Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Double-Constraint Diffusion Model with Nuclear Regularization for Ultra-low-dose PET Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Steering a frozen pre-trained diffusion model with two small modules reconstructs ultra-low-dose PET at down to 1% of full dose and adapts to dose levels never seen in training.

desk verdict A useful parameter-efficient PET reconstruction idea with a table-level inconsistency that must be resolved before the headline claims are credible. read the letter →

arxiv 2509.00395 v1 pith:JF5RBRWW submitted 2025-08-30 cs.CV

classification cs.CV
keywords ultra-low-dosePETreconstructiondiffusionmodelnuclearnormregularizationlow-rankfeatureextractionTransformerdosereductionfactorunknown-dosegeneralizationmedicalimage
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 claims that you do not need to fine-tune a large diffusion model to adapt PET image reconstruction to a new radiation-dose level. In DCDM, the pre-trained diffusion network is frozen and two small trainable modules are attached: the Nuclear Transformer Constraint (NTC) compresses the low-dose image into low-rank, sparse features via nuclear-norm regularization, and the Encoding Nexus Constraint (ENC) injects those features into the frozen decoder as control signals at every denoising step. The authors report that this beats U-Net, Transformer, GAN, and diffusion baselines at all five dose levels of the public UDPET dataset and improves clinical lesion metrics on a ten-patient dataset with unknown dose levels, including reconstructions from 1% of the full dose. If true, the practical payoff is that image priors and dose adaptation are separable: one frozen generative backbone serves many dose regimes by training only the small controller, and the same features that guide reconstruction can route an image of unknown dose to the right controller.

What carries the argument

The load-bearing object is the double-constraint controller plugged into a frozen diffusion model. NTC (Nuclear Transformer Constraint) is a Transformer whose training objective combines the nuclear norm of the feature matrix—a convex proxy for rank—with a sparsity term and a lossy-coding-rate term; its attention layers are derived as unrolled gradient steps on that objective, and a classification head lets the same features label the image's dose level. ENC (Encoding Nexus Constraint) is six encoding-nexus blocks that fuse the NTC features with the diffusion timestep into a modulation feature and inject it into the frozen decoder through zero-initialized convolutions (ZeroConv), which keep

What would settle it

Give NTC held-out low-dose images with dose levels it never trained on (e.g., DRF 30, 15, or 200), report a confusion matrix and per-class accuracy, and compare reconstruction quality when the image is routed through the selected ENC versus through a deliberately wrong one. If misrouting does not measurably hurt PSNR/SSIM, then the adaptive-routing mechanism is not what carries the generalization result.

Watch

Extended reading notes

Core claim

Frozen diffusion weights plus a two-module controller suffice, the paper claims, for ultra-low-dose PET reconstruction. NTC, a Transformer with nuclear-norm regularization—a convex stand-in for matrix rank—squeezes the low-dose image into low-rank, sparse features that suppress noise but keep structure. ENC fuses those features with the diffusion timestep and injects them into the frozen decoder through zero-initialized convolutions, so its influence starts at zero. One ENC per dose level serves DRFs 100, 50, 20, 10, 4; for unknown dose levels, NTC's classifier selects the matching ENC. On the UDPET dataset and a ten-patient clinical set from another scanner, it reports beating full-training

Load-bearing premise

That the NTC classifier, trained on five discrete dose levels, routes a real low-dose image whose dose level was not among them to the correct one of five controllers; the paper shows t-SNE cluster separation but reports no classification accuracy, and the only clinical support is ten patients.

Editorial extensions

If this is right

  • A single frozen diffusion backbone can serve multiple dose levels: adapting to a new DRF trains only the small ENC controller, not the full model.
  • Ultra-low-dose reconstruction at 1% of the full scan (DRF 100) is achievable with this controller, with the paper reporting the best PSNR/SSIM/FID/LPIPS among the compared methods.
  • The same low-rank features that steer reconstruction can classify an image's dose level, which is what enables the unknown-DRF pipeline.
  • Because the backbone is untouched, the image prior and the dose adaptor evolve independently: a better frozen prior or a better controller improves reconstruction without redoing the other.

Reading between the lines

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

  • A natural testable extension: instead of a hard one-of-five ENC choice, blend the control signals of neighboring ENCs weighted by NTC class probabilities; for dose levels between trained DRFs (say 30 or 15) this could reconstruct more smoothly than selecting the nearest class.
  • The mechanism is not obviously PET-specific: if nuclear-norm compression is what protects count-starved images, the same controller design could transfer to other low-count modalities such as low-dose CT or SPECT, where the noise model is similar.
  • The paper's evidence for correct routing in real scans is t-SNE separation plus a ten-patient clinical set; measuring NTC's classification accuracy and confusion matrix across a continuous range of DRFs would reveal the dose levels where the adaptive selection starts to misroute, since a wrong ENC means the image is steered by the wrong dose-specific control.
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

4 major / 5 minor

Summary. The paper proposes DCDM, a double-constraint diffusion framework for ultra-low-dose PET reconstruction. A pre-trained diffusion model is frozen, and two lightweight trainable modules are injected: the Nuclear Transformer Constraint (NTC), which extracts low-rank and sparse compressed feature representations from low-dose images using a nuclear-norm regularized Transformer, and the Encoding Nexus Constraint (ENC), which injects these representations into the decoder of the pre-trained diffusion model. The authors claim state-of-the-art performance on the UDPET public dataset across known dose reduction factors (DRF 4, 10, 20, 50, 100), and generalization to unknown DRF scenarios by using NTC as a classifier to select one of five dose-specific ENC modules. Experiments include comparisons with U-Net, MPRNet, ViT, Pix2Pix, IDDPM, and ControlNet, ablation studies, and a clinical evaluation on 10 patients with clinical metrics such as ΔSUV, SNR, CoV, and CR.

Significance. If the claims hold, DCDM would be a practically valuable contribution: it adapts a single frozen diffusion model to multiple dose levels by training only a small controller, which is parameter-efficient and flexible for clinical scenarios where the dose reduction factor is unknown. The open-source code, public-dataset evaluation, and the idea of combining low-rank Transformer features with a controllable diffusion decoder are strengths. However, the validity of the central claim depends on the consistency and statistical reliability of the reported numbers, and on evidence that the NTC classifier actually selects the correct ENC in unseen conditions. The table-level inconsistency identified below is load-bearing and must be resolved.

major comments (4)
  1. [§IV.B (Table I) vs §IV.C (Table III)] The same condition (DRF=100, UDPET, full DCDM) is reported with different values: Table I gives PSNR=40.12, SSIM=0.9725, FID=21.40, LPIPS=0.0356; Table III gives PSNR=38.24, SSIM=0.9607, FID=33.11, LPIPS=0.0344. The PSNR gap is 1.88 dB and the FID gap is 11.71, far beyond rounding. If Table III is correct, DCDM trails ControlNet (39.73 dB) and IDDPM (39.57 dB) at DRF=100, directly contradicting the abstract's state-of-the-art claim on known DRFs. The paper must specify which evaluation is used, why the two tables differ (test split, slice count, sampling schedule, checkpoint), and correct the tables or the claims accordingly.
  2. [Abstract and §IV.B, Table I] The abstract claims DCDM 'outperforms state-of-the-art methods on known dose reduction factors' as a blanket statement over all metrics. Table I shows this is not true for all metrics: at DRF=20, DCDM has FID=22.02 while ControlNet has FID=20.56, and LPIPS values tie at 0.0300; at DRF=4, LPIPS also ties at 0.0200. The claim should be qualified to specific metrics (e.g., PSNR/SSIM), or the paper should report which metrics DCDM does and does not improve.
  3. [§III.C and §IV.B (Clinical dataset)] The unknown-DRF mechanism relies on NTC's classification output to select one of five ENC modules (DRF 100, 50, 20, 10, 4), but no classification accuracy or confusion matrix is reported anywhere. Fig. 7 only shows t-SNE separation, which is not a quantitative measure. The clinical validation uses only 10 patients, and the unknown DRF data are obtained by a different list-mode resampling procedure than the UDPET training data. Without evidence that misclassification does not route an image to the wrong ENC, the paper's central generalization claim is unsupported. Please report per-class accuracy, a confusion matrix, and ideally a sensitivity analysis of ENC selection.
  4. [§IV.A and Tables I–II] All reported metrics are point estimates without error bars, confidence intervals, or significance tests on the 1,346-slice test set and the 10-patient clinical set. Several advantages over ControlNet are small (e.g., DRF=100 PSNR 40.12 vs 39.73, a 0.39 dB difference), and diffusion sampling is stochastic. Given the table inconsistency, it is critical to report variance across slices/patients and perform paired statistical tests (e.g., Wilcoxon signed-rank) for the headline metrics.
minor comments (5)
  1. [Equations (4), (5), Algorithm 1] The notation in Eqs. (4) and (5) is garbled (β, α, and Σ symbols are mixed), and Algorithm 1 line 8 has unbalanced parentheses. Also, the text introduces 'MSSA' in Eq. (11) but later calls it 'MHSA'; please unify.
  2. [Eq. (13) and Eq. (15)] The derivation assumes D^T D ≈ I in Eq. (13) and that learned FFN operators H, Γ, Φ converge to ADMM subproblem solutions in Eq. (15). These are strong assumptions; please justify them or provide empirical validation that the unrolled network matches the optimization objective.
  3. [References] Reference [46] is cited for LPIPS but points to Diff-Restorer; the original LPIPS paper (Zhang et al., 2018) should be cited. Several other references are formatted inconsistently.
  4. [Fig. 2] The claim that NTC shows 'superiority' in rank and sparsity is based on visual 3D plots without quantitative rank or sparsity values. Please report numerical summaries to support this claim.
  5. [Table I] The bold/italic convention for optimal and suboptimal values is difficult to discern in the text version; please ensure the best and second-best values are clearly distinguishable in the final PDF.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the DCDM derivation and evaluation are self-contained.

full rationale

The paper's central derivation chain is not circular. The pre-trained diffusion model is trained internally on full-dose PET images (Section III.B), not imported from the authors' prior work, and the double-constraint controller (NTC/ENC) is trained on paired low-dose/full-dose slices and evaluated on held-out slices plus a separate clinical dataset against an external full-dose reference. The NTC module's construction from rate-distortion and nuclear-norm objectives (Eqs. 7-15) is an unrolled-optimization architecture derivation based on external work (ref. [42], White-box transformers via sparse rate reduction), and the paper does not claim the trained network's outputs are equivalent to the optimization objective by construction; rather, it uses the unrolled steps as a network design. The unknown-DRF selection mechanism uses the NTC classifier to choose among five ENC controllers trained at discrete DRFs; this is an empirical generalization claim, not a fitted parameter renamed as a prediction. The self-citation [25] appears only in a contextual list of diffusion-based PET methods and is not load-bearing for any theorem or fitted constant. The notable discrepancy between Table I and Table III for the same DRF=100 DCDM condition (PSNR 40.12 vs 38.24, FID 21.40 vs 33.11) is a serious internal-consistency/correctness issue that should be investigated, but it is not a circularity: the evaluation still compares the model's output against ground-truth full-dose images rather than reducing to the training objective. Overall, the derivation is self-contained against external benchmarks, and no circular step can be exhibited from the paper's own equations.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The core empirical claim rests on several unstated hyperparameters and domain assumptions. The paper's own contribution is the architecture and controller training; the diffusion backbone, rate-reduction background, and DRF-validity assumptions are imported without independent verification. The most fragile inputs are the learned-FFN approximation to the nuclear-norm problem and the unmeasured classifier accuracy for unknown-DRF routing.

free parameters (5)
  • NTC loss weights lambda_1, lambda_2 = not reported
    Weights of the nuclear norm and sparsity terms in Eq. (7); the paper gives no values and says hyper-parameter selection is 'tricky'.
  • ADMM penalty rho and inner learning rate eta = not reported
    Appear in Eq. (14) as 1/(2rho) and eta, and are needed for the unrolled optimization; no values are given.
  • NTC architecture hyperparameters (patch size p, feature dim d, number of NT blocks L) = not reported
    Define the tokenization and depth of the nuclear transformer; not specified in the text.
  • Number of ENC controller modules = 5
    One controller per DRF (100, 50, 20, 10, 4); a design choice that fixes the granularity of unknown-DRF routing.
  • Training iterations and learning rate = 300k/100k/100k, 1e-4
    Chosen training budget; not derived and likely tuned on the validation split.
assumptions (7)
  • standard math DDPM forward and reverse Markov chain equations (Eqs. 1-5) from Ho et al. [28]
    The pre-trained diffusion model and the sampling loop in Algorithm 1 rely on this background result without re-deriving it.
  • standard math Sparse rate reduction objective and its equivalence to multi-head self-attention (Eqs. 7-11) from Yu et al. [42]
    The NTC architecture is presented as an unrolled solution of this objective; the paper takes the equivalence as given.
  • ad hoc to paper Near-orthonormal dictionary assumption D^T D approx I in Eq. (13)
    Needed to replace the gradient of the rate R(Z) with a simpler update; no construction or validation of D is given.
  • ad hoc to paper Learned FFN operators H, Gamma, Phi converge to the ADMM subproblem solutions in Eq. (15)
    The paper replaces the optimization in Eq. (14) by deep networks without convergence or approximation guarantees; this is the bridge used to claim NTC is a nuclear-regularized transformer.
  • domain assumption Paired UDPET slices created by subsampling full scans are a valid proxy for true clinical ultra-low-dose acquisitions
    The known-DRF experiments are trained and evaluated on this synthetic degradation; the clinical set is the only real-dose test and has only 10 patients.
  • domain assumption Low-rank and sparse feature representations of low-dose PET images retain the information needed to guide reconstruction
    The design of NTC rests on this assumption; it is supported only by downstream metrics and t-SNE clusters, not by an independent measurement.
  • domain assumption NTC class assignments select the correct ENC for unknown DRF images
    The adaptive pipeline in Section III.C relies on classifier accuracy, which is never reported; without it the unknown-DRF claim is unsupported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Double-Constraint Diffusion Model with Nuclear Regularization for Ultra-low-dose PET Reconstruction." pith.science (2026). https://pith.science/paper/JF5RBRWW

@misc{pith2026250900395,
  author       = {Pith},
  title        = {Pith review of: Double-Constraint Diffusion Model with Nuclear Regularization for Ultra-low-dose PET Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JF5RBRWW}},
  note         = {Machine review of arXiv:2509.00395}
}
read the original abstract

Ultra-low-dose positron emission tomography (PET) reconstruction holds significant potential for reducing patient radiation exposure and shortening examination times. However, it may also lead to increased noise and reduced imaging detail, which could decrease the image quality. In this study, we present a Double-Constraint Diffusion Model (DCDM), which freezes the weights of a pre-trained diffusion model and injects a trainable double-constraint controller into the encoding architecture, greatly reducing the number of trainable parameters for ultra-low-dose PET reconstruction. Unlike full fine-tuning models, DCDM can adapt to different dose levels without retraining all model parameters, thereby improving reconstruction flexibility. Specifically, the two constraint modules, named the Nuclear Transformer Constraint (NTC) and the Encoding Nexus Constraint (ENC), serve to refine the pre-trained diffusion model. The NTC leverages the nuclear norm as an approximation for matrix rank minimization, integrates the low-rank property into the Transformer architecture, and enables efficient information extraction from low-dose images and conversion into compressed feature representations in the latent space. Subsequently, the ENC utilizes these compressed feature representations to encode and control the pre-trained diffusion model, ultimately obtaining reconstructed PET images in the pixel space. In clinical reconstruction, the compressed feature representations from NTC help select the most suitable ENC for efficient unknown low-dose PET reconstruction. Experiments conducted on the UDPET public dataset and the Clinical dataset demonstrated that DCDM outperforms state-of-the-art methods on known dose reduction factors (DRF) and generalizes well to unknown DRF scenarios, proving valuable even at ultra-low dose levels, such as 1% of the full dose.

Figures

Figures reproduced from arXiv: 2509.00395 by the authors.

Figure 1
Figure 1. , which respectively leverage low-dose PET images as constraints to maintain semantic consistency and transform them into feature representations to ensure fine-grained detail accuracy. Notably, NTC enforces the low-rank and sparse prop￾erties of these feature representations. These properties enable the feature representations to suppress substantial interfering in￾formation inherent in low-dose PET images while re… view at source ↗
Figure 2
Figure 2. displays a comparative 3D visualization evaluating the rank and sparsity of NTC and other recognition models in their compressed feature embeddings, synchronized with the feature extraction progression. The sequence of three sparsity-focused plots illustrates how zero-count distributions evolve across samples for the three models during feature extraction progres￾sion, unveiling changing sparsity patterns. The right… view at source ↗
Figure 3
Figure 3. The detailed structure of the encoding nexus (EN) block in the ENC module. Given the time step t from the pre-trained diffusion model (details provided in Section Ⅲ. B) and the compressed feature representation Z from NTC, these two inputs are separately processed by a FFN and a FFN equipped with a zero linear layer. The zero linear, in which both the weight matrix and bias are initialized to zero, is introduced to … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparative illustration of a traditional deep learning-based model and the DCDM framework for unknown DRF reconstruction. IV. EXPERIMENTS A. Experimental Setup In this section, the performance of DCDM is compared with state-of-the-arts at varying dose levels, includin…
Figure 5
Figure 5. Figure 5: presents a visual comparison of the reconstruction re￾sults on the UDPET public dataset across different known dose levels. The proposed DCDM shows significant advantages over other state-of-the-art methods. In terms of overall image quality, the PET images reconstruct…
Figure 6
Figure 6. Figure 6: Reconstruction results on the Clinical dataset at an unknown DRF. The three rows present Full-dose, unknown low-dose, and reconstructions by U-Net, MPRNet, ViT-Rec, Pix2Pix, IDDPM, ControlNet, and DCDM (Ours). Table Ⅱ assesses the performance of state-of-the-art method…
Figure 7
Figure 7. Figure 7: t-SNE visualizations of compressed feature representation process across different methods such as ViT-Cls, ResNet and NTC (Ours). VI. CONCLUSIONS In this study, we presented a novel framework incorporating a nuclear Transformer for ultra-low-dose PET reconstruction. T…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

  1. [1]

    Automated brain tumor detection and segmentation for treatment response assessment using amino acid PET,

    R. Gutsche, C. Lowis, K. Ziemons, M. Kocher, G. Ceccon, C. R. Brambilla et al., “Automated brain tumor detection and segmentation for treatment response assessment using amino acid PET,” Journal of Nuclear Medicine, vol. 64, no. 10, pp 9, 2023

  2. [2]

    PET imaging of neuroinflammation in neurological disorders,

    W. C. Kreisl, M. J. Kim, J. M. Coughlin, I. D. Henter, and R. B . Innis, “PET imaging of neuroinflammation in neurological disorders,” Lancet Neurology, vol. 19, no. 11, pp. 940–950, 2020

  3. [3]

    The basic principles of FDG-PET/CT imaging,

    S. Basu, S. Hess, P.E. N. Braad, B. B. Olsen, S. Inglev, and P. F. Høilund- Carlsen, “The basic principles of FDG-PET/CT imaging,” PET Clinics, vol. 9, no. 4, pp. 355–370, 2014

  4. [4]

    Petformer netw ork enables ultra-low-dose tota l-body PET imaging without structural prior,

    Y. Li and Y. Li, “Petformer netw ork enables ultra-low-dose tota l-body PET imaging without structural prior,” Physics in Medicine & Biology , vol. 69, no. 7, pp 075030, 2024

  5. [5]

    Fast anisotropic gauss filtering,

    J.M. Geusebroek, A. W. Smeulders , and J. Van De Weijer, “Fast anisotropic gauss filtering,” IEEE Trans. Image Processing , vol. 12, no. 8, pp. 938–943, 2003

  6. [6]

    Low dose PET reconstructio n with total variation regularization,

    C. Wang, Z. Hu, P. Shi, and H. Liu, “Low dose PET reconstructio n with total variation regularization,” in 2014 36th Annual International Conf. IEEE Engr. Medicine and Biology Society, pp. 1917–1920, 2014

  7. [7]

    Low dose PET image reconstruction 10 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. xx, NO. x, 2025 with total variation using alternating direction method,

    X. Yu, C. Wang, H. Hu, and H. Liu, “Low dose PET image reconstruction 10 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. xx, NO. x, 2025 with total variation using alternating direction method,” Plos One, vol. 11, no. 12, pp e0166871, 2016

  8. [8]

    Non-lo cal means denoising of dynamic PET images,

    D. Joyita, R. M. Leahy, L. Quanzheng, and M. B. Arrate, “Non-lo cal means denoising of dynamic PET images,” Plos One, vol. 8, no. 12, pp e81390, 2013

Show all 46 references
  1. [9]

    Spatially guided nonlocal mean approach f o r denoising of PET images,

    H. Arabi and H. Zaidi, “Spatially guided nonlocal mean approach f o r denoising of PET images,” Medical Physics , vol. 47, no. 4, pp. 1656– 1669, 2020

  2. [10]

    Image denoi sing with block-matching and 3d filtering,

    K. Dabov, A. Foi, V. Katkovnik, a nd K. Egiazarian, “Image denoi sing with block-matching and 3d filtering,” in Image Processing: Algorithms and Systems, Neural Networks, and Machine Learning , vol. 6064, pp. 354–365, 2006

  3. [11]

    Anatomically guided PET image reconstruction using cond itional weakly-supervised multi- task learning integrati ng self-attention,

    B. Yang, K. Gong, H. Liu, Q. Li, and W. Zhu, “Anatomically guided PET image reconstruction using cond itional weakly-supervised multi- task learning integrati ng self-attention,” IEEE Trans. Medical Imaging , vol. 43, no. 6, pp. 2098–2112, 2024

  4. [12]

    Deep generalized learning model fo r PET image reconstruction,

    Q. Zhang, Y. Hu, Y. Zhao, J. Cheng, W. Fan, D. Hu et al ., “Deep generalized learning model fo r PET image reconstruction,” IEEE Trans. Medical Imaging, vol. 43, no. 1, pp. 122–134, 2023

  5. [13]

    200x low-dose PET reconstruction using deep learning,

    J. Xu, E. Gong, J. Pauly, and G. Zaharchuk, “200x low-dose PET reconstruction using deep learning,” arXiv:1712.04119, 2017

  6. [14]

    Multi-stage progressive image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.H. Yang et al., “Multi-stage progressive image restoration,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) , pp. 14821–14831, 2021

  7. [15]

    Preliminary deep learning-based low dose whole body PET denoising incorpora ting CT information,

    Z. Peng, F. Zhang, J. Sun, Y. Du, Y. Wang, and G. S. Mok, “Preliminary deep learning-based low dose whole body PET denoising incorpora ting CT information,” in 2022 IEEE Conf. Nuclear Science Symposium and Medical Imaging (NSS/MIC), pp. 1–2, 2022

  8. [16]

    A total-body ultralow-dose PET r econstruction method via image space shuffle u-net and body sampling,

    G. Chen, S. Liu, W. Ding, L. Lv, C. Zhao, F. Weng et al., “A total-body ultralow-dose PET r econstruction method via image space shuffle u-net and body sampling,” IEEE Trans. Radiation and Plasma Medical Sciences, vol. 8, no. 4, pp. 357–365, 2023

  9. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zha i, T. Unterthiner et al ., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv:2010.11929, 2020

  10. [18]

    Iterative PET image reconstruction using convolutional neural network representation,

    K. Gong, J. Guan, K. Kim, X. Zhang, J. Yang, Y. Seo et al., “Iterative PET image reconstruction using convolutional neural network representation,” IEEE Trans. Medical Imaging , vol. 38, no. 3, pp. 675– 685, 2018

  11. [19]

    Ultra-low-dose PET reconstruction using generativ e adversarial network with fe ature matching and task-specific perceptual loss,

    O. Jiahong, K. T. Chen, G. Enhao, P. John, and Z. Greg, “Ultra-low-dose PET reconstruction using generativ e adversarial network with fe ature matching and task-specific perceptual loss,” Medical Physics, no. 8, pp 46, 2019

  12. [20]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farl ey, S. Ozair et al., “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

  13. [21]

    Wasserstein generativ e ad- versarial networks,

    M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generativ e ad- versarial networks,” in International Conf. on Machine Learning , pp. 214–223, 2017

  14. [22]

    Improved training of wasserstein gans,

    I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Cou rville, “Improved training of wasserstein gans,” Advances in Neural Information Processing Systems, vol. 30, 2017

  15. [23]

    3d multi- modality transformer-gan for high-quality PET reconstruction,

    Y. Wang, Y. Luo, C. Zu, B. Zhan, Z. Jiao, X. Wu et al ., “3d multi- modality transformer-gan for high-quality PET reconstruction,” Medical Image Analysis, vol. 91, pp 102983, 2024

  16. [24]

    Prior knowledge-guided triple-domain transformer-gan for direct PET reconstruction from low-count sinograms,

    J. Cui, P. Zeng, X. Zeng, Y. Xu, P. Wang, J. Zhou et al ., “Prior knowledge-guided triple-domain transformer-gan for direct PET reconstruction from low-count sinograms,” IEEE Trans. Medical Imaging, no. 12, pp 43, 2024

  17. [25]

    Diffusion transf ormer model with compact prior for low-dose PET reconstruction,

    B. Huang, X. Liu, L. Fang, Q. Liu, and B. Li, “Diffusion transf ormer model with compact prior for low-dose PET reconstruction,” Physics in Medicine and Biology, 2024

  18. [26]

    Bidirectiona l condition diffusion probabilistic models for PET image denoising,

    C . S h e n , C . T i e , Z . Y a n g , N . Z hang, and Y. Zhang, “Bidirectiona l condition diffusion probabilistic models for PET image denoising,” IEEE Trans. Radiation and Plasma Medical Sciences , vol. 8, no. 4, pp. 402– 415, 2024

  19. [27]

    PET imag e denoising based on denoising di ffusion probabilistic model,

    K. Gong, K. Johnson, G. El Fakhri, Q. Li, and T. Pan, “PET imag e denoising based on denoising di ffusion probabilistic model,” European Journal of Nuclear Medicine and Molecular Imaging , vol. 51, no. 2, pp. 358–368, 2024

  20. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems , vol. 33, pp. 6840– 6851, 2020

  21. [29]

    Denoising diffusion implicit m odels,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit m odels,” arXiv:2010.02502, 2020

  22. [30]

    Improved denoising diffusion probabilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” in International Conf. on Machine Learning , pp. 8162–8171, 2021

  23. [31]

    PET-diffusion: Unsupervised PET enhancement based on the latent diffusion model,

    C. Jiang, Y. Pan, M. Liu, L. Ma, X. Zhang, J. Liu et al., “PET-diffusion: Unsupervised PET enhancement based on the latent diffusion model,” in International Conf. on Medical Image Computing and Computer- Assisted Intervention (MICCAI), pp. 3–12, 2023

  24. [32]

    Contrastive diffusion model with auxiliary guidance for coarse-to-fine PET reconstruction,

    Z. Han, Y. Wang, L. Zhou, P. Wang, B. Yan, J. Zhou et al., “Contrastive diffusion model with auxiliary guidance for coarse-to-fine PET reconstruction,” in International Conf. on Medical Image Computing and Computer Assisted Intervention (MICCAI), pp. 239–249, 2023

  25. [33]

    Brain PET synthesis from MRI using joint probability d istribution of diffusion mode l at ultrahigh fields,

    T. Xie, C. Cao, Z. Cui, F. Li, Z. Wei, Y. Zhu et al., “Brain PET synthesis from MRI using joint probability d istribution of diffusion mode l at ultrahigh fields,” arXiv:2211.08901, 2022

  26. [34]

    Synthesizing PET images from high-field and ultra-high-field MR images using joint diffusion attention model,

    T. Xie, C. Cao, Z.X. Cui, Y. Guo, C. Wu, X. Wang et al., “Synthesizing PET images from high-field and ultra-high-field MR images using joint diffusion attention model,” Medical Physics , vol. 51, no. 8, pp. 5250– 5269, 2024

  27. [35]

    Joint diffusion: mutual consistency-dr iven diffusion model for PET-MR I co- reconstruction,

    T . X i e , Z . X . C u i , C . L u o , H . W a n g , C . L i u , Y . Z h a n g et al ., “Joint diffusion: mutual consistency-dr iven diffusion model for PET-MR I co- reconstruction,” Physics in Medicine & Biology , vol. 69, no. 15, pp.155019, 2024

  28. [36]

    Full- dose whole-body PET synthesis from low-dose PET using high-efficiency denoising diffusion probabilistic model: PET consistency model,

    S. Pan, E. Abouei, J. Peng, J. Qian, J. F. Wynne, T. Wang et al., “Full- dose whole-body PET synthesis from low-dose PET using high-efficiency denoising diffusion probabilistic model: PET consistency model, ” Medical Physics, vol. 51, no. 8, pp. 5468–5478, 2024

  29. [37]

    Dose-aware diffusion model for 3d low-dose PET: multi-instituti onal validation with reader study and real low-dose data,

    H. Xie, W. Gan, B. Zhou, M.K. Chen, M. Kulon, A. Boustani et al ., “Dose-aware diffusion model for 3d low-dose PET: multi-instituti onal validation with reader study and real low-dose data,” arXiv:2405.12996, 2024

  30. [38]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. No rouzi, “Image super-resolution via iterative refinement,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4713–4726, 2022

  31. [39]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” Advances in Neural Information Processing Systems, vol. 34, pp. 8780–8794, 2021

  32. [40]

    Ma sked autoencoders are scalable vision learners,

    K . H e , X . C h e n , S . X i e , Y . L i , P. Doll´ar, and R. Girshick, “Ma sked autoencoders are scalable vision learners,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR) , pp. 16000–16009, 2022

  33. [41]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European Conference on Computer Vision, pp. 213–229, 2020

  34. [42]

    White-box transformers via sparse rate reduction,

    Y. Yu, S. Buchanan, D. Pai, T. Chu, Z. Wu, S. Tong et al., “White-box transformers via sparse rate reduction,” Advances in Neural Information Processing Systems, vol. 36, pp. 9422–9457, 2023

  35. [43]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer Assisted Intervention (MICCAI), pp. 234–241, 2015

  36. [44]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J.Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), pp. 1125–1134, 2017

  37. [45]

    Adding conditional control to text- to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text- to-image diffusion models,” in Proc. IEEE/CVF International Conference on Computer Vision (ICCV), pp. 3836–3847, 2023

  38. [46]

    Diff- restorer: Unleashing visual prompts for diffusion-based univers al image restoration,

    Y. Zhang, H. Zhang, X. Chai, Z. Cheng, R. Xie, L. Song et al., “Diff- restorer: Unleashing visual prompts for diffusion-based univers al image restoration,” arXiv:2407.03636, 2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.