Pith. sign in

REVIEW 4 major objections 7 minor 89 references

The paper argues that low-resolution, noisy-latent, and text tokens should evolve together in the same diffusion-transformer attention blocks for medical super-resolution, rather than the low-resolution input being injected one-way from out

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-01 09:52 UTC pith:ZELSB3SH

load-bearing objection Genuine tri-stream design idea with broad experiments, but overclaimed priority and missing error bars make the SOTA claim provisional. the 4 major comments →

arxiv 2607.20598 v1 pith:ZELSB3SH submitted 2026-07-22 eess.IV

MedDiT4SR: Tri-Stream Joint Adaptation of Pre-Trained Diffusion Transformers for Medical Image Super-Resolution

classification eess.IV
keywords medical image super-resolutiondiffusion transformermultimodal diffusionlow-resolution token streamsemantic alignmentdomain generalizationsuper-resolution adapter
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MedDiT4SR argues that the standard way of adapting a large pre-trained diffusion transformer to medical image super-resolution—feeding the low-resolution scan in through a separate one-way branch—is weaker than letting the low-resolution evidence evolve inside the model's joint attention alongside the noise and text streams. The paper proposes a tri-stream adaptation that turns the low-resolution scan into a third token stream in every multimodal diffusion transformer block, then adds a Super-Resolution Adapter for local scale-aware structure and a Semantic Alignment Refiner that gates low-resolution features with text-conditioned semantics. Across five modalities (cardiac MRI, abdominal CT, fundus, histology, breast ultrasound) and in within-modality cross-dataset tests, MedDiT4SR reports the best quantitative results and the best downstream segmentation scores, with strong performance using only 10% of the training data. A sympathetic reader should take the claim as: large natural-image diffusion priors can be steered for medical super-resolution if the degraded image is made a co-evolving participant in the denoising process rather than a fixed condition.

Core claim

On the paper's own terms, the central discovery is that joint three-stream attention is a more effective control mechanism than one-way injection. The low-resolution image is bicubically upsampled, encoded by the pre-trained VAE, and patched into a low-resolution token stream; those tokens are concatenated with noisy latent tokens and text tokens before each joint attention operation, so degradation information is updated at every denoising step. The SR Adapter performs scale-dependent strided-convolution fusion to compress interpolation-induced redundancy, and the SA Refiner uses a zero-initialized depth-wise convolution plus a text-conditioned token gate to modulate low-resolution features

What carries the argument

The central mechanism is a multimodal diffusion transformer block in which image and text tokens share joint attention; MedDiT4SR inserts a third low-resolution token stream into that joint attention so the LR, noise, and text representations update together. The SR Adapter adds local, scale-aware fusion through zero-initialized strided and transposed convolutions, while the SA Refiner produces a token-wise, text-conditioned gate from a zero-initialized depth-wise convolution and lightweight projections, applied inside the FFN.

Load-bearing premise

The load-bearing premise is that a pre-trained natural-image VAE and multimodal diffusion transformer, fed with bicubically upsampled single-channel medical images, produce a latent and feature space compatible enough for meaningful adaptation; if medical inputs are too far out of distribution, the reported gains may reflect the evaluation setup rather than the tri-stream design.

What would settle it

A concrete test would compare VAE and DiT feature statistics between natural images and single-channel medical images, and then train the tri-stream model with a medical-domain-adapted VAE; if performance does not degrade when the domain-adapted VAE replaces the original, the underlying domain-gap assumption is not load-bearing. Equally, an equal-parameter one-way ControlNet-style baseline that matches the tri-stream model's capacity would falsify the 'joint update' explanation if it matches the reported PSNR and SSIM.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Across five imaging modalities, MedDiT4SR achieves the best or competitive PSNR, SSIM, and FID at 4x super-resolution, suggesting the architecture transfers across MRI, CT, fundus, histology, and ultrasound without modality-specific redesign.
  • In within-modality cross-dataset tests (e.g., trained on BUSI and evaluated on BUSBRA and BUSUC), MedDiT4SR improves PSNR over all baselines at both 2x and 4x, indicating robustness to scanner and protocol shifts.
  • Downstream nnU-Net segmentation on MedDiT4SR-reconstructed images yields Dice and IoU closest to ground truth on ACDC and BUSI, suggesting the reconstructed images preserve clinically relevant structures.
  • With only 10% of the ACDC training data, MedDiT4SR outperforms the strongest baseline by about 3.4 SSIM points, indicating efficient use of large pre-trained priors under limited supervision.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the tri-stream design turns the low-resolution input into a learned, evolving condition rather than a fixed external signal; the same pattern may transfer to other conditional restoration tasks such as deblurring or inpainting, where the condition should be refined over diffusion steps.
  • Editorial inference: the ablation does not equalize parameter count between the one-way ControlNet-style baseline and the tri-stream model, so part of the reported gain could come from additional capacity; a matched-parameter one-way baseline would isolate whether joint updating itself is the cause.
  • Editorial inference: because captions are generated from low-resolution inputs by a vision-language model, prompt quality is intertwined with the semantic path; comparing oracle or expert-written prompts against generated prompts would quantify how much of the gain depends on caption fidelity.
  • Editorial inference: the SR Adapter keys its fusion to the upsampling scale, suggesting a natural extension to arbitrary-scale super-resolution by making the adapter's stride or scale a continuous input.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This paper proposes MedDiT4SR, an adaptation of the pre-trained Stable Diffusion 3 multimodal diffusion transformer (MM-DiT) to medical image super-resolution. The method augments the standard noise and text streams with a third low-resolution (LR) token stream that participates in the same joint attention, allowing LR evidence to co-evolve with denoising and semantic representations. Two lightweight modules are introduced: an SR Adapter that performs scale-aware local token aggregation to suppress bicubic interpolation redundancy, and a Semantic Alignment Refiner that gates LR features using text-derived scalar scores. The authors evaluate on five modalities (cardiac MRI, abdominal CT, fundus, histopathology, breast ultrasound) under both in-domain and within-modality cross-dataset settings, reporting PSNR/SSIM improvements over existing MedSR baselines including an SD3-ControlNet one-way conditioning baseline. Additional experiments cover ablations, data efficiency, prompt robustness, and downstream segmentation.

Significance. The paper has clear strengths: it ships code; the architecture is well motivated and the components are simple and plausible; the evaluation is broad, spanning five modalities and cross-dataset transfer; and the ablations and downstream segmentation evaluation are useful. The central claim — that tri-stream joint attention outperforms one-way ControlNet-style conditioning — is interesting and, if established with rigorous statistics, would be a valuable contribution to the growing literature on adapting large pre-trained diffusion models to medical imaging. Significance is currently tempered by the absence of error bars on the headline tables, the risk that the SD3-ControlNet baseline is undertuned, and the fully synthetic (bicubic) degradation setting. These issues should be resolved before the empirical claims can be considered robust.

major comments (4)
  1. [Tables 1-3] The main empirical claims rest entirely on point estimates without error bars, confidence intervals, or significance tests. Table 4 reports standard deviations of about 0.03 SSIM for ACDC; if similar variability applies to cross-dataset PSNR, a margin of 0.7-0.8 dB (e.g., Table 2, BUSI→BUSBRA ×2: MedDiT4SR 30.4 vs MedSRGAN 29.7) may not be statistically significant. Please report mean±std over at least three independent runs for every method and metric in Tables 1-3, and apply a paired significance test (e.g., Wilcoxon signed-rank or paired bootstrap) for the key comparisons to SD3-ControlNet and the best per-dataset baseline.
  2. [Table 3, SDK-ControlNet baseline] The ControlNet-style configuration is the critical control for the tri-stream contribution, yet row 1 of Table 3 shows a 2.0 dB PSNR and 2.5-point SSIM jump when moving to Tri-Stream + residual. This unusually large jump suggests the baseline may not have been tuned to its full potential. The paper does not state whether the ControlNet branch used zero-conv initialization, how many blocks were duplicated, or whether it was trained with the same steps, learning rate, batch size, and captions as MedDiT4SR. Please provide full implementation details for all baselines, including trainable parameter counts and compute, and consider grid-searching the ControlNet baseline (e.g., longer training, different LR, zero-conv init) to demonstrate that the tri-stream advantage is robust.
  3. [Experiments, Implementation Details] All LR–HR pairs are generated by bicubic downsampling (the MedSRGAN pipeline). The abstract and conclusion emphasize clinical relevance and faithful preservation of anatomical details, but real medical SR involves anisotropic acquisition, noise, and motion artifacts. The current validation therefore only supports a synthetic degradation scenario. Please add at least one experiment with a more realistic degradation model (e.g., Gaussian blur plus noise, or blind SR), or explicitly scope the claims to bicubic-downsampled SR. This is load-bearing for the clinical framing of the paper.
  4. [Implementation Details (Text Prompt Design)] Captions are generated by LLaVA-Med from the low-resolution input itself ('use LLaVA-Med to generate captions solely from the LR inputs'). The text stream therefore encodes information already present in the LR stream, so the SA Refiner's 'semantic guidance' is not independent external knowledge. The prompt-robustness experiment (Table 5) only compares variants of this LR-derived caption; it does not test whether an informative, externally grounded caption would improve results. Please clarify what additional information the text stream provides, and ideally compare against a fixed template, an empty prompt, and captions generated from the HR reference, to distinguish genuine semantic conditioning from a second, weaker copy of the LR signal.
minor comments (7)
  1. [Table 1] FID is reported only for ACDC, not for the other four modalities. Please report it consistently across all datasets or explicitly state why it is omitted for some modalities.
  2. [Introduction (Contributions)] The claim of being the 'first to investigate the adaptation of large-scale pre-trained diffusion transformers to medical image super-resolution' is too strong, since DiTMSR (Tu et al. 2025) and DiT4SR (Duan et al. 2025) already explore DiT-based SR. Please soften or clarify the novelty.
  3. [Figure 2] The caption mentions 'LLM Dataset Enc Dec C' but the text does not define what the 'Dataset Enc/Dec' branch represents. Please clarify the figure.
  4. [Method] The number of adapted MedDiT4SR blocks N is not specified. Please state which SD3 variant is used (e.g., SD3.5-Medium vs Large) and how many blocks are made trainable.
  5. [Table 3] The row labelled 'Linear' in the Semantic Guidance column is not described in the text. Please define what this baseline is, or remove it if it is the same as the SA Refiner without gating.
  6. [Table 4] Only SSIM is reported for the data-efficiency study. Adding PSNR would strengthen the conclusion.
  7. [General] No inference time or parameter count comparison is provided. Such information would help calibrate the practical cost of the proposed adapters relative to baselines.

Circularity Check

0 steps flagged

No circularity found: the paper is an empirical architecture study with held-out evaluations and no load-bearing self-citations.

full rationale

MedDiT4SR is an empirical systems paper; it contains no theoretical derivation chain whose predictions could reduce by construction to fitted inputs. The central claim—that tri-stream joint attention with the SR Adapter and SA Refiner outperforms ControlNet-style conditioning and prior MedSR methods—is supported by held-out test-set evaluations (Tables 1–2), cross-dataset generalization tests, ablations (Table 3), and downstream segmentation (Table 6). No equation in the paper defines an output in terms of a target quantity, and no fitted parameter is renamed as a prediction. The authors do not cite their own previous work, so no self-citation chain is load-bearing. The only mild self-reference is that text captions are generated by LLaVA-Med from the LR inputs themselves, making the text stream partially a re-encoding of the same evidence rather than independent anatomical knowledge. However, this is a data-processing design choice, not a circular derivation: the architecture is still trained and tested against held-out HR ground truth, and the text stream is only one of three input streams. The claim of 'semantic guidance' may be overstated, but that is an interpretation concern, not a circularity in the derivation. Baseline tuning and statistical significance are potential weaknesses but are outside the circularity rubric. The paper is self-contained against external benchmarks, so the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The method rests on transfer assumptions rather than derivations. The most consequential are: SD3's natural-image VAE/DiT is used on medical images without an explicit channel or domain adaptation step; LR pairs are synthetic bicubic degradations; and semantic conditioning is generated by LLaVA-Med from the LR image itself, so the 'text' stream is not independent of the image evidence. No new physical or mathematical entities are introduced; the LR Stream, SR Adapter, and SA Refiner are learned architectural components.

free parameters (5)
  • Semantic gate scale alpha = learned, initialized to 0
    In SA Refiner, G = 1 + alpha * tanh(W_L eta(L) + W_C Avg(C)); alpha is a trainable scalar controlling how strongly text-conditioned modulation affects LR features and is fit during training.
  • Classifier-free guidance scale = 8
    Set to 8 for all experiments; reconstruction quality and text-conditioning strength depend on this value, which is not ablated.
  • Number of sampling steps = 40
    SD3.5 default schedule with 40 steps; affects PSNR/SSIM/FID and is chosen without sensitivity analysis.
  • Learning rate = 3e-5
    Constant learning rate for all training; no schedule or sensitivity analysis reported.
  • Batch size = 32
    Training configuration; effect on stability and results is not investigated.
axioms (6)
  • domain assumption SD3's pre-trained MM-DiT and VAE priors transfer to medical imaging domains
    The entire method assumes natural-image priors help MRI/CT/fundus/histopathology/ultrasound reconstruction; no domain-adaptation or channel-mismatch analysis is provided (Method overview, Figure 2).
  • domain assumption Bicubic downsampling with a known scale is an adequate degradation model for MedSR
    LR-HR pairs are synthesized by downsampling HR to 256x256 (x2) or 128x128 (x4) and then bicubically upsampling to 512x512 (Implementation Details); real clinical degradations include noise, anisotropic resolution, motion, and scanner-specific blur.
  • ad hoc to paper A scalar, token-averaged text embedding is sufficient semantic conditioning for LR feature modulation
    SA Refiner averages text tokens over M tokens and projects to a scalar gate G, discarding per-token spatial and positional semantic information; this design is unvalidated and specific to this paper (Eq. G = 1 + alpha tanh(...)).
  • domain assumption LLaVA-Med captions generated from LR inputs are valid semantic conditions
    For datasets without captions, text prompts are generated solely from LR images (Implementation Details); there is no validation that these captions reliably describe HR anatomy.
  • ad hoc to paper Joint attention over concatenated LR/noise/text tokens preserves the pre-trained denoising behavior
    Adding an extra stream changes attention distributions and can shift pre-trained representations; stability is asserted via zero-init of adapter layers but not analyzed (Method, Figure 3).
  • domain assumption The SD3 VAE latent space handles single-channel medical images losslessly
    SD3's VAE expects RGB inputs, while medical images are often single-channel; the paper never describes the channel-mapping step, implying an unstated assumption that replication or conversion loses no relevant information (Method, VAE encoding).

pith-pipeline@v1.3.0-alltime-deepseek · 12289 in / 14237 out tokens · 110799 ms · 2026-08-01T09:52:17.921722+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of MedDiT4SR: Tri-Stream Joint Adaptation of Pre-Trained Diffusion Transformers for Medical Image Super-Resolution." pith.science (2026). https://pith.science/paper/ZELSB3SH

@misc{pith2026260720598,
  author       = {Pith},
  title        = {Pith review of: MedDiT4SR: Tri-Stream Joint Adaptation of Pre-Trained Diffusion Transformers for Medical Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZELSB3SH}},
  note         = {Machine review of arXiv:2607.20598}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Medical image super-resolution (MedSR) requires recovering fine anatomical structures from degraded observations while avoiding unsupported details introduced by generative priors. Large-scale pre-trained multimodal diffusion transformers provide strong visual priors, but their adaptation to MedSR remains non-trivial. In conventional ControlNet-style adaptation, the low-resolution (LR) image is processed as an external condition and injected into the denoising stream through one-way connections. Consequently, LR anatomical evidence cannot be jointly updated with the evolving denoising and semantic representations. We propose MedDiT4SR, a tri-stream adaptation framework that integrates the LR, noisy latent, and text representations into the same multimodal diffusion-transformer blocks. To complement global token interaction, we introduce a Super-Resolution Adapter (SR Adapter) that aggregates scale-dependent local tokens and suppresses interpolation-induced redundancy. We further propose a Semantic Alignment Refiner (SA Refiner) that calibrates local LR responses using prompt-conditioned semantic information. Experiments under both in-domain and within-modality cross-dataset settings demonstrate the effectiveness of adapting large-scale pre-trained DiT models to medical image super-resolution across diverse imaging domains.

Figures

Figures reproduced from arXiv: 2607.20598 by Le Zhang, Zhi Chen.

Figure 1
Figure 1. Figure 1: SD3-ControlNet processes LR features in a sepa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed MedDiT4SR framework for medical image super-resolution. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: MedDiT4SR block architecture. The LR, Noise, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of MedDiT4SR with representative baselines presented in Table 1 for [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 18 linked inside Pith

  1. [1]

    Ahmad, W.; Ali, H.; Shah, Z.; and Azmat, S. 2022. A new generative adversarial network for medical images super resolution. Scientific Reports, 12(1): 9533

  2. [2]

    Al-Dhabyani, W.; Gomaa, M.; Khaled, H.; and Fahmy, A. 2020. Dataset of breast ultrasound images. Data in brief, 28: 104863

  3. [3]

    Bernard, O.; Lalande, A.; Zotti, C.; Cervenansky, F.; Yang, X.; Heng, P.-A.; Cetin, I.; Lekadir, K.; Camara, O.; Ballester, M. A. G.; et al. 2018. Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37(11): 2514--2525

  4. [4]

    M.; Gkontra, P.; Izquierdo, C.; Martin-Isla, C.; Sojoudi, A.; Full, P

    Campello, V. M.; Gkontra, P.; Izquierdo, C.; Martin-Isla, C.; Sojoudi, A.; Full, P. M.; Maier-Hein, K.; Zhang, Y.; He, Z.; Ma, J.; et al. 2021. Multi-centre, multi-vendor and multi-disease cardiac segmentation: the M&Ms challenge. IEEE Transactions on Medical Imaging, 40(12): 3543--3554

  5. [5]

    Chen, H.; Wang, Y.; Guo, T.; Xu, C.; Deng, Y.; Liu, Z.; Ma, S.; Xu, C.; Xu, C.; and Gao, W. 2021. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12299--12310

  6. [6]

    Chen, X.; Wang, X.; Zhou, J.; Qiao, Y.; and Dong, C. 2023. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22367--22377

  7. [7]

    Cherti, M.; Beaumont, R.; Wightman, R.; Wortsman, M.; Ilharco, G.; Gordon, C.; Schuhmann, C.; Schmidt, L.; and Jitsev, J. 2023. Reproducible Scaling Laws for Contrastive Language-Image Learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2818–2829. IEEE

  8. [8]

    Dai, T.; Cai, J.; Zhang, Y.; Xia, S.-T.; and Zhang, L. 2019. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11065--11074

  9. [9]

    C.; He, K.; and Tang, X

    Dong, C.; Loy, C. C.; He, K.; and Tang, X. 2014. Learning a deep convolutional network for image super-resolution. In European conference on computer vision, 184--199. Springer

  10. [10]

    Duan, M.; Qu, L.; Yang, Z.; Wang, M.; Zhang, C.; and Song, Z. 2024. An efficient dual-branch framework via implicit self-texture enhancement for arbitrary-scale histopathology image super-resolution. arXiv:2401.15613

  11. [11]

    S.; Guo, C.-L.; and Li, C

    Duan, Z.-P.; Zhang, J.; Jin, X.; Zhang, Z.; Xiong, Z.; Zou, D.; Ren, J. S.; Guo, C.-L.; and Li, C. 2025. DiT4SR: Taming Diffusion Transformer for Real-World Image Super-Resolution. arXiv:2503.23580

  12. [12]

    Esser, P.; Kulal, S.; Blattmann, A.; Entezari, R.; Müller, J.; Saini, H.; Levi, Y.; Lorenz, D.; Sauer, A.; Boesel, F.; Podell, D.; Dockhorn, T.; English, Z.; Lacey, K.; Goodwin, A.; Marek, Y.; and Rombach, R. 2024. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. arXiv:2403.03206

  13. [13]

    J.; and Coelho de Albuquerque Pereira, W

    G \'o mez-Flores, W.; Gregorio-Calas, M. J.; and Coelho de Albuquerque Pereira, W. 2024. BUS-BRA: A breast ultrasound dataset for assessing computer-aided diagnosis systems. Medical physics, 51(4): 3110--3123

  14. [14]

    Gu, Y.; Zeng, Z.; Chen, H.; Wei, J.; Zhang, Y.; Chen, B.; Li, Y.; Qin, Y.; Xie, Q.; Jiang, Z.; et al. 2020. MedSRGAN: medical images super-resolution using generative adversarial networks. Multimedia Tools and Applications, 79(29): 21815--21840

  15. [15]

    Iqbal, A.; and Sharif, M. 2024. Memory-efficient transformer network with feature fusion for breast tumor segmentation and classification task. Engineering Applications of Artificial Intelligence, 127: 107292

  16. [16]

    F.; Kohl, S

    Isensee, F.; Jaeger, P. F.; Kohl, S. A.; Petersen, J.; and Maier-Hein, K. H. 2021. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2): 203--211

  17. [17]

    Ji, Y.; Bai, H.; Ge, C.; Yang, J.; Zhu, Y.; Zhang, R.; Li, Z.; Zhanng, L.; Ma, W.; Wan, X.; et al. 2022. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems, 35: 36722--36732

  18. [18]

    Kong, D.; Li, F.; Wang, Z.; Xu, J.; Pei, R.; Li, W.; and Ren, W. 2025. Dual Prompting Image Restoration with Diffusion Transformers. arXiv:2504.17825

  19. [19]

    Labs, B. F. 2024. FLUX. https://github.com/black-forest-labs/flux

  20. [20]

    Landman, B.; Xu, Z.; Igelsias, J.; Styner, M.; Langerak, T.; and Klein, A. 2015. Miccai multi-atlas labeling beyond the cranial vault--workshop and challenge. In Proc. MICCAI multi-atlas labeling beyond cranial vault—workshop challenge, volume 5, 12. Munich, Germany

  21. [21]

    X.; and Mou, L

    Li, C.; Shi, Y.; Hu, H.; Hu, J.; Zhu, X. X.; and Mou, L. 2025. Taming Stable Diffusion for Computed Tomography Blind Super-Resolution. arXiv:2506.11496

  22. [22]

    Li, C.; Wong, C.; Zhang, S.; Usuyama, N.; Liu, H.; Yang, J.; Naumann, T.; Poon, H.; and Gao, J. 2023. LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day. arXiv:2306.00890

  23. [23]

    Liu, K.; Ma, Y.; Xiong, H.; Yan, Z.; Zhou, Z.; Fang, P.; and Liu, C. 2019. Medical image super-resolution method based on dense blended attention network. arXiv:1905.05084

  24. [24]

    Liu, T.; Han, S.; Xie, L.; Xing, W.; Liu, C.; Li, B.; and Ta, D. 2024. Super-resolution reconstruction of ultrasound image using a modified diffusion model. Physics in Medicine & Biology, 69(12): 125026

  25. [25]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. arXiv:2103.14030

  26. [26]

    Lu, H.; Mei, J.; Qiu, Y.; Li, Y.; Hao, F.; Xu, J.; and Tang, L. 2025. Information sparsity guided transformer for multi-modal medical image super-resolution. Expert Systems with Applications, 261: 125428

  27. [27]

    Ma, J.; Zhang, Y.; Gu, S.; Ge, C.; Mae, S.; Young, A.; Zhu, C.; Yang, X.; Meng, K.; Huang, Z.; et al. 2024. Unleashing the strengths of unlabelled data in deep learning-assisted pan-cancer abdominal organ quantification: the flare22 challenge. The Lancet Digital Health, 6(11): e815--e826

  28. [28]

    Mo, W.; Xia, Y.; Yan, Y.; Zhou, H.; Zhang, L.; and Gao, H. 2026. Foveated-Imaging Geometry CT Architecture and Seeded Diffusion Model Enabling Global Super-Resolution Reconstruction. arXiv:2606.10390

  29. [29]

    I.; Fu, H.; Breda, J

    Orlando, J. I.; Fu, H.; Breda, J. B.; Van Keer, K.; Bathula, D. R.; Diaz-Pinto, A.; Fang, R.; Heng, P.-A.; Kim, J.; Lee, J.; et al. 2020. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical image analysis, 59: 101570

  30. [30]

    Peebles, W.; and Xie, S. 2023. Scalable Diffusion Models with Transformers. arXiv:2212.09748

  31. [31]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020

  32. [32]

    Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2023. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. arXiv:1910.10683

  33. [33]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. arXiv:2112.10752

  34. [34]

    Shin, J.; Hwang, A.; Kim, Y.; Kim, D.; and Park, J. 2025. Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing. arXiv:2508.07519

  35. [35]

    P.; Chen, H.; Qi, X.; Heng, P.-A.; Guo, Y

    Sirinukunwattana, K.; Pluim, J. P.; Chen, H.; Qi, X.; Heng, P.-A.; Guo, Y. B.; Wang, L. Y.; Matuszewski, B. J.; Bruni, E.; Sanchez, U.; et al. 2017. Gland segmentation in colon histology images: The glas challenge contest. Medical image analysis, 35: 489--502

  36. [36]

    Tu, X.; Li, G.; Fan, Z.; Ding, X.; and Liu, Y. 2025. Taming diffusion transformers for high-fidelity MRI super-resolution. Computers in Biology and Medicine, 198: 111261

  37. [37]

    von Platen, P.; Patil, S.; Lozhkov, A.; Cuenca, P.; Lambert, N.; Rasul, K.; Davaadorj, M.; Nair, D.; Paul, S.; Berman, W.; Xu, Y.; Liu, S.; and Wolf, T. 2022. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers

  38. [38]

    Woo, S.; Park, J.; Lee, J.-Y.; and Kweon, I. S. 2018. CBAM: Convolutional Block Attention Module. arXiv:1807.06521

  39. [39]

    Xin, B.; Ye, M.; Axel, L.; and Metaxas, D. N. 2023. Fill the K-Space and Refine the Image: Prompting for Dynamic and Multi-Contrast MRI Reconstruction. arXiv:2309.13839

  40. [40]

    W.; Saha, P

    You, C.; Cong, W.; Vannier, M. W.; Saha, P. K.; Hoffman, E. A.; Wang, G.; Li, G.; Zhang, Y.; Zhang, X.; Shan, H.; Li, M.; Ju, S.; Zhao, Z.; and Zhang, Z. 2020. CT Super-Resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE). IEEE Transactions on Medical Imaging, 39(1): 188–203

  41. [41]

    Zhang, L.; Rao, A.; and Agrawala, M. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. arXiv:2302.05543

  42. [42]

    E.; Pham, D

    Zhao, C.; Dewey, B. E.; Pham, D. L.; Calabresi, P. A.; Reich, D. S.; and Prince, J. L. 2020. SMORE: a self-supervised anti-aliasing and super-resolution algorithm for MRI using deep learning. IEEE transactions on medical imaging, 40(3): 805--817

  43. [43]

    Zhao, L.; Chi, H.; Zhong, T.; and Jia, Y. 2024. Perception-oriented generative adversarial network for retinal fundus image super-resolution. Computers in Biology and Medicine, 168: 107708

  44. [44]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Second-order attention network for single image super-resolution , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  45. [45]

    European conference on computer vision , pages=

    Learning a deep convolutional network for image super-resolution , author=. European conference on computer vision , pages=. 2014 , organization=

  46. [46]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Pre-trained image processing transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  47. [47]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Activating more pixels in image super-resolution transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [48]

    IEEE transactions on medical imaging , volume=

    SMORE: a self-supervised anti-aliasing and super-resolution algorithm for MRI using deep learning , author=. IEEE transactions on medical imaging , volume=. 2020 , publisher=

  49. [49]

    and Saha, Punam K

    You, Chenyu and Cong, Wenxiang and Vannier, Michael W. and Saha, Punam K. and Hoffman, Eric A. and Wang, Ge and Li, Guang and Zhang, Yi and Zhang, Xiaoliu and Shan, Hongming and Li, Mengzhou and Ju, Shenghong and Zhao, Zhen and Zhang, Zhuiyang , year=. CT Super-Resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE)...

  50. [50]

    Scientific Reports , volume=

    A new generative adversarial network for medical images super resolution , author=. Scientific Reports , volume=. 2022 , publisher=

  51. [51]

    2023 , eprint=

    Scalable Diffusion Models with Transformers , author=. 2023 , eprint=

  52. [52]

    2023 , eprint=

    PixArt- : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis , author=. 2023 , eprint=

  53. [53]

    2024 , eprint=

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author=. 2024 , eprint=

  54. [54]

    2024 , howpublished=

    Black Forest Labs , title=. 2024 , howpublished=

  55. [55]

    2025 , eprint=

    Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing , author=. 2025 , eprint=

  56. [56]

    2025 , eprint=

    DiT4SR: Taming Diffusion Transformer for Real-World Image Super-Resolution , author=. 2025 , eprint=

  57. [57]

    2025 , eprint=

    Dual Prompting Image Restoration with Diffusion Transformers , author=. 2025 , eprint=

  58. [58]

    2019 , eprint=

    Decoupled Weight Decay Regularization , author=. 2019 , eprint=

  59. [59]

    2022 , eprint=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. 2022 , eprint=

  60. [60]

    GitHub repository , howpublished =

    Patrick von Platen and Suraj Patil and Anton Lozhkov and Pedro Cuenca and Nathan Lambert and Kashif Rasul and Mishig Davaadorj and Dhruv Nair and Sayak Paul and William Berman and Yiyi Xu and Steven Liu and Thomas Wolf , title =. GitHub repository , howpublished =. 2022 , publisher =

  61. [61]

    2023 , eprint=

    Adding Conditional Control to Text-to-Image Diffusion Models , author=. 2023 , eprint=

  62. [62]

    2021 , eprint=

    Learning Transferable Visual Models From Natural Language Supervision , author=. 2021 , eprint=

  63. [63]

    Reproducible Scaling Laws for Contrastive Language-Image Learning , url=

    Cherti, Mehdi and Beaumont, Romain and Wightman, Ross and Wortsman, Mitchell and Ilharco, Gabriel and Gordon, Cade and Schuhmann, Christoph and Schmidt, Ludwig and Jitsev, Jenia , year=. Reproducible Scaling Laws for Contrastive Language-Image Learning , url=. doi:10.1109/cvpr52729.2023.00276 , booktitle=

  64. [64]

    2023 , eprint=

    Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer , author=. 2023 , eprint=

  65. [65]

    2021 , eprint=

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows , author=. 2021 , eprint=

  66. [66]

    2018 , eprint=

    CBAM: Convolutional Block Attention Module , author=. 2018 , eprint=

  67. [67]

    IEEE transactions on medical imaging , volume=

    Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved? , author=. IEEE transactions on medical imaging , volume=. 2018 , publisher=

  68. [68]

    Miccai multi-atlas labeling beyond the cranial vault--workshop and challenge , author=. Proc. MICCAI multi-atlas labeling beyond cranial vault—workshop challenge , volume=. 2015 , organization=

  69. [69]

    Medical image analysis , volume=

    Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs , author=. Medical image analysis , volume=. 2020 , publisher=

  70. [70]

    Medical image analysis , volume=

    Gland segmentation in colon histology images: The glas challenge contest , author=. Medical image analysis , volume=. 2017 , publisher=

  71. [71]

    Data in brief , volume=

    Dataset of breast ultrasound images , author=. Data in brief , volume=. 2020 , publisher=

  72. [72]

    Medical physics , volume=

    BUS-BRA: A breast ultrasound dataset for assessing computer-aided diagnosis systems , author=. Medical physics , volume=. 2024 , publisher=

  73. [73]

    Engineering Applications of Artificial Intelligence , volume=

    Memory-efficient transformer network with feature fusion for breast tumor segmentation and classification task , author=. Engineering Applications of Artificial Intelligence , volume=. 2024 , publisher=

  74. [74]

    Advances in neural information processing systems , volume=

    Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation , author=. Advances in neural information processing systems , volume=

  75. [75]

    IEEE Transactions on Medical Imaging , volume=

    Multi-centre, multi-vendor and multi-disease cardiac segmentation: the M&Ms challenge , author=. IEEE Transactions on Medical Imaging , volume=. 2021 , publisher=

  76. [76]

    The Lancet Digital Health , volume=

    Unleashing the strengths of unlabelled data in deep learning-assisted pan-cancer abdominal organ quantification: the flare22 challenge , author=. The Lancet Digital Health , volume=. 2024 , publisher=

  77. [77]

    arXiv preprint arXiv:2601.03267 , year=

    Openai gpt-5 system card , author=. arXiv preprint arXiv:2601.03267 , year=

  78. [78]

    Computers in Biology and Medicine , volume=

    Taming diffusion transformers for high-fidelity MRI super-resolution , author=. Computers in Biology and Medicine , volume=. 2025 , publisher=

  79. [79]

    2023 , eprint=

    Fill the K-Space and Refine the Image: Prompting for Dynamic and Multi-Contrast MRI Reconstruction , author=. 2023 , eprint=

  80. [80]

    2026 , eprint=

    Foveated-Imaging Geometry CT Architecture and Seeded Diffusion Model Enabling Global Super-Resolution Reconstruction , author=. 2026 , eprint=

Showing first 80 references.