arxiv: 2601.15416 · v2 · submitted 2026-01-21 · 💻 cs.CV · cs.AI

Recognition: 1 theorem link

· Lean Theorem

DuFal: Dual-Frequency-Aware Learning for High-Fidelity Extremely Sparse-view CBCT Reconstruction

Cuong Tran Van , Trong-Thang Pham , Ngoc-Son Nguyen , Duy Minh Ho Nguyen , Ngan Le

Authors on Pith no claims yet

Pith reviewed 2026-05-16 11:58 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords sparse-view reconstructioncone-beam CThigh-frequency learningFourier neural operatormedical image reconstructiondual-path architecturecross-attention fusion

0 comments

The pith

DuFal recovers fine anatomical details in CT scans from extremely few X-ray projections using dual frequency processing.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes DuFal, a framework for reconstructing high-fidelity cone-beam CT images from very limited X-ray projections. It addresses the issue that conventional methods lose high-frequency details due to bias toward low frequencies. By using a dual-path architecture with global and local high-frequency Fourier neural operators, plus cross-attention fusion, it integrates frequency and spatial information. Experiments show superior performance on lung and dental datasets in sparse-view settings. This matters because it could enable lower radiation doses in medical imaging while maintaining diagnostic quality.

Core claim

DuFal integrates frequency-domain and spatial-domain processing through a High-Local Factorized Fourier Neural Operator consisting of global and local high-frequency enhanced branches, combined via cross-attention, to recover high-frequency anatomical features from undersampled projections and reconstruct accurate CT volumes.

What carries the argument

The High-Local Factorized Fourier Neural Operator, which uses global frequency pattern capture and local patch processing to preserve spatial details lost in global analysis.

Load-bearing premise

The global and local high-frequency branches combined with cross-attention will recover fine details from limited projections without creating artifacts on unseen clinical scans.

What would settle it

Running DuFal on a new clinical dataset with ground-truth dense projections and checking if fine structures like small vessels or tooth details match the full-view reconstruction without extra blurring or false features.

Figures

Figures reproduced from arXiv: 2601.15416 by Cuong Tran Van, Duy Minh Ho Nguyen, Ngan Le, Ngoc-Son Nguyen, Trong-Thang Pham.

**Figure 2.** Figure 2: Overview of the proposed HiLocFFNO block within the Frequency Encoder. Each HiLocFFNO block comprises two major branches: The gHiF is illustrated at the top using solid blue line (—) and arrows (→), and the lHiF is depicted at the bottom using dashed blue line (- - -) and arrows (99K). Both branches employ SCF to replace the full complex weight with separate channel-mixing and spectral-weighting kernels. A… view at source ↗

**Figure 3.** Figure 3: The CAFF flowchart. This diagram shows the fusion process between spatial features [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Visualization of 10-view reconstructed chest CT (from top to bottom: axial, coronal, and sagittal [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Visualization of 10-view reconstructed dental CT (from top to bottom: axial, coronal, and sagittal [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 6.** Figure 6: Performance vs. Inference Speed Trade-off on LUNA16 Dataset (6-View). Scatter plots comparing reconstruction quality against computational efficiency for different methods. Left: PSNR vs. Inference speed. Right: SSIM vs. Inference speed. DuFal achieves the highest reconstruction quality in both metrics while maintaining good inference speed. PSNR: 12.12 W-PSNR: 12.12 PSNR: 11.78 W-PSNR: 38.15 PSNR: 10.59 W… view at source ↗

**Figure 7.** Figure 7: (a) Original image and region of interest in mask format; (b) Two types of noise: Gaussian noise [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Comparison of LungMask segmentation performance on full-view versus 10-view reconstructed CT [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

read the original abstract

Sparse-view Cone-Beam Computed Tomography reconstruction from limited X-ray projections remains a challenging problem in medical imaging due to the inherent undersampling of fine-grained anatomical details, which correspond to high-frequency components. Conventional CNN-based methods often struggle to recover these fine structures, as they are typically biased toward learning low-frequency information. To address this challenge, this paper presents DuFal (Dual-Frequency-Aware Learning), a novel framework that integrates frequency-domain and spatial-domain processing via a dual-path architecture. The core innovation lies in our High-Local Factorized Fourier Neural Operator, which comprises two complementary branches: a Global High-Frequency Enhanced Fourier Neural Operator that captures global frequency patterns and a Local High-Frequency Enhanced Fourier Neural Operator that processes spatially partitioned patches to preserve spatial locality that might be lost in global frequency analysis. To improve efficiency, we design a Spectral-Channel Factorization scheme that reduces the Fourier Neural Operator parameter count. We also design a Cross-Attention Frequency Fusion module to integrate spatial and frequency features effectively. The fused features are then decoded through a Feature Decoder to produce projection representations, which are subsequently processed through an Intensity Field Decoding pipeline to reconstruct a final Computed Tomography volume. Experimental results on the LUNA16 and ToothFairy datasets demonstrate that DuFal significantly outperforms existing state-of-the-art methods in preserving high-frequency anatomical features, particularly under extremely sparse-view settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DuFal's dual global/local high-frequency FNO branches with cross-attention fusion look like a targeted fix for sparse-view CBCT detail loss, but the abstract supplies zero numbers or ablations to show the gains are real.

read the letter

The paper's core move is splitting high-frequency recovery into a global FNO branch and a local patch-wise FNO branch, then using spectral-channel factorization to keep the parameter count down and cross-attention to blend the streams before decoding to the CT volume. That specific pairing is not in the earlier single-branch FNO work for this task, so the architecture itself is the new piece. It correctly identifies that standard CNNs drop high-frequency anatomical edges under extreme undersampling and tries to counter it with explicit frequency paths rather than hoping a deeper network will learn them. The description of the Intensity Field Decoding pipeline is also straightforward and ties the projection features back to the final volume in a way that matches how CBCT is actually acquired. Those parts read cleanly and give a reviewer something concrete to check. The main weakness is the evidence. The abstract claims clear outperformance on LUNA16 and ToothFairy for high-frequency preservation, yet it contains no PSNR, SSIM, or error-bar numbers, no ablation table, and no mention of training details or statistical tests. Without those, the central claim cannot be assessed. The stress-test concern about unseen clinical data is also on point: both datasets are public and relatively clean; nothing in the provided text indicates cross-scanner or real-acquisition validation, which is where frequency-specific artifacts usually appear. This work is aimed at researchers already using neural operators for medical reconstruction who need a practical way to keep fine bone or vessel detail at low dose. A reader who wants to implement the dual-branch idea would find the description usable, but anyone needing proven clinical reliability would still be waiting for the numbers. It is worth sending to peer review because the architecture is specific enough to critique and the problem is real, but the referees will need to insist on full quantitative results and at least one external test set before acceptance.

Referee Report

2 major / 2 minor

Summary. The paper proposes DuFal, a dual-frequency-aware framework for extremely sparse-view CBCT reconstruction that combines a High-Local Factorized Fourier Neural Operator (with global and local high-frequency branches) and a Cross-Attention Frequency Fusion module to recover high-frequency anatomical details that CNNs typically miss, followed by an Intensity Field Decoding pipeline. Experiments on LUNA16 and ToothFairy datasets are claimed to show significant outperformance over SOTA methods in preserving fine structures under extreme undersampling.

Significance. If the quantitative results and ablations hold, the work would represent a meaningful advance in frequency-aware reconstruction for low-dose CBCT, potentially enabling reliable recovery of diagnostic high-frequency features with fewer projections and lower radiation dose.

major comments (2)

[§4] §4 (Experiments) and Table 1: the central claim of significant outperformance lacks reported PSNR/SSIM values, error bars, ablation studies on the global vs. local branches, or training details; without these the magnitude and robustness of the improvement cannot be assessed.
[§3.2] §3.2 (Cross-Attention Frequency Fusion) and §5 (Discussion): the assumption that the dual-branch FNO plus fusion recovers fine details without introducing hallucinations or new artifacts on unseen clinical data is load-bearing but untested; no cross-dataset evaluation or real-acquisition protocol results are provided despite known domain-shift risks in CBCT frequency content.

minor comments (2)

[§3.1] Notation for the Spectral-Channel Factorization scheme is introduced without an explicit equation or complexity analysis; adding a parameter-count comparison table would clarify the efficiency claim.
[Figure 3] Figure 3 caption and axis labels should explicitly state the number of views used in each sparse-view setting for reproducibility.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We have carefully reviewed the feedback and provide point-by-point responses below, outlining the revisions we will implement to address the concerns raised.

read point-by-point responses

Referee: [§4] §4 (Experiments) and Table 1: the central claim of significant outperformance lacks reported PSNR/SSIM values, error bars, ablation studies on the global vs. local branches, or training details; without these the magnitude and robustness of the improvement cannot be assessed.

Authors: We agree that the current presentation of results in §4 and Table 1 requires strengthening for full transparency. In the revised version, we will expand Table 1 to report mean PSNR and SSIM values accompanied by standard deviation error bars computed over multiple training runs with different random seeds. We will also insert a new ablation subsection that isolates the performance of the global high-frequency branch versus the local high-frequency branch (and their combination), including quantitative metrics and qualitative visualizations. Finally, we will add a dedicated paragraph (or supplementary section) detailing all training hyperparameters, including optimizer choice, learning-rate schedule, batch size, number of epochs, and data augmentation strategy. revision: yes
Referee: [§3.2] §3.2 (Cross-Attention Frequency Fusion) and §5 (Discussion): the assumption that the dual-branch FNO plus fusion recovers fine details without introducing hallucinations or new artifacts on unseen clinical data is load-bearing but untested; no cross-dataset evaluation or real-acquisition protocol results are provided despite known domain-shift risks in CBCT frequency content.

Authors: We acknowledge that explicit validation against domain shift and potential hallucinations is important. Although LUNA16 and ToothFairy already span distinct anatomical domains (pulmonary versus dental), we agree that dedicated cross-dataset experiments (training on one dataset and testing on the other) are missing. In the revision we will add these cross-dataset results together with frequency-domain residual analysis to check for spurious high-frequency content. We will also expand §5 with a more thorough discussion of hallucination risks and domain-shift mitigation strategies. However, because we do not currently possess raw real-acquisition CBCT projection data acquired under clinical protocols, we can only discuss this limitation rather than present new empirical results on such data. revision: partial

standing simulated objections not resolved

Absence of real clinical acquisition protocol results for final validation (we can discuss the limitation but cannot generate new empirical results on such data within the revision timeframe).

Circularity Check

0 steps flagged

No significant circularity in architectural proposal or claims

full rationale

The paper presents DuFal as a new dual-path neural architecture incorporating a High-Local Factorized Fourier Neural Operator (with global and local branches), Spectral-Channel Factorization, and Cross-Attention Frequency Fusion, followed by a Feature Decoder and Intensity Field Decoding pipeline. These elements are introduced as explicit design choices rather than mathematical derivations or fitted quantities. No equations or steps in the abstract reduce claimed performance gains to parameters tuned on the same data or to self-citations that bear the central load. Experimental results on LUNA16 and ToothFairy are presented as independent empirical validation of the architecture's ability to preserve high-frequency features, not as predictions forced by construction from the inputs. The derivation chain is therefore self-contained as an engineering proposal plus benchmark evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The framework rests on the assumption that Fourier Neural Operators are suitable for modeling frequency content in projection data and that high-frequency anatomical features can be recovered by explicit dual-path processing; no new physical entities are postulated.

axioms (1)

domain assumption Fourier Neural Operators can capture frequency patterns in medical imaging data
Invoked when the paper builds the Global and Local High-Frequency Enhanced Fourier Neural Operators on top of existing FNO literature.

invented entities (2)

High-Local Factorized Fourier Neural Operator no independent evidence
purpose: To jointly model global frequency patterns and spatially local high-frequency details
New architectural component introduced to address limitations of standard FNOs in sparse-view settings
Cross-Attention Frequency Fusion module no independent evidence
purpose: To integrate spatial and frequency-domain features
New fusion mechanism described in the abstract

pith-pipeline@v0.9.0 · 5563 in / 1423 out tokens · 21040 ms · 2026-05-16T11:58:44.606848+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean, IndisputableMonolith/Foundation/DimensionForcing.lean washburn_uniqueness_aczel, alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

High-Local Factorized Fourier Neural Operator... Global High-Frequency Enhanced Fourier Neural Operator... Local High-Frequency Enhanced Fourier Neural Operator... Spectral-Channel Factorization... Cross-Attention Frequency Fusion

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 3 internal anchors

[1]

doi: https: //doi.org/10.1016/0161-7346(84)90008-7

ISSN 0161-7346. doi: https: //doi.org/10.1016/0161-7346(84)90008-7. URL https://www.sciencedirect.com/science/article/ pii/0161734684900087. Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. InCVPR,

work page doi:10.1016/0161-7346(84)90008-7
[2]

Abril Corona-Figueroa, Jonathan Frawley, Sam Bond Taylor, Sarath Bethapudi, Hubert P

doi: 10.1109/ ACCESS.2022.3144840. Abril Corona-Figueroa, Jonathan Frawley, Sam Bond Taylor, Sarath Bethapudi, Hubert P. H. Shum, and Chris G. Willcocks. Mednerf: Medical neural radiance fields for reconstructing 3d-aware ct-projections from a single x-ray. In2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (...

work page arXiv 2022
[3]

Real-time Pilot Crew’s Mental Workload and Arousal Assessment During Simulated Flights for Training Evaluation: a Case Study,

doi: 10.1109/EMBC48229.2022.9871757. Esther Decabooter, Guido C. Hilgers, Joke De Rouck, Koen Salvo, Jacobus Van Wingerden, Hilde Bosmans, Brent van der Heyden, Sima Qamhiyeh, Chrysi Papalazarou, Robert Kaatee, Geert Pittomvils, and Evelien Bogaert. Survey on fan-beam computed tomography for radiotherapy: Imaging for dose calculation and delineation.Physi...

work page doi:10.1109/embc48229.2022.9871757 2022
[4]

doi: https://doi.org/10.1016/j.phro.2023.100522

ISSN 2405-6316. doi: https://doi.org/10.1016/j.phro.2023.100522. URL https://www.sciencedirect.com/science/article/ pii/S2405631623001136. L. A. Feldkamp, L. C. Davis, and J. W. Kress. Practical cone-beam algorithm.Journal of the Optical Society of America A, 1(6):612–619, June

work page doi:10.1016/j.phro.2023.100522 2023
[5]

Deep Residual Learning for Compressed Sensing CT Reconstruction via Persistent Homology Analysis

doi: 10.1364/JOSAA.1.000612. Yo Seob Han, Jaejun Yoo, and Jong Chul Ye. Deep residual learning for compressed sensing ct reconstruction via persistent homology analysis.arXiv preprint arXiv:1611.06391,

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1364/josaa.1.000612
[6]

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision appli- cations.ArXiv, abs/1704.04861,

work page internal anchor Pith review Pith/arXiv arXiv
[7]

Yun Su Jeong, Hye Bin Yoo, and Il Yong Chun

URLhttps://api.semanticscholar.org/CorpusID:12670695. Yun Su Jeong, Hye Bin Yoo, and Il Yong Chun. Dx2ct: Diffusion model for 3d ct reconstruction from bi or mono-planar 2d x-ray(s). InICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5,

work page 2025
[8]

W., Gorp, H

doi: 10.1109/ICASSP49660.2025.10888986. 18 Published in Transactions on Machine Learning Research (12/2025) Kyong Hwan Jin, Michael T. McCann, Emmanuel Froustey, and Michael Unser. Deep convolutional neural network for inverse problems in imaging.IEEE Transactions on Image Processing, 26(9):4509–4522,

work page doi:10.1109/icassp49660.2025.10888986 2025
[9]

Marimuthu Kalimuthu, David Holzmuller, and Mathias Niepert

doi: 10.1109/TIP.2017.2713099. Marimuthu Kalimuthu, David Holzmuller, and Mathias Niepert. Loglo-fno: Efficient learning of local and global features in fourier neural operators,

work page doi:10.1109/tip.2017.2713099 2017
[10]

Mahyar Khayatkhoei and Ahmed Elgammal

URLhttps://arxiv.org/abs/2504.04260. Mahyar Khayatkhoei and Ahmed Elgammal. Spatial frequency bias in convolutional generative adversarial networks.Proceedings of the AAAI Conference on Artificial Intelligence, 36(7):7152–7159, Jun

work page arXiv
[11]

URLhttps://ojs.aaai.org/index.php/AAAI/article/view/20675

doi: 10.1609/aaai.v36i7.20675. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/20675. Daeun Kyung, Kyungmin Jo, Jaegul Choo, Joonseok Lee, and Edward Choi. Perspective projection-based 3d ct reconstruction from biplanar x-rays. InICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5,

work page doi:10.1609/aaai.v36i7.20675 2023
[12]

Yunxiang Li, Hua-Chieh Shao, Xiao Liang, Liyuan Chen, Ruiqi Li, Steve Jiang, Jing Wang, and You Zhang

doi: 10.1109/ICASSP49357.2023.10096296. Yunxiang Li, Hua-Chieh Shao, Xiao Liang, Liyuan Chen, Ruiqi Li, Steve Jiang, Jing Wang, and You Zhang. Zero-shot medical image translation via frequency-guided diffusion models.IEEE transactions on medical imaging, 43(3):980–993, 2023a. Yunxiang Li, Hua-Chieh Shao, Xiaoxue Qian, and You Zhang. Fddm: Unsupervised med...

work page doi:10.1109/icassp49357.2023.10096296 2023
[13]

Fourier Neural Operator for Parametric Partial Differential Equations

URLhttps://arxiv.org/abs/2010.08895. Wei-An Lin, Haofu Liao, Cheng Peng, Xiaohang Sun, Jingdan Zhang, Jiebo Luo, Rama Chellappa, and Shaohua Kevin Zhou. Dudonet: Dual domain network for ct metal artifact reduction. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10504–10513,

work page internal anchor Pith review Pith/arXiv arXiv 2010
[14]

Yiqun Lin, Zhongjin Luo, Wei Zhao, and Xiaomeng Li

doi: 10.1109/ CVPR.2019.01076. Yiqun Lin, Zhongjin Luo, Wei Zhao, and Xiaomeng Li. Learning deep intensity field for extremely sparse-view cbct reconstruction. In Hayit Greenspan, Anant Madabhushi, Parvin Mousavi, Septimiu Salcudean, James Duncan, Tanveer Syeda-Mahmood, and Russell Taylor (eds.),Medical Image Computing and Computer Assisted Intervention –...

work page arXiv 2019
[15]

ISBN 978-3-031-43999-5

Springer Nature Switzerland. ISBN 978-3-031-43999-5. Yiqun Lin, Hualiang Wang, Jixiang Chen, and Xiaomeng Li. Learning 3D Gaussians for Extremely Sparse- View Cone-Beam CT Reconstruction . Inproceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, volume LNCS 15007. Springer Nature Switzerland, October 2024a. Yiqun Lin, Jie...

work page 2024
[16]

URLhttps://doi.org/10.24963/ijcai.2022/101

doi: 10.24963/ijcai.2022/101. URLhttps://doi.org/10.24963/ijcai.2022/101. Main Track. Jiachen Liu and Xiangzhi Bai. VolumeNeRF: CT Volume Reconstruction from a Single Projection View . In proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, volume LNCS 15007. Springer Nature Switzerland, October

work page doi:10.24963/ijcai.2022/101 2022
[17]

Image reconstruction for accelerated mr scan with faster fourier convolutional neural networks.IEEE Transactions on Image Processing, 2024a

19 Published in Transactions on Machine Learning Research (12/2025) Xiaohan Liu, Yanwei Pang, Xuebin Sun, Yiming Liu, Yonghong Hou, Zhenchang Wang, and Xuelong Li. Image reconstruction for accelerated mr scan with faster fourier convolutional neural networks.IEEE Transactions on Image Processing, 2024a. Xuhui Liu, Zhi Qiao, Runkun Liu, Hong Li, Juan Zhang...

work page 2025
[18]

Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister

doi: 10.1007/978-3-031-43999-5_24. Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister. Dynamic high-pass filtering and multi-spectral attention for image super-resolution. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4268–4277,

work page doi:10.1007/978-3-031-43999-5_24
[19]

2021.00425

doi: 10.1109/ICCV48922. 2021.00425. Duy M. H. Nguyen, Hoang Nguyen, Nghiem T. Diep, Tan N. Pham, Tri Cao, Binh T. Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, Daniel Sonntag, and Mathias Niepert. Lvm-med: learning large-scale self-supervised vision models for medical imaging via second-order graph matching. In Proceedings of the 37th Inte...

work page doi:10.1109/iccv48922 2021
[20]

Image processing for 3d reconstruction using a modified fourier transform profilometry method

Jesus Carlos Pedraza Ortega, Jose Wilfrido Rodriguez Moreno, Leonardo Barriga Rodriguez, Efren Gorrosti- eta Hurtado, Tomas Salgado Jimenez, Juan Manuel Ramos Arreguin, and Angel Rivas. Image processing for 3d reconstruction using a modified fourier transform profilometry method. InMICAI 2007: Advances in Artificial Intelligence: 6th Mexican International...

work page 2007
[21]

doi: https://doi.org/10.1016/j.media.2017.06.015

ISSN 1361-8415. doi: https://doi.org/10.1016/j.media.2017.06.015. URLhttps: //www.sciencedirect.com/science/article/pii/S1361841517301020. Liyue Shen, John Pauly, and Lei Xing. Nerp: implicit neural representation learning with prior embedding for sparsely sampled image reconstruction.IEEE Transactions on Neural Networks and Learning Systems, 2022a. Liyue...

work page doi:10.1016/j.media.2017.06.015 2017
[22]

Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information processing systems, 33:7537–7547,

20 Published in Transactions on Machine Learning Research (12/2025) Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information processing system...

work page 2025
[23]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pp

doi: 10.1109/CVPR42600.2020.00181. Xingde Ying, Heng Guo, Kai Ma, Jian Wu, Zhengxin Weng, and Yefeng Zheng. X2ct-gan: Reconstructing ct from biplanar x-rays with generative adversarial networks. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10611–10620,

work page doi:10.1109/cvpr42600.2020.00181 2020
[24]

Ruyi Zha, Yanhao Zhang, and Hongdong Li

doi: 10.1109/CVPR.2019.01087. Ruyi Zha, Yanhao Zhang, and Hongdong Li. Naf: Neural attenuation fields for sparse-view cbct reconstruction. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 442–452. Springer, 2022a. Ruyi Zha, Yanhao Zhang, and Hongdong Li. Naf: Neural attenuation fields for sparse-view cbct recon...

work page doi:10.1109/cvpr.2019.01087 2019