Pith. sign in

REVIEW 6 major objections 6 minor 42 references

Pyramid Hierarchical Masked Diffusion Model for Imaging Synthesis

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that PHMDiff, a pyramid hierarchical masked diffusion model, synthesizes missing medical imaging modalities with higher PSNR and SSIM than GAN and diffusion baselines, and does so with fewer training steps.

desk verdict A plausible new architecture for medical image synthesis, but the ablation table reproduces a baseline row verbatim, so the quantitative claims need major repair before they can be believed. read the letter →

arxiv 2507.16579 v1 pith:GVWA7WMZ submitted 2025-07-22 eess.IV cs.AIcs.CV

classification eess.IVcs.AIcs.CV
keywords medicalimagesynthesisdiffusionmodelsmaskedautoencoderpyramidhierarchymulti-scalemaskingcross-granularityregularizationMRICT
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

PHMDiff is proposed as a way to synthesize missing medical imaging modalities, such as producing a T2-weighted MRI from a T1-weighted MRI or CT from MRI, so that patients need fewer scans. The central claim is that a pyramid of progressively down-scaled images, each passed through a masked transformer-based diffusion model, preserves overall anatomy and fine detail better than GANs, plain diffusion models, and previous masked diffusion hybrids. On the BraTS dataset the method reports the best PSNR and SSIM for T1-to-T2 and FLAIR-to-T1, and it reports strong scores on the pelvic MRI-to-CT task. The paper also claims that multi-scale random masking makes training efficient enough that 500 timesteps outperform DiT and CoLa-Diff trained with 1000 timesteps. The authors argue this would make scan replacement faster and more practical and would improve downstream segmentation when synthetic images are added to real training data.

What carries the argument

The load-bearing object is the pyramid of multi-scale images: each level is produced by resizing the previous one with scaling factor $\alpha = 0.5$, and synthesis moves from the lowest resolution upward so coarse structure is fixed before fine detail. At each level, the diffused image is patched and a fraction of patches are randomly masked; the visible patches are encoded by a Vision Transformer, and a conditional DiT decoder predicts the noise in the masked region. The cross-granularity regularization loss (Eq. 6) uses maximum mean discrepancy to match the predicted noise distributions across the low, middle, and high resolution levels, which the paper says preserves mutual-information consistency across granularities and improves pixel-level perceptual accuracy.

What would settle it

Run the released evaluation code with the ground-truth target image completely removed, so no target-derived tensor from Eq. (6) and no target input reaches the network at test time, and compare the resulting PSNR and SSIM with Table I; a clear drop would show the reported scores were aided by direct target access. A second check is to inspect the forward pass in the repository to see whether $\hat{Y}$ or any target-derived feature appears anywhere in the evaluation graph.

Watch

Extended reading notes

Core claim

The paper's core claim is that medical image synthesis is best organized as a coarse-to-fine pyramid. The source image is repeatedly halved in resolution, and the lowest-resolution level is denoised first, then upsampled and fused into the next level. At every level, a random high-proportion mask hides most patches, a Vision Transformer encodes only the visible patches, and a DiT-style decoder predicts the noise in the masked region while conditioning on the visible region. A cross-granularity regularization loss built from maximum mean discrepancy encourages the noise statistics at low, middle, and high resolutions to stay mutually consistent. The paper reports that this combination yields PSNR of 28.32 ± 1.16 dB and SSIM of 92.42 ± 1.53% for T1-to-T2 on BraTS, above all compared baselines, and that ablations show each component contributes to the final score.

Load-bearing premise

The paper's reported synthesis quality depends on the ground-truth target image $\hat{Y}$ entering the cross-granularity regularization statistic in Eq. (6), and it never states that $\hat{Y}$ is withheld from the model during inference; if $\hat{Y}$ is available when metrics are computed, the results do not actually measure image synthesis.

Editorial extensions

If this is right

  • If the reported numbers hold, clinical workflows could replace a missing or low-quality scan with a synthesized one while preserving anatomical structure, reducing scan time and contrast-agent exposure.
  • The 500-versus-1000 timestep result implies that hierarchical masking can roughly halve diffusion training cost at equal synthesis quality on these datasets.
  • The ablation study supports the conclusion that every component, pyramid hierarchy, MAE-style masking, diffusion, transformer backbone, and cross-granularity regularization, contributes to the top PSNR and SSIM scores.
  • The segmentation experiment implies that synthetic images from PHMDiff are useful training data, since combining them with real data raises Dice scores and lowers HD95 error.

Reading between the lines

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

  • A natural extension the paper does not explore is treating the per-level masking ratio as a tunable hyperparameter: the reported 500-step advantage suggests even more aggressive schedules might work for high-resolution 3D volumes.
  • The same coarse-to-fine masked diffusion template could be applied to other dense prediction tasks such as super-resolution, denoising, or missing-slice imputation, which share the need to preserve global structure while recovering local detail.
  • A targeted robustness test would remove the ground-truth target term from Eq. 6 and retrain; if the metrics hold, the regularization is a genuine synthesis prior, and if they drop, the reported gains partly encode direct target information.
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

6 major / 6 minor

Summary. The paper proposes PHMDiff, a pyramid hierarchical masked diffusion model for medical image synthesis. The method decomposes the input into a multi-resolution pyramid, applies adaptive random masking at each scale, and uses a Transformer-based diffusion model with a cross-granularity regularization (CGR) loss built on maximum mean discrepancy. The authors evaluate on the BraTS 2021 and pelvic MRI-CT datasets, reporting PSNR and SSIM improvements over GAN, MAE, and diffusion baselines, along with ablation studies, a timestep-efficiency comparison, and a downstream segmentation experiment. The source code is available at a GitHub repository.

Significance. If the reported results are accurate, the paper addresses a practically important problem: synthesizing missing medical imaging modalities with a coarse-to-fine pyramid diffusion approach that may reduce training cost while improving fidelity. The paper includes several strengths: a public code link, comparisons with a broad set of baselines, ablation of the main components, a downstream segmentation evaluation, and a t-SNE analysis. The claimed advantage of reaching or exceeding 1000-step diffusion baselines with 500 steps is also potentially valuable. However, the quantitative evidence is currently undermined by internal inconsistencies in the experimental tables and an unclear validation protocol, so the significance cannot be assessed reliably from the manuscript as written.

major comments (6)
  1. [Section IV-C4, Table II] The row labeled 'w/o MAE' in Table II is numerically identical to the Uni-GAN row in Table I for both tasks (26.46±1.47 / 87.31±1.15 and 26.12±1.25 / 87.04±0.93). Removing the MAE component from a pyramid diffusion model cannot be expected to reproduce a separate GAN baseline to four decimal places. This duplication means Table II cannot support the stated ablation conclusions or the SOTA comparison as reported. Please provide corrected ablation results and, if possible, the original per-fold experimental logs.
  2. [Section IV-C4, Table II] The caption of Table II states the tasks are T1→T2 and T1→T1ce, but the column headers both read 'T1→T1ce'. Moreover, the first PHMDiff entry in the table (28.32±1.16 / 92.42±1.53) is actually the T1→T2 result from Table I, while the second PHMDiff entry (29.49±1.34 / 93.58±0.87) does not appear anywhere else in the paper. The table must be fully reconciled with the caption and with Table I before any ablation claim can be evaluated.
  3. [Section IV-A and Section IV-B] The validation protocol is described inconsistently: Section IV-A says the pelvic dataset uses a split of 9 training, 2 validation, and 4 testing subjects, whereas Section IV-B states that a 5-fold cross-validation approach was employed. It is unclear which protocol produced Tables I, II, and III, and how the reported mean ± std values were computed across folds or subjects. Please specify the exact evaluation protocol for each dataset and each table.
  4. [Section III-B5, Eq. (6)] The CGR loss uses the symbol \hat Y inside the definition of m, but \hat Y is never defined explicitly. Earlier in the paper \hat Y is introduced as the synthesized output image. If \hat Y denotes the ground-truth target image, please state clearly that it is used only during training and that it is withheld at inference; if \hat Y denotes the model's own output, please explain how it is available inside the regularization term without circularity. The current text leaves the inference-time protocol ambiguous, which is critical for interpreting the reported synthesis metrics.
  5. [Section IV-C2 and Section IV-C3] For the MRI→CT task on the Pelvic dataset, the paper provides only a radar chart (Fig. 5) and no numeric PSNR/SSIM values with standard deviations. Since the paper claims superior performance on two datasets and reports statistical significance, the Pelvic results should be presented in a table with the same level of detail as the BraTS results.
  6. [Section III-C5 and Fig. 8] The claim that PHMDiff trained with 500 timesteps surpasses DiT and CoLa-Diff trained with 1000 timesteps is supported only by a single plot without error bars or statistical tests. Please provide quantitative values, the number of runs, and the test details, or moderate the claim accordingly.
minor comments (6)
  1. [Fig. 4 caption] The caption mentions T1→FLAIR and T1ce→T2 visual results, but Table I reports only T1→T2 and FLAIR→T1. Please clarify whether the visual results are representative and where the quantitative results for the additional tasks are reported.
  2. [Section IV-C2] The paper states that p-values are less than 0.05 but does not report actual p-values or the precise form of the paired t-test (e.g., paired across subjects or across slices, and whether multiple comparisons were corrected). Please include these details.
  3. [Section III-B5, Eq. (6)] The notation K(m, m') is not explained; please define m' as an independent sample and specify the kernel used in the experiments.
  4. [Abstract and Section I] The abstract says 'The source code will be released with the paper' and also 'The source code is available at ...'. Please align these statements so it is clear whether the code is currently available or will be released upon acceptance.
  5. [Section III-B2] The 'Encoder' paragraph repeats the full text of the 'Multi-scale Masking' paragraph verbatim. Please remove the duplication and present the masking and encoder descriptions separately.
  6. [Section III-B4] The subsection title 'Reserve diffusion process' appears to be a typo for 'Reverse diffusion process'; please correct it.

Circularity Check

1 steps flagged · score 3.0 of 10

Table II's 'w/o MAE' row is exactly the Uni-GAN row from Table I, so the ablation evidence for MAE's contribution is a relabeled baseline rather than a derived result; the main SOTA comparison is otherwise independent.

  1. renaming known result [Table II (Section IV.C.4 ablation study) vs Table I (Section IV.C.2 comparison with SOTA)]
    ""w/o MAE 26.46±1.47 87.31±1.15 26.12±1.25 87.04±0.93" (Table II); "Uni-GAN 26.46±1.47 87.31±1.15 26.12±1.25 87.04±0.93" (Table I); text: "The elimination of either the diffusion component or MAE resulted in lower scores""

    The row labeled 'w/o MAE' in the ablation table is numerically identical, to every reported decimal place and on both tasks, to the Uni-GAN baseline row in the main comparison table. A PHMDiff variant with the MAE component removed is a different architecture from Uni-GAN, so the identical statistics cannot be a measured ablation result; the row is the known Uni-GAN result relabeled as 'w/o MAE'. The paper then uses this row to conclude that removing MAE 'resulted in lower scores', i.e., to attribute a performance contribution to MAE. That attribution is therefore not derived from an ablation of PHMDiff but from a renamed external baseline, so the claimed ablation support reduces to a relabeling rather than to a measurement of the stated component removal.

full rationale

This is an empirical model paper rather than a derivation chain of analytic predictions from first principles, so most circularity categories do not apply. There is no load-bearing self-citation: the authors' prior diffusion paper [30] appears only as a related-work citation and is not used to justify PHMDiff's architecture, and no uniqueness theorem is invoked. The central quantitative claim in Table I compares PHMDiff against external baselines on BraTS and pelvic datasets; that comparison is self-contained and does not reduce to fitted constants or to the authors' own prior results. The main circularity-adjacent finding is in Table II: the 'w/o MAE' ablation row duplicates the Uni-GAN row from Table I exactly, and the table header repeats 'T1→T1ce' while omitting the stated T1→T2 column, so the ablation claim that each component contributes is not verifiable as reported. This is a partial, non-central reduction by relabeling: it undermines the component-attribution claim but not the independent SOTA comparison. Separately, Eq. (6) constructs the cross-granularity statistic m using the term (1−√ᾱ)Ŷ without stating whether Ŷ is the ground-truth target available only during training or the model output at inference; if the target were available at test time, the synthesis metrics would not measure synthesis, but the text introduces Ŷ as the synthesized output, so this is a clarity and soundness concern rather than a demonstrated circular step. Overall, the paper's core empirical comparison is not circular, but the ablation evidence contains one concrete relabeling of a known baseline, giving a score of 3.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper's central contribution is a new architecture, so its free parameters are mostly training hyperparameters; the main concern is the unstated role of the target modality in the CGR loss.

free parameters (4)
  • Pyramid scaling factor α = 0.5
    Set to 0.5 in Eq. (1) to halve resolution at each pyramid level; no sensitivity analysis provided.
  • Adaptive masking ratio r = Not specified
    Adaptive per pyramid level but the mapping from resolution/complexity to r is not defined; the paper claims high-proportion masks speed training without experiments varying r.
  • Number of diffusion timesteps T = 500 for PHMDiff, 1000 for DiT/CoLa-Diff
    Used for the faster-training claim (Fig. 8); no timing or compute comparison is given.
  • Training hyperparameters = lr=1e-6, batch=10, Adam
    Standard choices, but no sensitivity analysis; central claim does not depend heavily on them.
assumptions (3)
  • domain assumption Paired co-registered multi-modal images exist for training
    The method is supervised cross-modal synthesis; requires aligned source-target pairs, which limits applicability to unpaired clinical data.
  • standard math Gaussian diffusion forward process with a fixed variance schedule
    Assumed from DDPM literature; standard in diffusion models.
  • ad hoc to paper Target image Ŷ availability during training
    Eq. (6) includes Ŷ in the CGR computation; the paper does not state that Ŷ is excluded from the model inputs at inference.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pyramid Hierarchical Masked Diffusion Model for Imaging Synthesis." pith.science (2026). https://pith.science/paper/GVWA7WMZ

@misc{pith2026250716579,
  author       = {Pith},
  title        = {Pith review of: Pyramid Hierarchical Masked Diffusion Model for Imaging Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GVWA7WMZ}},
  note         = {Machine review of arXiv:2507.16579}
}
read the original abstract

Medical image synthesis plays a crucial role in clinical workflows, addressing the common issue of missing imaging modalities due to factors such as extended scan times, scan corruption, artifacts, patient motion, and intolerance to contrast agents. The paper presents a novel image synthesis network, the Pyramid Hierarchical Masked Diffusion Model (PHMDiff), which employs a multi-scale hierarchical approach for more detailed control over synthesizing high-quality images across different resolutions and layers. Specifically, this model utilizes randomly multi-scale high-proportion masks to speed up diffusion model training, and balances detail fidelity and overall structure. The integration of a Transformer-based Diffusion model process incorporates cross-granularity regularization, modeling the mutual information consistency across each granularity's latent spaces, thereby enhancing pixel-level perceptual accuracy. Comprehensive experiments on two challenging datasets demonstrate that PHMDiff achieves superior performance in both the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), highlighting its capability to produce high-quality synthesized images with excellent structural integrity. Ablation studies further confirm the contributions of each component. Furthermore, the PHMDiff model, a multi-scale image synthesis framework across and within medical imaging modalities, shows significant advantages over other methods. The source code is available at https://github.com/xiaojiao929/PHMDiff

Figures

Figures reproduced from arXiv: 2507.16579 by the authors.

Figure 1
Figure 1. Challenge in modeling reliable synthesized medical images. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our proposed framework. [12], [14]–[20]. However, GANs often encounter issues with unstable training dynamics and mode collapse, which impact the diversity and fidelity of the synthesized images [3]. B. Diffusion Model In response to these limitations, Denoising Diffusion Prob￾abilistic Models (DDPMs) have recently emerged as an effec￾tive alternative. DDPMs utilize a Markov chain-based process to it… view at source ↗
Figure 3
Figure 3. Illustration of our proposed masked diff architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustrative instances of synthetic images were demonstrated on the BraTS dataset for T1 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Quantitative comparison with other synthesis methods in MRI [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Illustrative instances of synthetic images on the Pelvic dataset for MRI [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: The t-SNE feature space visualization for the different model’s [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: experimental results of our PHMDiff compared with DiT and CoLa [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 29 canonical work pages

  1. [1]

    A review on medical imaging synthesis using deep learning and its clinical applications,

    T. Wang, Y . Lei, Y . Fu, J. F. Wynne, W. J. Curran, T. Liu, and X. Yang, “A review on medical imaging synthesis using deep learning and its clinical applications,” Journal of applied clinical medical physics, vol. 22, no. 1, pp. 11–36, 2021

  2. [2]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial networks,” Communications of the ACM , vol. 63, no. 11, pp. 139–144, 2020

  3. [3]

    On the convergence and mode collapse of gan,

    Z. Zhang, M. Li, and J. Yu, “On the convergence and mode collapse of gan,” in SIGGRAPH Asia 2018 Technical Briefs , 2018, pp. 1–4

  4. [4]

    Six-channel image representation for cross-domain object detection,

    T. Zhang, W. Ma, and G. Wang, “Six-channel image representation for cross-domain object detection,” in the 11th International Conference on Image and Graphics , 2021, pp. 171–184

  5. [5]

    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

  6. [6]

    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

  7. [7]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 000–16 009

  8. [8]

    Enhancenet: Single image super-resolution through automated texture synthesis,

    M. S. Sajjadi, B. Scholkopf, and M. Hirsch, “Enhancenet: Single image super-resolution through automated texture synthesis,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 4491– 4500

Show all 42 references
  1. [9]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

  2. [10]

    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 Proceedings of CVPR , 2017, pp. 1125–1134

  3. [11]

    Image synthesis in multi-contrast mri with conditional generative adversarial networks,

    S. U. Dar, M. Yurt, L. Karacan, A. Erdem, E. Erdem, and T. Cukur, “Image synthesis in multi-contrast mri with conditional generative adversarial networks,” IEEE transactions on medical imaging , vol. 38, no. 10, pp. 2375–2388, 2019

  4. [12]

    Un- paired mr to ct synthesis with explicit structural constrained adversarial learning,

    Y . Ge, D. Wei, Z. Xue, Q. Wang, X. Zhou, Y . Zhan, and S. Liao, “Un- paired mr to ct synthesis with explicit structural constrained adversarial learning,” in 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019) . IEEE, 2019, pp. 1096–1099

  5. [13]

    7t mri super-resolution with generative adversarial network,

    H. Q. Do, P. Bourdon, D. Helbert, M. Naudin, and R. Guillevin, “7t mri super-resolution with generative adversarial network,” in IS&T Electronic Imaging 2021 Symposium , 2021

  6. [14]

    Unpaired deep cross- modality synthesis with fast training,

    L. Xiang, Y . Li, W. Lin, Q. Wang, and D. Shen, “Unpaired deep cross- modality synthesis with fast training,” in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, in Conjunction with MICCAI

  7. [15]

    Missing mri pulse sequence synthesis using multi-modal generative adversarial network,

    A. Sharma and G. Hamarneh, “Missing mri pulse sequence synthesis using multi-modal generative adversarial network,” IEEE transactions on medical imaging , vol. 39, no. 4, pp. 1170–1183, 2019

  8. [16]

    Ea-gans: edge-aware generative adversarial networks for cross-modality mr image synthesis,

    B. Yu, L. Zhou, L. Wang, Y . Shi, J. Fripp, and P. Bourgeat, “Ea-gans: edge-aware generative adversarial networks for cross-modality mr image synthesis,” IEEE transactions on medical imaging , vol. 38, no. 7, pp. 1750–1762, 2019

  9. [17]

    Synthesize high-quality multi- contrast magnetic resonance imaging from multi-echo acquisition using multi-task deep generative model,

    G. Wang, E. Gong, S. Banerjee, et al. “Synthesize high-quality multi- contrast magnetic resonance imaging from multi-echo acquisition using multi-task deep generative model,” IEEE transactions on medical imag- ing, vol. 39, no. 10, pp. 3089–3099, 2020

  10. [18]

    Autoencoder- based collaborative attention gan for multi-modal image synthesis,

    B. Cao, H. Cao, J. Liu, P. Zhu, C. Zhang, and Q. Hu, “Autoencoder- based collaborative attention gan for multi-modal image synthesis,” IEEE Transactions on Multimedia , vol. 26, pp. 995–1010, 2023

  11. [19]

    Unified multi-modal image synthesis for missing modality imputation,

    Y . Zhang, C. Peng, Q. Wang, D. Song, K. Li, and S. K. Zhou, “Unified multi-modal image synthesis for missing modality imputation,” arXiv preprint arXiv:2304.05340, 2023

  12. [20]

    Unified multi-modal image synthesis for missing modality im- putation,

    ——, “Unified multi-modal image synthesis for missing modality im- putation,” IEEE Transactions on Medical Imaging , 2024

  13. [21]

    Srdiff: Single image super-resolution with diffusion probabilistic mod- els,

    H. Li, Y . Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y . Chen, “Srdiff: Single image super-resolution with diffusion probabilistic mod- els,” Neurocomputing, vol. 479, pp. 47–59, 2022

  14. [22]

    Implicit diffusion models for continuous super-resolution,

    S. Gao, X. Liu, B. Zeng, S. Xu, Y . Li, X. Luo, J. Liu, X. Zhen, and B. Zhang, “Implicit diffusion models for continuous super-resolution,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 10 021–10 030

  15. [23]

    A multimodal com- parison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis,

    G. M ¨uller-Franzes, J. M. Niehues, F. Khader, et al., “A multimodal com- parison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis,” Scientific Reports , vol. 13, no. 1, p. 12098, 2023

  16. [24]

    Denoising diffusion probabilistic models for 3d medical image generation,

    F. Khader, G. M ¨uller-Franzes, S. Tayebi Arasteh, T. Han, C. Haarburger, M. Schulze-Hagen, P. Schad, S. Engelhardt, B. Baeßler, S. Foersch et al. , “Denoising diffusion probabilistic models for 3d medical image generation,” Scientific Reports, vol. 13, no. 1, p. 7303, 2023

  17. [25]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2022, pp. 10 684–10 695

  18. [26]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” arXiv preprint arXiv:1706.03762, 2017

  19. [27]

    Vtgan: Semi-supervised retinal image synthesis and disease prediction using vision transformers,

    S. A. Kamran, K. F. Hossain, A. Tavakkoli, S. L. Zuckerbrod, and S. A. Baker, “Vtgan: Semi-supervised retinal image synthesis and disease prediction using vision transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 3235–3245

  20. [28]

    Ganbert: Generative adversarial networks with bidirectional encoder representations from transformers for mri to pet synthesis,

    H.-C. Shin, A. Ihsani, S. Mandava, S. T. Sreenivas, C. Forster, J. Cha, and A. D. N. Initiative, “Ganbert: Generative adversarial networks with bidirectional encoder representations from transformers for mri to pet synthesis,” arXiv preprint arXiv:2008.04393 , 2020

  21. [29]

    Ptnet: a high-resolution infant mri synthesizer based on transformer,

    X. Zhang, X. He, J. Guo, N. Ettehadi, N. Aw, D. Semanek, J. Posner, A. Laine, and Y . Wang, “Ptnet: a high-resolution infant mri synthesizer based on transformer,” arXiv preprint arXiv:2105.13993 , 2021

  22. [30]

    FgC2F-UDiff: Frequency-guided and Coarse-to-fine Unified Diffusion Model for Multi-modality Missing MRI Synthesis,

    X. Xiao, Q. Hu, and G. Wang, “FgC2F-UDiff: Frequency-guided and Coarse-to-fine Unified Diffusion Model for Multi-modality Missing MRI Synthesis,” IEEE Transactions on Computational Imaging , vol.10 1815 – 1828, 2024

  23. [31]

    Diffusion models as masked autoencoders,

    C. Wei, K. Mangalam, P.-Y . Huang, Y . Li, H. Fan, H. Xu, H. Wang, C. Xie, A. Yuille, and C. Feichtenhofer, “Diffusion models as masked autoencoders,” in Proceedings of CVPR , 2023, pp. 16 284–16 294

  24. [32]

    2d medical image syn- thesis using transformer-based denoising diffusion probabilistic model,

    S. Pan, T. Wang, R. L. Qiu, M. Axente, C.-W. Chang, J. Peng, A. B. Patel, J. Shelton, S. A. Patel, J. Roper et al. , “2d medical image syn- thesis using transformer-based denoising diffusion probabilistic model,” Physics in Medicine & Biology , vol. 68, no. 10, p. 105004, 2023

  25. [33]

    Cola-diff: Conditional latent diffusion model for multi-modal mri synthesis,

    L. Jiang, Y . Mao, X. Wang, X. Chen, and C. Li, “Cola-diff: Conditional latent diffusion model for multi-modal mri synthesis,” in International Conference on Medical Image Computing and Computer-Assisted Inter- vention. Springer, 2023, pp. 398–408

  26. [34]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

  27. [35]

    Generative moment matching networks,

    Y . Li, K. Swersky, and R. Zemel, “Generative moment matching networks,” in International conference on machine learning . PMLR, 2015, pp. 1718–1727

  28. [36]

    Mr and ct data with multiobserver delineations of organs in the pelvic area—part of the gold atlas project,

    T. Nyholm, S. Svensson, S. Andersson, J. Jonsson, M. Sohlin, C. Gustafsson, E. Kjell ´en, K. S ¨oderstr¨om, P. Albertsson, L. Blomqvist et al., “Mr and ct data with multiobserver delineations of organs in the pelvic area—part of the gold atlas project,” Medical physics , vol. ...

  29. [37]

    The rsna- asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,

    U. Baid, S. Ghodasara, S. Mohan, M. Bilello, et al. , “The rsna- asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification,” arXiv preprint arXiv:2107.02314 , 2021

  30. [38]

    The multimodal brain tumor image segmentation benchmark (brats),

    B. H. Menze, A. Jakab, S. Bauer, et al., “The multimodal brain tumor image segmentation benchmark (brats),” IEEE transactions on medical imaging, vol. 34, no. 10, pp. 1993–2024, 2014

  31. [39]

    Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features,

    S. Bakas, H. Akbari, A. Sotiras, M. Bilello, M. Rozycki, J. S. Kirby, J. B. Freymann, K. Farahani, and C. Davatzikos, “Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features,” Scientific data, vol. 4, no. 1, pp. 1–13, 2017

  32. [40]

    Transunet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,” arXiv preprint arXiv:2102.04306 , 2021

  33. [41]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research , vol. 9, no. 11, 2008

  34. [42]

    Lf-synthseg: Label-free brain tissue-assisted tumor synthesis and segmentation,

    P. Xu, J. Lyu, L. Lin, P. Cheng, and X. Tang, “Lf-synthseg: Label-free brain tissue-assisted tumor synthesis and segmentation,” IEEE Journal of Biomedical and Health Informatics , 2024

Pith tools

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