading between the lines

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

  • A likely boundary of the temporal gains is that all CFD cases share the same inlet pressure waveform, so the full-sequence prediction task is essentially learning one repeated pulse; real scans vary in waveform and heart rate, and a test with patient-like waveform variability would reveal how much of the temporal advantage is pulse memorization rather than learned hemodynamics.
  • Because training data are simulated, the velocity-to-wall-shear-stress mapping inherits the CFD assumptions, including the prescribed inlet pulse and vessel rigidity; clinical users should expect those assumptions to enter predictions for real patients.
  • A natural testable extension would be to condition the network explicitly on the inlet pressure waveform or heart rate, or to extend the output to pressure and oscillatory shear index, which the current architecture does not predict.
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 / 4 minor

Summary. The paper introduces LoFNO, a 3D neural architecture that combines an EDSR-based upsampling block, a DAFNO Fourier operator layer with domain localization, and Laplacian eigenvector geometric priors to upsample and denoise 4D flow hemodynamic data. The authors train on synthetic CFD simulations of aneurysm geometries from the Aneurisk dataset and report test errors on 10 unseen geometries for spatial super-resolution (scales 2, 3, 4) and temporal upsampling (32^3×12, 32^3×6, and 32^3×1 to 32^3×24), comparing against interpolation methods, SRCNN, EDSR, and FNO with an EDSR front-end. The central claim is that LoFNO achieves lower velocity and WSS reconstruction errors than all baselines on this synthetic test set.

Significance. If the benchmark result is robust, the paper makes a useful methodological contribution: it demonstrates a practical architecture for joint spatial and temporal upsampling of hemodynamic fields on irregular, unseen vascular geometries, with code and hyperparameters publicly available. The careful use of Laplacian eigenvectors as geometric priors and the localization mechanism from DAFNO are sensible design choices, and the spatial super-resolution results show consistent improvements across all scales. The temporal claim, however, is substantially weakened by the shared-pressure-pulse experimental setup, and the WSS evaluation is not reproducible as described. The work is therefore a promising step but does not yet support the clinical-transfer implications stated in the abstract.

major comments (5)
  1. [Section 3, 'Results' (after Table 2)] The temporal upsampling experiments are compromised by the fact that all 95 CFD simulations use the same pressure pulse at the vessel inlet, as the paper itself states. The claim that models were 'trained on slight pulse variations' is not supported by any description of such variations; if the pulse is identical across all training and test cases, the temporal task can be solved by memorizing a single temporal waveform. The Table 2 gains, especially the 32^3×1 to 32^3×24 full-sequence prediction, therefore do not demonstrate generalization to the heart-rate and waveform variability present in real 4D flow MRI. Please either describe the pulse variations used or add an experiment with varied pulses to support the temporal generalization claim.
  2. [Section 3, Tables 1 and 2] The paper reports WSS prediction errors for all methods but never specifies how WSS is computed from the predicted velocity fields. WSS depends on the wall shear rate and viscosity, and the conversion from volumetric velocity to surface WSS involves gradient estimation and boundary treatment; without this specification the WSS numbers in Tables 1 and 2 cannot be reproduced or interpreted. Please state the exact WSS computation formula, the viscosity value used, and the numerical scheme for surface extraction.
  3. [Section 3, Tables 1 and 2] The test results are reported as point averages over the 10 unseen geometries, without error bars, confidence intervals, or statistical significance tests. The statement that LoFNO 'consistently outperform[s] the other methods for every individual test case' is not substantiated by any per-case results. Please report the distribution of errors (e.g., mean ± std or per-case boxplots) and perform paired statistical comparisons across geometries, at least for the main table entries.
  4. [Section 2, 'Localizing to the Relevant Domains and Geometric Prior'] The graph Laplacian eigenvectors are described as 'the first k=32 eigenvectors, corresponding to the largest nonzero eigenvalues,' with eigenvalues 'sorted in descenting order.' For the normalized graph Laplacian, the smallest nonzero eigenvalues are the standard choice for geometric embeddings, and using the largest eigenvalues would select high-frequency modes. Please clarify the sorting direction and the intended selection, and justify the choice with respect to the geometric-prior role; if the implementation actually uses the smallest eigenvalues, correct the text.
  5. [Section 3, dataset description] The dataset description is inconsistent: the text says '95 pulse flows imposed on geometries from the Aneurisk dataset' and later 'We selected 80 aneurysm-containing regions.' Please clarify whether there are 95 simulations, 80 geometries, or some other count, and how these are split into training and test sets (the test set is stated to contain 10 geometries).
minor comments (4)
  1. [Abstract] The abstract claims the model can predict WSS 'directly from clinical imaging data,' but the method is trained on CFD data and no clinical or in-vivo validation is performed; the Limitations section appropriately acknowledges the CFD dependence. Please soften the abstract to avoid overclaiming clinical readiness.
  2. [Section 2] The word 'descenting' should be 'descending' in the eigenvector description, and 'adoptimplicit' appears as a typo in the Related Works paragraph.
  3. [Section 3] The training loss is defined as a relative error, but the test error is defined as err_f(x,t) = ||f(x,t) - hat f(x,t)||_2. Please clarify the exact relationship between the training loss and the reported L_test values in Tables 1 and 2.
  4. [Figure 3] The caption says the WSS vector lengths are 'colored black for the smallest to yellow for the largest,' but this is vague; please specify the colormap and the scale used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LoFNO's benchmark claims rest on supervised learning against independently simulated CFD ground truth, with geometric priors serving as inputs and all load-bearing architecture components coming from external prior work.

full rationale

The central claim is empirical test-error superiority on a synthetic CFD dataset. The network is trained with a supervised relative loss comparing its output to the simulated noiseless ground truth, and no parameter is fitted to the test set or renamed as a prediction. The Laplacian eigenvectors, the domain characteristic function, and the low-resolution velocity fields are inputs to the model, not outputs that could make the result true by definition. The main building blocks, DAFNO, EDSR, and implicit FNOs, are cited from external works rather than from self-citations that carry the argument. The paper's authors do cite their own prior works, but these citations are contextual and not load-bearing for the claimed upsampling performance. The strongest limitation identified in the paper, that all CFD simulations share the same pressure pulse at the vessel inlet, is a generalization and benchmarking confound for temporal upsampling, not a circularity: the temporal outputs are still evaluated against separately simulated ground truth, and the shared pulse does not make the network's prediction equal to its input by construction. The limitations section itself acknowledges the dependence on CFD training data and the lack of real-world variability, which further supports treating the benchmark as an in-domain empirical comparison rather than a derived or self-justifying result. No equation in the paper defines an output in terms of the claimed prediction, no fitted parameter is presented as a prediction, and no uniqueness theorem is imported from the authors' own prior work to force the chosen formulation. The paper is therefore self-contained with respect to circularity concerns.

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

The paper introduces no new physical entities. Its load-bearing assumptions are the fidelity of CFD ground truth, the single-pulse temporal setup, and the adequacy of synthetic white noise as a proxy for clinical MRI noise. The main free parameters are architectural choices tuned by ablation, including the number of Laplacian eigenvectors.

free parameters (3)
  • k (number of Laplacian eigenvectors) = 32
    The selection of the first 32 eigenvectors as geometric prior is a design choice optimized via ablation, not derived from theory or data. Section 2.
  • DAFNO/FNO hyperparameters (modes N_m, layers L, hidden dim d_h) = not reported
    The paper states the architecture and hyperparameters were optimized via ablation studies (Section 3 and Fig. 2 caption), but exact values are not given in the text.
  • SNR of added white noise = 10
    White noise with SNR 10 is a simulation choice assumed to represent clinical 4D flow MRI noise; it is not derived from measured noise characteristics. Section 3.
assumptions (5)
  • domain assumption Navier-Stokes equations with LES/WALE subgrid model produce accurate ground-truth hemodynamics.
    Section 3: simulations are used as ground truth for training and testing; no validation against in vivo measurements is provided.
  • domain assumption Blood is Newtonian and incompressible with density 1060 kg/m3 and kinematic viscosity 3.5e-3 Pa·s.
    Section 3: standard hemodynamic modeling choices, not verified against patient-specific blood properties.
  • domain assumption All simulations impose the same pressure pulse at the vessel inlet.
    Section 3 temporal upsampling paragraph: this enables the temporal prediction task but limits generalization to variable clinical pulse waveforms.
  • ad hoc to paper White Gaussian noise with SNR 10 approximates 4D flow MRI noise.
    Section 3: no real noise characterization is used, and real data analysis is declared infeasible in Section 1.
  • domain assumption Laplacian eigenvectors computed on the vessel surface mesh and resampled to the Cartesian grid provide a transferable geometric prior for unseen geometries.
    Section 2: supported only by the ablation table, not by a theoretical guarantee or independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Localized FNO for Spatiotemporal Hemodynamic Upsampling in Aneurysm MRI." pith.science (2026). https://pith.science/paper/5BHDZU56

@misc{pith2026250713789,
  author       = {Pith},
  title        = {Pith review of: Localized FNO for Spatiotemporal Hemodynamic Upsampling in Aneurysm MRI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5BHDZU56}},
  note         = {Machine review of arXiv:2507.13789}
}
read the original abstract

Hemodynamic analysis is essential for predicting aneurysm rupture and guiding treatment. While magnetic resonance flow imaging enables time-resolved volumetric blood velocity measurements, its low spatiotemporal resolution and signal-to-noise ratio limit its diagnostic utility. To address this, we propose the Localized Fourier Neural Operator (LoFNO), a novel 3D architecture that enhances both spatial and temporal resolution with the ability to predict wall shear stress (WSS) directly from clinical imaging data. LoFNO integrates Laplacian eigenvectors as geometric priors for improved structural awareness on irregular, unseen geometries and employs an Enhanced Deep Super-Resolution Network (EDSR) layer for robust upsampling. By combining geometric priors with neural operator frameworks, LoFNO de-noises and spatiotemporally upsamples flow data, achieving superior velocity and WSS predictions compared to interpolation and alternative deep learning methods, enabling more precise cerebrovascular diagnostics.

Figures

Figures reproduced from arXiv: 2507.13789 by the authors.

Figure 1
Figure 1. Method pipeline: Geometry and boundary conditions are extracted from cere￾bral angiograms. Simulations are run, the relevant section is segmented, Laplacian eigenvectors are computed, and noise is added. Flow and eigenvectors are sampled on a voxel grid for preprocessing. The model is trained to predict noise-free flow on a high￾resolution grid and tested on unseen data. In clinical use, it can be applied directly t… view at source ↗
Figure 2
Figure 2. Computational graph for LoFNO. Both e(x), a selection of eigenvectors, and u(x), the flow velocity with added noise, are sampled on a regular grid, while uˆ(x) (pre￾diction for noiseless flow), χ(x) (domain characteristic function), and hl(x) (Fourier layer input/output, l ∈ {0, . . . , L}) are sampled on a higher resolution grid. The lifting layer P and projection layer Q are MLPs, mapping (3 + Ne) to dh and dh to … view at source ↗
Figure 3
Figure 3. Time-snapshot of hemodynamics upsampling for Scale ×4: Top: Flow velocity components u, v, and w in the x, y, and z directions for five models compared against the ground truth. Bottom: WSS evaluated on the surface geometry, with the length of vectors colored black for the smallest to yellow for the largest. We selected 80 aneurysm-containing regions, computed Laplacian eigenvec￾tors on the surface mesh, and resampl… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

  1. [1]

    CoRRabs/2104.13478(2021), https://arxiv.org/abs/2104.13478

    Bronstein, M.M., Bruna, J., Cohen, T., Velickovic, P.: Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. CoRRabs/2104.13478(2021), https://arxiv.org/abs/2104.13478

    arXiv 2021

  2. [2]

    Acta Mechanica Sinica37(12), 1727–1738 (2021)

    Cai, S., Mao, Z., Wang, Z., Yin, M., Karniadakis, G.E.: Physics-informed neural networks (pinns) for fluid mechanics: A review. Acta Mechanica Sinica37(12), 1727–1738 (2021)

    work page 2021

  3. [3]

    Neurology International16(1), 74–94 (2024)

    Deshmukh, A.S., Priola, S.M., Katsanos, A.H., Scalia, G., Costa Alves, A., Srivas- tava, A., Hawkes, C.: The management of intracranial aneurysms: Current trends and future directions. Neurology International16(1), 74–94 (2024)

    work page 2024

  4. [4]

    IEEE transactions on pattern analysis and machine intelligence 38(2), 295–307 (2015)

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convo- lutional networks. IEEE transactions on pattern analysis and machine intelligence 38(2), 295–307 (2015)

    2015

  5. [5]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    Flouris, K., Konukoglu, E.: Canonical normalizing flows for manifold learning. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 27294–27314. Curran Associates, Inc. (2023)

    work page 2023

  6. [6]

    Flouris, K., Volokitin, A., Bredell, G., Konukoglu, E.: Explicit and data-efficient encoding via gradient flow (2025), https://arxiv.org/abs/2412.00864

    work page Pith review arXiv 2025

  7. [7]

    International Journal of Surgery86, 15–23 (2021)

    Han, P., Jin, D., Wei, W., Song, C., Leng, X., Liu, L., Yu, J., Li, X.: The prog- nostic effects of hemodynamic parameters on rupture of intracranial aneurysm: A systematic review and meta-analysis. International Journal of Surgery86, 15–23 (2021)

    work page 2021

  8. [8]

    Radiologic Clinics 58(4), 815–830 (2020) 10 K

    Hussein, A., Malguria, N.: Imaging of vascular malformations. Radiologic Clinics 58(4), 815–830 (2020) 10 K. Flouris et al

    work page 2020

Show all 29 references
  1. [9]

    In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super- resolution using a generative adversarial network. In: 2017 IEEE Conference on Computer Vision and Pattern Reco...

    2017

  2. [10]

    In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T

    Li, X.J., Yang, J., Zhang, F.L.: Laplacian mesh transformer: Dual attention and topology aware network for 3d mesh classification and segmentation. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision – ECCV 2022. pp. 541–560. Springer Na...

    2022

  3. [11]

    Li, Z., Huang, D.Z., Liu, B., Anandkumar, A.: Fourier neural operator with learned deformationsforpdesongeneralgeometries.JournalofMachineLearningResearch 24(388), 1–26 (2023)

    2023

  4. [12]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.: Fourier neural operator for parametric partial differential equa- tions (2021)

    2021

  5. [13]

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

    Lim,B.,Son,S.,Kim,H.,Nah,S.,MuLee,K.:Enhanceddeepresidualnetworksfor single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136–144 (2017)

    2017

  6. [14]

    Advances in Neural Information Processing Systems36(2024)

    Liu, N., Jafarzadeh, S., Yu, Y.: Domain agnostic fourier neural operators. Advances in Neural Information Processing Systems36(2024)

    2024

  7. [15]

    Communi- cations Physics7(1) (Jan 2024)

    Liu,X.Y., Zhu, M., Lu,L., Sun, H.,Wang, J.X.: Multi-resolutionpartial differential equations preserved learning framework for spatiotemporal dynamics. Communi- cations Physics7(1) (Jan 2024)

    2024

  8. [16]

    Journal of the Indian Institute of Science pp

    Maramkandam, E.B., Kannan, A., Valeti, C., Manjunath, N., Panneerselvam, N.K., Alagan, A.K., Panchal, P.M., Kannath, S.K., Darshan, H., Nekkanti, R.K., et al.: Review of cfd based simulations to study the hemodynamics of cerebral aneurysms. Journal of the Indian Institute of S...

    2024

  9. [17]

    Frontiers in Cardiovascular Medicine9, 793072 (2022)

    Maupu, C., Lebas, H., Boulaftali, Y.: Imaging modalities for intracranial aneurysm: more than meets the eye. Frontiers in Cardiovascular Medicine9, 793072 (2022)

    2022

  10. [18]

    Journal of Cerebral Blood Flow & Metabolism41(2), 206–218 (2021)

    Morgan, A.G., Thrippleton, M.J., Wardlaw, J.M., Marshall, I.: 4d flow mri for non-invasive measurement of blood flow in the brain: a systematic review. Journal of Cerebral Blood Flow & Metabolism41(2), 206–218 (2021)

    2021

  11. [19]

    World Neurosurgery179, 68–76 (2023)

    Nico, E., Hossa, J., McGuire, L.S., Alaraj, A.: Rupture-risk stratifying patients with cerebral arteriovenous malformations using quantitative hemodynamic flow measurements. World Neurosurgery179, 68–76 (2023)

    2023

  12. [20]

    Journal of Cerebral Blood Flow & Metabolism41(12), 3141–3156 (2021)

    Pan, P., Weinsheimer, S., Cooke, D., Winkler, E., Abla, A., Kim, H., Su, H.: Review of treatment and therapeutic targets in brain arteriovenous malformation. Journal of Cerebral Blood Flow & Metabolism41(12), 3141–3156 (2021)

    2021

  13. [21]

    Electronic Journal of Statistics8(2), 1879 – 1890 (2014)

    Sangalli, L.M., Secchi, P., Vantini, S.: AneuRisk65: A dataset of three-dimensional cerebral vascular geometries. Electronic Journal of Statistics8(2), 1879 – 1890 (2014)

    2014

  14. [22]

    Journal of Complex Networks8(2), cnaa007 (Mar 2020)

    Torres, L., Chan, K.S., Eliassi-Rad, T.: GLEE: Geometric laplacian eigenmap em- bedding. Journal of Complex Networks8(2), cnaa007 (Mar 2020)

    2020

  15. [23]

    The Lancet 369(9558), 306–318 (2007)

    Van Gijn, J., Kerr, R.S., Rinkel, G.J.: Subarachnoid haemorrhage. The Lancet 369(9558), 306–318 (2007)

    2007

  16. [24]

    Wang, R., Kashinath, K., Mustafa, M., Albert, A., Yu, R.: Towards physics- informed deep learning for turbulent flow prediction (2020)

    2020

  17. [25]

    In: Mukhopadhyay, A., Oksuz, I., Engelhardt, S., Mehrof, D., Yuan, Y

    Wang, Y., Lee, Y.Y.R., Dolfini, A., Reischl, M., Konukoglu, E., Flouris, K.: Energy- basedpriorlatentspacediffusionmodelforreconstructionoflumbarvertebraefrom thick slice mri. In: Mukhopadhyay, A., Oksuz, I., Engelhardt, S., Mehrof, D., Yuan, Y. (eds.) Deep Generative Models. ...

    2025

  18. [26]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wei, M., Zhang, X.: Super-Resolution Neural Operator . In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18247– 18256. IEEE Computer Society, Los Alamitos, CA, USA (Jun 2023)

    2023

  19. [27]

    Computer Methods in Applied Mechanics and Engineering398, 115296 (2022)

    You, H., Zhang, Q., Ross, C.J., Lee, C.H., Yu, Y.: Learning deep implicit fourier neural operators (ifnos) with applications to heterogeneous material modeling. Computer Methods in Applied Mechanics and Engineering398, 115296 (2022)

    2022

  20. [28]

    arXiv preprint arXiv:2211.15188 (2022)

    Zhao, J., George, R.J., Zhang, Y., Li, Z., Anandkumar, A.: Incremental fourier neural operator. arXiv preprint arXiv:2211.15188 (2022)

    2022 arXiv

  21. [29]

    Scientific Re- ports7(1), 5331 (2017)

    Zhou, G., Zhu, Y., Yin, Y., Su, M., Li, M.: Association of wall shear stress with intracranial aneurysm rupture: systematic review and meta-analysis. Scientific Re- ports7(1), 5331 (2017)

    2017

Pith tools

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