Pith. sign in

REVIEW 5 major objections 5 minor 54 references

FD-DiT: Frequency Domain-Directed Diffusion Transformer for Low-Dose CT Reconstruction

T0 review · 5 major / 5 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 A frequency-split diffusion transformer improves low-dose CT reconstruction by separating noise from anatomy before denoising.

desk verdict Consistent empirical gains on two LDCT benchmarks, but the reconstruction-stage equations in the paper are internally inconsistent—the code, not the math, is the real spec. read the letter →

arxiv 2506.23466 v1 pith:VFXE6O3D submitted 2025-06-30 eess.IV cs.CVphysics.med-ph

Qiqing Liu , Guoquan Wei , Zekun Zhou , Yiyang Wen , Liu Shi , Qiegen Liu This is my paper · ORCID
classification eess.IVcs.CVphysics.med-ph
keywords low-doseCTdiffusionmodeltransformerfrequencydecompositionsinogramdenoisingsparselocalattentionimagereconstruction
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 FD-DiT, a diffusion-model method for low-dose CT reconstruction that works in the sinogram (projection) domain. Its central claim is that decomposing sinogram data into high-, low-, and full-frequency bands, then denoising each band with a dedicated module, suppresses quantum and electronic noise while preserving fine anatomical detail better than existing methods. The authors report the highest PSNR/SSIM and lowest MSE across all tested dose levels on both the public LDCT grand-challenge dataset and an anthropomorphic phantom dataset. A sympathetic reader would care because this targets the real trade-off in low-dose CT: reducing radiation exposure without losing diagnostically relevant structures.

What carries the argument

The load-bearing object is the generalized diffusion process defined by $x_t = \mathcal{P}(x_0, x_T, t) = \alpha_t x_0 + (1-\alpha_t) x_T$, where $x_T$ is the low-dose sinogram distribution. Each reverse step applies a Fourier-domain Gaussian decomposition, $G_L = \mathcal{F}^{-1}(\mathcal{F}(x_t) \odot \mathcal{K}_\sigma)$ and $G_H = \mathcal{F}(x_t) - \mathcal{F}(x_t) \odot \mathcal{K}_\sigma$, to produce low-, high-, and full-frequency branches processed by FLD, FHD, and FFD. The FHD branch couples multi-head self-attention with sliding sparse local attention (SSLA), a token-level sparse windowed attention with dilation rates $r=1,2,3$, to catch localized high-frequency noise; the LDF module then learns how to fuse the three denoised bands. This machinery is what lets the network isolate noise from anatomy and recombine components without fixed weights.

What would settle it

Measure the noise power spectrum of paired low-dose and full-dose sinograms after the Gaussian split of Eq. (10); if a substantial fraction of the noise variance lands in the low-frequency component $G_L$, or if the high-frequency branch alone accounts for most of the PSNR gain in the ablation, then the frequency-separation premise fails and the three-branch design loses its justification.

Watch

Extended reading notes

Core claim

FD-DiT's central claim is that frequency-directed decoupling improves diffusion-based LDCT reconstruction. Instead of diffusing toward Gaussian noise, the forward process interpolates from normal-dose CT sinograms toward the empirical low-dose CT distribution, and the reverse process denoises three Fourier-decomposed bands: a transformer-based high-frequency module, a U-Net low-frequency module, and a U-Net full-frequency compensation branch. Their outputs are combined by a learnable dynamic fusion network, and iterative reconstruction applies penalized weighted least-squares and total-variation constraints at every step. On both evaluation datasets, at all three photon-dose levels, FD-DiT reports the best PSNR, SSIM, and MSE among all compared methods, with the largest margins at the lowest dose.

Load-bearing premise

The load-bearing premise is that a Gaussian high/low frequency split of the sinogram separates noise from anatomy well enough that denoising the bands independently helps; the paper supports this only with a qualitative illustration.

Editorial extensions

If this is right

  • At identical dose, FD-DiT reports the highest PSNR/SSIM and lowest MSE on the public LDCT challenge dataset at 1e4, 5e4, and 1e5 photons.
  • The same model, applied without fine-tuning to an anthropomorphic phantom dataset, also leads all compared methods at every dose level, supporting cross-dataset generalization.
  • Ablation results show that the full FFD+FLD+FHD composition beats any subset, so each frequency branch contributes independently to reconstruction quality.
  • Replacing global-only attention with SSLA and replacing fixed summation fusion with LDF each yield measurable improvement, validating the two architectural choices.
  • Because diffusion targets the low-dose distribution rather than pure Gaussian noise, the reverse process avoids large sampling schedules while retaining detail.

Reading between the lines

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

  • A natural next test is to measure the actual noise power spectrum of paired low-dose and full-dose sinograms; if a large share of noise energy sits below the Gaussian cutoff, the architecture would need an adaptive bandwidth $\sigma$ rather than a fixed one.
  • The frequency-decoupling-plus-learnable-fusion recipe could transfer to other inverse problems with band-limited noise, such as low-dose PET, MRI undersampling, or photon-counting CT, where noise statistics differ by acquisition physics.
  • The authors note the transformer lacks downsampling between blocks; adding a hierarchical multi-scale pyramid to the high-frequency branch is a direct extension that could push the low-dose regime further.
  • If the reported margins hold in a reader study, the method's clinical relevance would be judged by whether radiologists can extract diagnostic information at the lowest photon count, not just by PSNR.
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 FD-DiT, a frequency-domain-directed diffusion transformer for low-dose CT (LDCT) sinogram denoising and reconstruction. The method decomposes sinogram data into high-, low-, and full-frequency components via Gaussian filtering, processes them with a transformer-based high-frequency denoising module (FHD), a U-Net-based low-frequency module (FLD), and a full-frequency U-Net (FFD), and fuses the outputs with a learnable dynamic fusion (LDF) module. A generalized diffusion process is trained to map normal-dose CT data toward the LDCT data distribution, and reconstruction adds PWLS and TV regularization steps at each reverse diffusion iteration. The paper reports PSNR/SSIM/MSE results on the AAPM Challenge dataset and a CIRS phantom at three dose levels, claiming state-of-the-art performance, and includes ablation studies on the frequency modules, attention mechanism, and fusion strategy.

Significance. If the results and method description were both sound, the paper would make a useful empirical contribution: it combines frequency-aware multi-branch processing with a diffusion transformer, releases code, and compares against several recent methods on two datasets. The ablations support the value of the three-branch design and of the sliding sparse local attention. However, the significance is substantially tempered by the single-patient AAPM test set, the absence of error bars or statistical tests in the main table, and several mathematical inconsistencies in the description of the forward and reconstruction processes. The central empirical claim may well be reproducible from the released code, but the paper as written does not provide a correct and complete specification of the algorithm it claims to evaluate.

major comments (5)
  1. [Sec. III-D, Eq. (24) and Algorithm 1, line 5] The update x_{t-1} = [W(y - x_t) + μR'(x_t)]/(W + μ) is not the minimizer of the PWLS objective in Eq. (23); the first-order condition of Eq. (23) is W(x - y) + μR'(x) = 0, and in the high-confidence limit W → ∞ the stated update approaches y - x_t rather than y, so the reconstruction loop drives iterates away from the measured sinogram and does not implement the claimed data fidelity.
  2. [Sec. III-D, Eq. (26) and Algorithm 1, line 6] The TV update is not well defined: x_{t-1} appears on both sides of the equation, the norm ‖x - x_{t-1}‖ uses an undefined variable x, and the expression is not a standard TV gradient step. As written, Algorithm 1 cannot execute this step, so the described reconstruction procedure is incomplete.
  3. [Sec. III-C, Eq. (9) (and Sec. II-B, Eq. (3))] The forward diffusion is only an interpolation between x0 and a fixed terminal sample x_T; at t = T the equation reduces to the identity x_T = x_T regardless of α_T, so it does not define how LDCT noise enters the process. Specify the schedule (including the value of α_T) and how x_T is sampled, or state that Eq. (9) is an interpolation rather than a noise-addition process.
  4. [Sec. IV-A and Table II] The AAPM Challenge evaluation is based on a single test patient, and Table II reports no standard deviations, confidence intervals, or statistical significance tests for either dataset; without these, the reported margins (e.g., 0.56 dB over CoreDiff at 1e4 on AAPM) cannot be distinguished from run-to-run or patient-to-patient variability, weakening the generalization claim.
  5. [Sec. III-D, Eq. (25)] The PWLS weight model W = diag{w_t} = diag{1/σ_x_t^2} = diag{(1/l1) e^{η/x_t}} with η = 22000 is numerically implausible for typical sinogram magnitudes, and the quantity l1 is never defined; please provide the exact weight model, including units and any normalization, or the PWLS step is not reproducible.
minor comments (5)
  1. [Sec. II-B, Eq. (5)] The training objective is missing a parenthesis: Rθ(D(x0,xT,t)-x0) should presumably be ∥Rθ(D(x0,xT,t)) - x0∥, and the sentence preceding Eq. (5) should be aligned with that reading.
  2. [Sec. IV-C, text near Table II] The sentence 'FD-DiT achieves the highest PSNR, SSIM, and MSE values' should read 'the lowest MSE values,' since lower MSE is better.
  3. [Sec. IV-D.2, heading] The heading 'Sliding Spare Local Attention' contains a typo; it should be 'Sliding Sparse Local Attention.'
  4. [Fig. 11 caption] The caption labels the third panel as '(b)' twice; the LDF panel should be labeled '(c).'
  5. [Algorithm 1, lines 2-8] The loop index convention is unclear: after the loop 'For t = T to 1,' the variable used in the final reconstruction is written as x0, but the loop body assigns to x_{t-1}; please clarify the indexing so that the final projection variable is explicitly defined.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the central claim is an externally benchmarked empirical result, and the self-citations are not load-bearing.

full rationale

FD-DiT's central claim is an empirical benchmark result (Table II) evaluated against external AAPM Challenge and CIRS phantom datasets, not an analytic derivation whose conclusion is fed back in as an assumption. The training objective Eq. (19) fits network parameters to paired NDCT/LDCT data, and the reported PSNR/SSIM/MSE are measured against held-out ground truth; no fitted constant is renamed as a prediction. The Gaussian frequency decomposition in Eq. (10) is a stated modeling choice rather than a result derived from itself, and the multi-module/LDF architecture is validated by ablations (Tables III-IV). Self-citations such as [25] appear only as background or comparison and carry none of the load-bearing argument; no uniqueness theorem is imported from the authors' prior work. The PWLS/TV reconstruction equations (23)-(26) are not circular, although Eq. (24) is not the actual minimizer of Eq. (23) and would be a correctness, not circularity, concern. Overall, the paper is self-contained against external benchmarks, and no circular step reduces a prediction to its own input.

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

The central claim rests on a set of domain assumptions about CT noise being high-frequency and about the validity of paired-data diffusion interpolation, plus several hand-set hyperparameters. No new physical entities are introduced; the modules FHD, FLD, FFD, and LDF are neural network components, not hypothesized physical objects.

free parameters (5)
  • Gaussian filter bandwidth sigma = not reported
    Controls the high/low frequency split in Eq. (10); no ablation or selection criterion given.
  • PWLS calibration coefficient eta = 22000
    Hand-set in Eq. (25) to weight data fidelity; no sensitivity analysis.
  • Diffusion schedule alpha_t and total steps T = not fully specified; PWLS iterations set to 22000
    The forward process in Eq. (9) requires a schedule; the paper states alpha_t < alpha_{t-1} but not the actual values.
  • Attention window size omega and dilation rates r = r=1,2,3; omega unspecified
    Hyperparameters of SSLA (Eq. 12-13); only the dilation rates are given.
  • Learned network weights theta = trained on AAPM training set
    The denoising output (Eq. 18) is produced by a neural network with millions of trained parameters; the claimed reconstruction capability is empirical, not derived.
assumptions (5)
  • domain assumption Fourier-domain Gaussian filtering cleanly splits sinogram data into high- and low-frequency components (Eq. 10).
    The decomposition is standard signal processing, but its adequacy for CT noise separation is assumed.
  • domain assumption Most LDCT noise is concentrated in the high-frequency band while anatomy is low-frequency (Sec. III-A, Fig. 1).
    The entire multi-branch design depends on this empirical claim, supported only by a qualitative figure.
  • domain assumption Paired NDCT/LDCT sinogram data exist and the interpolation path in Eq. (9) is a valid degradation process for training.
    The generalized diffusion formulation requires paired samples and assumes the convex combination is a meaningful degradation.
  • ad hoc to paper Transformer with local and global attention is better than U-Net for high-frequency denoising, and U-Net is better for low-frequency denoising (Sec. III-A).
    This architectural choice is motivated heuristically and validated only through the final ablation.
  • ad hoc to paper The LDF concatenation of three denoised outputs is an optimal or near-optimal fusion (Eq. 16-17).
    The learned fusion strategy is proposed as an improvement over summation, but no theoretical justification is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FD-DiT: Frequency Domain-Directed Diffusion Transformer for Low-Dose CT Reconstruction." pith.science (2026). https://pith.science/paper/VFXE6O3D

@misc{pith2026250623466,
  author       = {Pith},
  title        = {Pith review of: FD-DiT: Frequency Domain-Directed Diffusion Transformer for Low-Dose CT Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFXE6O3D}},
  note         = {Machine review of arXiv:2506.23466}
}
read the original abstract

Low-dose computed tomography (LDCT) reduces radiation exposure but suffers from image artifacts and loss of detail due to quantum and electronic noise, potentially impacting diagnostic accuracy. Transformer combined with diffusion models has been a promising approach for image generation. Nevertheless, existing methods exhibit limitations in preserving finegrained image details. To address this issue, frequency domain-directed diffusion transformer (FD-DiT) is proposed for LDCT reconstruction. FD-DiT centers on a diffusion strategy that progressively introduces noise until the distribution statistically aligns with that of LDCT data, followed by denoising processing. Furthermore, we employ a frequency decoupling technique to concentrate noise primarily in high-frequency domain, thereby facilitating effective capture of essential anatomical structures and fine details. A hybrid denoising network is then utilized to optimize the overall data reconstruction process. To enhance the capability in recognizing high-frequency noise, we incorporate sliding sparse local attention to leverage the sparsity and locality of shallow-layer information, propagating them via skip connections for improving feature representation. Finally, we propose a learnable dynamic fusion strategy for optimal component integration. Experimental results demonstrate that at identical dose levels, LDCT images reconstructed by FD-DiT exhibit superior noise and artifact suppression compared to state-of-the-art methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 53 canonical work pages

  1. [1]

    Computed tomography —An increasing source of radiation exposure,

    D. J. Brenner and E. J. Hall, “Computed tomography —An increasing source of radiation exposure,” New Engl. J. Med. , vol. 357, no. 22, pp. 2277-2284, 2007

    work page 2007

  2. [2]

    Radiation- related cancer risk associated with surveillance imaging for metastasis from choroidal melanoma,

    J. C. Wen， V. Sai, B. R. Straatsma, and T. A. McCannel , "Radiation- related cancer risk associated with surveillance imaging for metastasis from choroidal melanoma," JAMA ophthalmology, vol. 131, no. 1, pp. 58– 63, 2013

    work page 2013

  3. [3]

    Health risks from exposure to low levels of ionizing radiation,

    National Research Council et al. "Health risks from exposure to low levels of ionizing radiation," National Academies Press, Washington, DC, 2006

    work page 2006

  4. [4]

    Bismuth shielding, organ-based tube current modulation, and global reduction of tube current for dose reduction to the eye at head CT,

    J. Wang et al., "Bismuth shielding, organ-based tube current modulation, and global reduction of tube current for dose reduction to the eye at head CT," Radiology, vol. 35, no. 2, pp. 170–175, 2012

    work page 2012

  5. [5]

    Strategies for CT radiation dose optimization,

    M. K. Kalra et al. , “Strategies for CT radiation dose optimization,” Radiology., vol. 230, no. 3, pp. 1000-1008, 2004

    work page 2004

  6. [6]

    Optimizing radiation dose and image quality,

    C. Catalano, M. Francone, A. Ascarelli, M. Mangia, I. Iacucci, and R. Passariello, "Optimizing radiation dose and image quality," European Radiology Supplements, vol. 17, pp. 26–32, 2007

    work page 2007

  7. [7]

    Principles of CT: radiation dose and image quality,

    L. W. Goldman, "Principles of CT: radiation dose and image quality," Journal of Nuclear Medicine Technology , vol. 35, no. 4, pp. 213 –225, 2007

    work page 2007

  8. [8]

    Sinogram noise reduction for low - dose CT by statistics-based nonlinear filters,

    J. Wang, H. Lu, T. Li, and Z. Liang, "Sinogram noise reduction for low - dose CT by statistics-based nonlinear filters," in Medical Imaging 2005: Image Processing, vol. 5747, SPIE, 2005, pp. 2058–2066

    work page 2005

Show all 54 references
  1. [9]

    Nonlinear sinogram smoothing for low-dose X-ray CT,

    T. Li, X. Li, J. Wang, J. Wen, H. Lu, J. Hsieh, and Z. Liang, "Nonlinear sinogram smoothing for low-dose X-ray CT," IEEE Trans. Nucl. Sci., vol. 51, no. 5, pp. 2505–2513, 2004

    2004

  2. [10]

    Penalized-likelihood sinogram restoration for computed tomography,

    P. J. La Riviè re, J. Bian, and P. A. Vargas, "Penalized-likelihood sinogram restoration for computed tomography," IEEE Trans. Med. Imaging , vol. 25, no. 8, pp. 1022–1036, 2006

    2006

  3. [11]

    Iterative reconstruction methods in X-ray CT,

    M. Beister, D. Kolditz, and W. A. Kalender, "Iterative reconstruction methods in X-ray CT," Physica Medica, vol. 28, no. 2, pp. 94–108, 2012

    2012

  4. [12]

    Model -based iterative reconstruction technique for ultralow-dose computed tomography of the lung: a pilot study,

    Y. Yamada et al., "Model -based iterative reconstruction technique for ultralow-dose computed tomography of the lung: a pilot study," Investigative Radiology, vol. 47, no. 8, pp. 482–489, 2012

    2012

  5. [13]

    Artifact suppressed dictionary learning for low -dose CT image processing,

    Y. Chen et al., "Artifact suppressed dictionary learning for low -dose CT image processing," IEEE Trans. Med. Imaging, vol. 33, no. 12, pp. 2271– 2292, 2014

    2014

  6. [14]

    Learning-based artifact removal via image decomposition for low -dose CT image processing,

    X. Y. Cui, Z. G. Gui, Q. Zhang, H. Shangguan, and A. H. Wang, "Learning-based artifact removal via image decomposition for low -dose CT image processing," IEEE Trans. Nucl. Sci , vol. 63, no. 3, pp. 1860 – 1873, 2016

    2016

  7. [15]

    Thoracic low-dose CT image processing using an artifact suppressed large-scale nonlocal means,

    Y. Chen et al., "Thoracic low-dose CT image processing using an artifact suppressed large-scale nonlocal means," Physics in Medicine & Biology, vol. 57, no. 9, p. 2667, 2012

    2012

  8. [16]

    Low-dose CT with a residual encoder -decoder convolutional neural network,

    H. Chen, Y. Zhang, M. K. Kalra, F. Lin, Y. Chen, P. Liao, J. Zhou, an d G. Wang, “Low-dose CT with a residual encoder -decoder convolutional neural network,” IEEE Trans. Med. Imag ing., vol. 36, pp. 2524 –2535, 2017

    2017

  9. [17]

    Generative image modeling using style and structure adversarial networks,

    X. Wang and A. Gupta , “Generative image modeling using style and structure adversarial networks,” in Proc. Eur. Conf. Comput. Vis. (ECCV), Springer, 2016, pp. 318–335

    2016

  10. [18]

    SemanticGAN: Generative adversarial networks for semantic image to photo-realistic image translation,

    J. Liu, Y. Zou, and D. Yang, “SemanticGAN: Generative adversarial networks for semantic image to photo-realistic image translation,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP) , 2020, pp. 2528–2532

    2020

  11. [19]

    WBT -GAN: Wavelet based generative adversarial network for texture synthesis,

    S. S. T. Moghadam et al. , “WBT -GAN: Wavelet based generative adversarial network for texture synthesis,” in Proc. 2021 11th Int. Conf. Comput. Eng. Knowl. (ICCKE), 2021, pp. 441–446. 11

    2021

  12. [20]

    Low -dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss,

    Q. Yang et al. , “Low -dose CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss,” IEEE Trans. Med. Imaging, vol. 37, pp. 1348–1357, 2018

    2018

  13. [21]

    DU -GAN: Generative adversarial networks with dual -domain U -Net based discriminators for low-dose CT denoising,

    Z. Huang, J. Zhang, Y. Zhang, and H. Shan, “DU -GAN: Generative adversarial networks with dual -domain U -Net based discriminators for low-dose CT denoising,” IEEE Trans. Instrum. Meas., vol. 71, pp. 1–12, 2021

    2021

  14. [22]

    Score-based generative modeling through stochastic differential equations,

    Y. Song, J. Sohl -Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in Proc. Int. Conf. Learn. Represent., 2020

    2020

  15. [23]

    CoreDiff: Contextual error -modulated generalized diffusion model for low -dose CT denoising and generalization,

    Q. Gao et al., “CoreDiff: Contextual error -modulated generalized diffusion model for low -dose CT denoising and generalization,” IEEE Trans. Med. Imaging., vol. 43, no. 2, pp. 411–423, 2024

    2024

  16. [24]

    CoCoDiff: A contextual conditional diffusion model for low -dose CT image denoising,

    Q. Gao and H. Shan, “CoCoDiff: A contextual conditional diffusion model for low -dose CT image denoising,” in Developments in X -Ray Tomography XIV, vol. 12242, SPIE, 2022, pp. 92–98

    2022

  17. [25]

    One-sample diffusion modeling in projection domain for low-dose CT imaging,

    B. Huang, S. Lu, L. Zhang, B. Lin, W. Wu, and Q. Liu, "One-sample diffusion modeling in projection domain for low-dose CT imaging," IEEE Trans. Radiat. Plasma Med. Sci, 2024

    2024

  18. [26]

    Differential transformer,

    T. Ye, L. Dong, Y. Xia, Y. Sun, Y. Zhu, G. Huang, and F. Wei , “Differential transformer,” arXiv:2410.05258, 2024

    2024 arXiv

  19. [27]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y. Lu, S. Pan, W. Bo, and Y. Liu, "Roformer: Enhanced transformer with rotary position embedding," Neurocomputing, vol. 568, p. 127063, 2024

    2024

  20. [28]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 4195–4205

    2023

  21. [29]

    Swin-Unet: Unet-like pure transformer for medical image segmentation,

    H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-Unet: Unet-like pure transformer for medical image segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 205–222

    2022

  22. [30]

    All are worth words: A vit backbone for diffusion models,

    F. Bao, S. Nie, K. Xue, Y. Cao, C. Li, H. Su, and J. Zhu , “All are worth words: A vit backbone for diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vision Pattern Recognit., 2023, pp. 22669–22679

    2023

  23. [31]

    Domain specific convolution and high frequency reconstruction based unsupervised domain adaptation for medical image segmentation,

    S. Hu, Z. Liao, and Y. Xia, “Domain specific convolution and high frequency reconstruction based unsupervised domain adaptation for medical image segmentation,” in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent., 2022, pp. 650–659

    2022

  24. [32]

    Learning frequency-aware dynamic network for efficient super-resolution,

    W. Xie, D. Song, C. Xu, C. Xu, H. Zhang, and Y. Wang, “Learning frequency-aware dynamic network for efficient super-resolution,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 4288–4297

    2021

  25. [33]

    Intriguing findings of frequency selection for image deblurring,

    X. Mao, Y. Liu, F. Liu, Q. Li, W. Shen, and Y. Wang, “Intriguing findings of frequency selection for image deblurring,” in Proc. AAAI Conf. Artif. Intell. (AAAI), vol. 37, no. 2, 2023, pp. 1905–1913

    2023

  26. [34]

    Fast Fourier convolution,

    L. Chi, B. Jiang, and Y. Mu, “Fast Fourier convolution,” in Proc. Adv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 4479–4488

    2020

  27. [35]

    Variance analysis of X-ray CT sinograms in the presence of electronic noise background,

    J. Ma, Z. Liang, Y. Fan, Y. Liu, J. Huang, W. Chen, H. Lu, "Variance analysis of X-ray CT sinograms in the presence of electronic noise background," Med. Phys., vol. 39, no. 7, pp. 4051-4065, 2012

    2012

  28. [36]

    Robust low- dose CT sinogram preprocessing via exploiting noise -generating mechanism,

    Q. Xie, D. Zeng, Q. Zhao, D. Meng, Z. Xu, Z. Liang, J. Ma, "Robust low- dose CT sinogram preprocessing via exploiting noise -generating mechanism," IEEE Trans. Med. Imaging, vol. 36, no. 12, pp. 2487-2498, 2017

    2017

  29. [37]

    Penalized weighted least -squares approach to sinogram noise reduction and image reconstruction for low - dose X-ray computed tomography,

    J. Wang, T. Li, H. Lu, and Z. Liang , “Penalized weighted least -squares approach to sinogram noise reduction and image reconstruction for low - dose X-ray computed tomography,” IEEE Trans. Med. Imaging, vol. 25, no. 10, pp. 1272–1283, 2006

    2006

  30. [38]

    Noise simulation for low -dose computed tomography,

    Y. Fan, A. Zamyatin, S. Nakanishi, "Noise simulation for low -dose computed tomography," in Proc. IEEE Nucl. Sci. Symp. Med. Imag. Conf. (NSS/MIC), 2012, pp. 3641–3643

    2012

  31. [39]

    Few-view image reconstruction with fractional-order total variation,

    Y. Zhang, W. Zhang, Y. Lei, and Zhou J, “Few-view image reconstruction with fractional-order total variation,” J. Opt. Soc. Am., vol. 31, no. 5, pp. 981-995, 2014

    2014

  32. [40]

    Denoising of low -dose CT images via low -rank tensor modeling and total variation regularization,

    S. V. M. Sagheer, and S. N. George, “Denoising of low -dose CT images via low -rank tensor modeling and total variation regularization,” Artif. Intel. in Med., vol. 94, pp. 1-17, 2019

    2019

  33. [41]

    GPU-based fast cone beam CT reconstruction from undersampled and noisy projection data via total variation,

    X. Jia, Y. Lou, R. Li, W. Y. Song and S. B. Jiang, “GPU-based fast cone beam CT reconstruction from undersampled and noisy projection data via total variation,” Med. Phys., vol. 37, no. 4, pp. 1757-1760, 2010

    2010

  34. [42]

    Bert: Pre-training of deep bidirectional transformers for language understanding ,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "Bert: Pre-training of deep bidirectional transformers for language understanding ," in Proc. Conf. North Amer. Chapter Assoc. Comput . Linguistics (NAACL), 2019, pp. 4171-4186

    2019

  35. [43]

    Dilateformer: Multi-scale dilated transformer for visual recognition,

    J. Jiao, Y. M. Tang, K. Y. Lin, Y. Ding, W. Wu, and J. Yan, “Dilateformer: Multi-scale dilated transformer for visual recognition,” IEEE Trans. Multimedia., vol. 25, pp. 8906–8919, 2023

    2023

  36. [44]

    An image is worth 16×16 words: transformers for image recognition at scale,

    A. Dosovitskiy, et al, “An image is worth 16×16 words: transformers for image recognition at scale,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2020

    2020

  37. [45]

    Training data -efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jé gou, “Training data -efficient image transformers & distillation through attention,” in Proc. Int. Conf. Mach. Learn. (ICML), PMLR, 2021, pp. 10 347–10 357

    2021

  38. [46]

    Twins: Revisiting the design of spatial attention in vision transformers,

    X. Chu, Z. Tian, Y. Wang, B. Zhang, H. Ren, X. Wei, H. Xia, and C. Shen, “Twins: Revisiting the design of spatial attention in vision transformers,” in Proc. Annu. Conf. Neural Inf. Process. Syst. (NeurIPS), 2021, vol. 34, pp. 9355–93661

    2021

  39. [47]

    Neighborhood attention transformer,

    A. Hassani, S. Walton, J. Li, S. Li, and H. Shi, “Neighborhood attention transformer,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit (CVPR), 2023, pp. 6185–6194

    2023

  40. [48]

    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 Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 10 012–10 022

    2021

  41. [49]

    MaxVit: Multi-axis vision transformer,

    Z. Tu, H. Talebi, H. Zhang, F. Yang, B. Behnam, Y. Li, J. Luo, and R. Ba, “MaxVit: Multi-axis vision transformer,” in Proc. Eur. Conf. Comput.Vis. (ECCV), Springer, 2022, pp. 459–479

    2022

  42. [50]

    Crossformer: A versatile vision transformer based on cross -scale attention,

    W. Wang, L. Yao, L. Chen, X. Shao, J. Shi, Z. Gao, and A. Zhang, “Crossformer: A versatile vision transformer based on cross -scale attention,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 5, pp. 3123-3136, 2023

    2023

  43. [51]

    Dynamic convolution: Attention over convolution kernels,

    Y. Chen, X. Dai, M. Liu, D. Chen, L. Yuan, and Z. Liu, “Dynamic convolution: Attention over convolution kernels,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 11030–11039

    2020

  44. [52]

    Fast calculation of the exact radiological path fora three- dimensional CT array,

    R. L. Siddon, “Fast calculation of the exact radiological path fora three- dimensional CT array,” Med. Phys., vol. 12, no. 2, pp. 252–255, 1985

    1985

  45. [53]

    Low-rank angular prior guided multi-diffusion model for few-shot low-dose CT reconstruction,

    W. Zhang, B. Huang, S. Chen, X. Li, and Y. Wu, “Low-rank angular prior guided multi-diffusion model for few-shot low-dose CT reconstruction,” IEEE Trans. Comput. Imaging, 2024

    2024

  46. [54]

    WiTUnet: A U-shaped architecture integrating CNN and transformer for improved feature alignment and local information fusion,

    B. Wang, F. Deng, P. Jiang, X. Zhang, and L. Liu, “WiTUnet: A U-shaped architecture integrating CNN and transformer for improved feature alignment and local information fusion,” Sci. Rep., vol. 14, no. 1, p. 25525, 2024

    2024

Pith tools

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