Benchmarking GANs, Diffusion Models, and Flow Matching for T1w-to-T2w MRI Translation

Alec Sargood; Andrea Moschetto; Daniele Rav\`i; Francesco Guarnera; Lemuel Puglisi; Pierluigi Dell'Acqua; Sebastiano Battiato

REVIEW 4 major objections 5 minor 23 references

Benchmarking GANs, Diffusion Models, and Flow Matching for T1w-to-T2w MRI Translation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that Pix2Pix, a conditional GAN, outperforms diffusion and flow matching models for T1w-to-T2w MRI translation across three public datasets when all models share the same U-Net backbone and training budget.

desk verdict Useful reproducible benchmark, but Table 1's MSE and PSNR are mathematically inconsistent, so the central ranking is not credible until recomputed. read the letter →

arxiv 2507.14575 v1 pith:RUZU3IZ5 submitted 2025-07-19 cs.CV cs.AI

Andrea Moschetto , Lemuel Puglisi , Alec Sargood , Pierluigi Dell'Acqua , Francesco Guarnera , Sebastiano Battiato , Daniele Ravì This is my paper · ORCID

classification cs.CVcs.AI

keywords T1w-to-T2wtranslationMRIsynthesisgenerativeadversarialnetworksdiffusionmodelsflowmatchingimage-to-imagePix2Pixmedicalimagingbenchmark

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 sets out to settle which generative family, GANs, diffusion models, or flow matching, produces the most usable synthetic MRI contrast when one modality must be inferred from another. It builds all three families on the same 2D U-Net backbone, with comparable training budgets, and tests them on T1w-to-T2w axial-slice translation using three public datasets of healthy adults. Its central claim is that Pix2Pix, a conditional GAN, wins on every quantitative metric (SSIM 0.862, PSNR 22.915, MSE 0.0054), produces the most clinically acceptable visual outputs, and runs in about 0.05 seconds per slice, far faster than the iterative methods. A correct result would give practical guidance for clinical MRI workflows where a missing contrast must be synthesized cheaply, and it would caution against assuming diffusion or flow models are automatically better than GANs on small, simple, low-dimensional medical tasks.

What carries the argument

The comparison is carried by a deliberate equalization: every model sits on the same 2D U-Net encoder-decoder backbone with skip connections, two residual blocks per resolution level, channel widths of 32, 64, 64, and 64, and attention at the last two resolution levels, and every model gets the same subject-level 80/5/15 split, batch size of 6, and up to 300 epochs with early stopping on validation loss. What differs between families is the training loss and conditioning scheme: Pix2Pix minimizes an adversarial loss plus a weighted L1 reconstruction loss; the diffusion model minimizes noise-prediction error with the T1w slice concatenated or injected through a ControlNet-style branch; the flow matching model regresses a conditional velocity field along a straight line from noise or directly from the T1w slice. This shared architecture and training budget is what turns the results into an inter-family ranking rather than a comparison of unrelated implementations.

What would settle it

Re-run the benchmark on the same three datasets but without early stopping, extending training to more epochs, widening the backbone's channel counts, and tuning each family's noise schedule and step count; if flow matching or diffusion then reaches or exceeds Pix2Pix's SSIM of 0.862 and PSNR of 22.915, the reported superiority is an artifact of the shared budget. A faster check is to plot validation losses across training: if the FM and diffusion curves are still descending when early stopping fires, the comparison stopped those families before they converged.

Watch

Extended reading notes

Core claim

The paper's central discovery, on its own terms, is that Pix2Pix is not merely acceptable but dominant in this regime. With a shared 2D U-Net backbone and a 300-epoch, early-stopped budget, Pix2Pix reaches SSIM 0.862, PSNR 22.915 dB, and MSE 0.0054, while the best flow-matching variant reaches SSIM 0.732 and the best diffusion variant reaches SSIM 0.469. Pix2Pix inference takes 0.05 seconds; flow matching variants take about 11.6 to 11.9 seconds; diffusion variants take 41 to 58 seconds. Visually, Pix2Pix preserves sharp anatomical boundaries and transfers white-matter lesions, flow matching sometimes inpaints lesions rather than reconstructing them, and diffusion methods blur or fail to transfer lesions. The paper attributes the gap to flow-based models overfitting on small datasets and simpler tasks, and it explicitly cautions that the finding may not generalize to larger datasets, higher-dimensional data, or more complex tasks.

Load-bearing premise

The load-bearing premise is that a single 300-epoch, early-stopped training budget and one shared U-Net backbone with 32-64-64-64 channels are representative configurations for all three model families; if diffusion or flow matching models need more training, more capacity, or adjusted schedules to reach their typical performance, the reported ranking describes the recipe, not an intrinsic winner.

Editorial extensions

If this is right

In low-data, 2D-slice MRI translation, clinical deployment should favor a GAN with a reconstruction loss: Pix2Pix gives the best fidelity and a per-slice inference cost of 0.05 seconds, suitable for near-real-time use.
Flow matching is the middle ground: Direct FM's higher SSIM (0.732 versus 0.715) suggests the T1w slice is more useful as the flow source than as a concatenated condition, but both FM variants remain roughly two orders of magnitude slower than Pix2Pix.
The diffusion baselines, as configured with 1000 noise steps and a linear schedule, are not clinically usable here, with SSIM values of 0.469 and 0.363 indicating substantial anatomical distortion.
The overfitting explanation implies that data volume is the key lever: flow and diffusion methods may close the gap as datasets grow, so the reported ranking is a small-data result rather than a universal ordering.
The authors state their conclusion is limited to this setting and may not carry over to larger datasets, higher-dimensional 3D volumes, or more complex translation tasks.

Reading between the lines

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

An unstated consequence is that this benchmark should not be read as GANs being better than diffusion on MRI in general; it is evidence about one fixed recipe, and that recipe may underpower the iterative methods by stopping them before convergence.
A direct test of the overfitting story is to train each family on progressively larger subsets of the same data, for example 25, 50, and 100 percent, and check whether the FM-to-Pix2Pix gap shrinks monotonically; the paper's explanation predicts that it will.
Neighboring problems that would stress the ranking are 3D volumetric translation, translation between contrasts with weaker structural correlation, and evaluation by downstream tasks such as lesion segmentation or radiologist preference, which raw SSIM, MSE, and PSNR do not measure.
Because the paper reports only aggregate means and standard deviations, a per-subject or per-slice breakdown could reveal whether Pix2Pix's margin is consistent or dominated by a few hard cases.

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.

Desk Editor's Note

read the letter

Here's my take on arXiv:2507.14575. The paper does something genuinely useful: it runs a clean, reproducible comparison of GAN, diffusion, and flow-matching models for 2D T1w-to-T2w MRI translation, with public code, three public datasets, subject-level splits, and a shared U-Net backbone. That is real work and the qualitative figures are informative. If the numbers are right, the conclusion that Pix2Pix wins on this kind of small-data, 2D task is believable and consistent with prior work like Akbar et al.

But the quantitative centerpiece does not hold up as reported. Table 1 gives MSE and PSNR per subject as mean ± std. For a fixed image scale, PSNR is a deterministic function of MSE: PSNR_i = 10 log10(MAX^2 / MSE_i). The reported Pix2Pix row has MSE = 0.0054 ± 0.0000 and PSNR = 22.915 ± 3.428. If the per-subject MSE values all round to 0.0054 at four decimals, they lie in a band that forces a PSNR std far below 3.428. The same contradiction appears for Concat. FM and Direct FM. Either the metrics are computed on different scales or aggregations than the text states, or the numbers are wrong. Since Table 1 is the entire quantitative basis for the ranking, this is load-bearing, not cosmetic.

Other soft spots, in order of severity: no significance testing, so we cannot tell whether the gaps between the mid-tier methods are noise; the ControlNet comparison uses a pretrained frozen backbone, so it is not on equal footing with the other models; and the claim that flow models 'overfit' is inferred from prior literature, never measured. These would be addressable in revision.

I would not cite the numbers as they stand, but the benchmark protocol and the public release are worth engaging with. This paper deserves a serious referee: the inconsistency is plausible to fix, and a corrected version would be a solid practical reference for people doing small-scale 2D MRI translation.

Referee Report

4 major / 5 minor

Summary. This paper reports a benchmark of three generative model families—GANs (Pix2Pix), conditional diffusion models (with input concatenation and with a ControlNet module), and flow matching (with concatenated noise and with direct T1w initialization)—for 2D axial T1w-to-T2w MRI translation. Experiments use three public datasets (IXI, HCP, CamCAN), a shared U-Net backbone, and a common training budget of up to 300 epochs with early stopping. The authors report SSIM, MSE, and PSNR, show qualitative examples for two test subjects, and compare inference time, parameters, and memory. The central claim is that Pix2Pix outperforms diffusion- and FM-based methods on structural fidelity, image quality, and computational efficiency, with an additional interpretive claim that flow-based models overfit on small datasets.

Significance. If the quantitative results are correct, this is a useful practical benchmark for an active clinical imaging task: it compares three model families under a unified implementation, uses publicly available datasets, makes code and models available, and reports resource usage in addition to quality metrics. The contribution is empirical rather than methodological, so its value depends heavily on the reliability of Table 1 and on whether the 'comparable settings' claim is actually satisfied. The overfitting interpretation in the abstract and conclusion is not supported by the reported measurements and should be reframed as a hypothesis.

major comments (4)

[Table 1 / Section 4.3] The reported MSE and PSNR values are internally inconsistent under the stated aggregation. Section 4.3 says SSIM, MSE, and PSNR are used to assess performance, and the Table 1 caption says metrics are 'mean ± standard deviation across the test set.' For images normalized to [0,1], PSNR_i = -10*log10(MSE_i). Under this standard relation, Pix2Pix's MSE = 0.0054 ± 0.0000 cannot co-exist with PSNR = 22.915 ± 3.428: a zero MSE standard deviation at four decimals constrains the per-subject MSE to roughly [0.00535, 0.00545], which yields a PSNR spread of about 0.08 dB, not 3.428 dB. The same contradiction appears for Concat. FM (MSE 0.0097 ± 0.0000 with PSNR std 1.143) and Direct FM (MSE 0.0106 ± 0.0000 with PSNR std 1.361). The mean values are also inconsistent for the diffusion rows: under [0,1] normalization, MSE = 0.5110 corresponds to PSNR ≈ 2.9 dB, not 15.892 dB, and MSE = 0.3887 corresponds to ≈ 4.1 dB, not 14.800 dB. Either the metrics are computed on different scales or with different aggregations (e.g., PSNR averaged per image while MSE is pooled over the whole test set), contradicting the stated protocol, or the numbers are erroneous. Because Table 1 is the principal evidence for the central claim, this must be corrected before the ranking can be assessed.
[Section 3.2 / Section 4.3] The ControlNet configuration violates the paper's 'comparable settings' claim. Section 3.2 states that ControlNet is trained on top of a 'pre-trained, frozen unconditional diffusion backbone,' while Section 4.3 says that 'each model employs a 2D U-Net encoder-decoder backbone' and Section 4.4 says that 'we train all networks' for up to 300 epochs with early stopping. A frozen pre-trained backbone is not trained under the same protocol as the other models, and Table 2 shows ControlNet has 3,255,745 parameters versus 2,328,737 for the other diffusion and FM models. The authors should either train all diffusion variants from scratch under identical conditions or explicitly report the ControlNet backbone's training setup and restrict the comparability claim to the models that were actually trained under the shared protocol.
[Abstract / Section 5.1 / Section 6] The claim that flow-based models 'are prone to overfitting on small datasets and simpler tasks' is not supported by the measurements in this paper. Table 1 reports only test-set SSIM, MSE, and PSNR; there is no training-versus-test gap, no memorization metric, and no dataset-size ablation. Lower test performance alone does not establish overfitting, and references [4,1] are external evidence rather than evidence from this study. The authors should either provide direct evidence of overfitting (for example, train/test gap curves, sample-memorization metrics, or an ablation across dataset sizes) or soften the wording to state that this is a hypothesis consistent with prior work.
[Section 5.1] The quantitative comparison is presented only as aggregate means over three datasets. Because the three datasets differ in acquisition protocol, size, and subject population, and because the paper claims a 'comprehensive benchmark,' the authors should report per-dataset results and, where appropriate, paired significance tests or confidence intervals for the Pix2Pix-versus-FM and Pix2Pix-versus-diffusion comparisons. Without this, the word 'demonstrably' in the description of Pix2Pix's superiority is not statistically supported.

minor comments (5)

[Section 4.2 / Table 1] The normalization range used for PSNR is never stated. Please specify whether pixel values are in [0,1] or [0,255] and provide the explicit PSNR formula used, so that Table 1 can be interpreted unambiguously.
[Section 5.2] The qualitative claims, such as 'most clinically acceptable and realistic synthetic images,' are based on visual inspection of two test subjects. No reader study, blinded evaluation, or inter-rater assessment is described. I suggest tempering the language or adding a small reader study to support claims of clinical acceptability.
[Section 5.3 / Table 2] The inference-time comparison conflates the chosen number of sampling steps with the model family: diffusion uses 1000 steps and FM uses 300 Euler steps, so the 0.05 s versus 11–12 s comparison is partly a step-count artifact. Please report time per function evaluation or add a discussion of how the ranking changes with matched step counts.
[Various] There are several typographical and formatting issues, including 'T able 1' in the caption, missing spacing in 'Lcf m' in Section 3.3, and inconsistent use of 'FM' versus 'flow matching' in the conclusions. These do not affect the technical content but should be cleaned up.
[Abstract / Conclusion] The conclusion correctly notes that the findings 'may not generalize to other settings,' but the abstract states the ranking without this scope qualifier. Aligning the abstract with the stated limitations would improve accuracy.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark ranking is an empirical comparison against held-out test data, not a derivation from fitted inputs.

full rationale

The paper's central claim that Pix2Pix outperforms diffusion and flow-matching methods on T1w-to-T2w translation is supported by quantitative metrics (SSIM, MSE, PSNR) computed on a held-out test split with strict subject-level separation, as described in Section 4.3. The compared model families use standard, externally published training objectives (Pix2Pix [13], DDPM [10], Flow Matching [15]) and public datasets (IXI, HCP, CamCAN); no parameter is fitted to the test set and then renamed a prediction. The only interpretive statements citing prior work (overfitting of flow models on small datasets, refs [1,4,9]) are external support for the observed pattern, not inputs that force the reported ranking. The known internal inconsistency in Table 1 between MSE and PSNR values is a data-quality or aggregation issue, not a circularity: it does not make the conclusion equivalent to its inputs by construction. No self-citation appears in the reference list, and no uniqueness theorem or ansatz is imported from the authors' prior work. Therefore the circularity burden is essentially zero.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The benchmark depends on several domain assumptions about the representativeness of a single 2D slice, the validity of SSIM, MSE, and PSNR as quality measures, the fairness of the shared backbone, and the early-stopping protocol. No new physical or mathematical entities are introduced.

free parameters (5)

Diffusion forward steps (T) = 1000
Chosen as a standard configuration; the ranking may shift if diffusion is given more steps or a different schedule.
Diffusion noise schedule endpoints (beta) = 1e-4 to 2e-2
Standard linear schedule; not tuned per dataset.
Flow matching Euler solver steps = 300
Used for all FM inference; more or fewer steps change quality and speed.
Pix2Pix L1 weight (lambda) = 100
Standard value from Pix2Pix; affects the tradeoff between realism and fidelity.
Training budget (epochs) = up to 300 with early stopping
Chosen as a comparable budget across methods; diffusion models often need more epochs, so the comparison measures a fixed-budget outcome.

assumptions (4)

domain assumption The central axial slice is a sufficient proxy for whole-volume T1w-to-T2w translation quality.
The entire benchmark and all conclusions are based on one 2D slice per subject (Section 4.2). If slice-level performance does not reflect volumetric clinical utility, the findings do not transfer to 3D workflows.
domain assumption SSIM, MSE, and PSNR capture image fidelity, structural preservation, and quality for this task.
Section 4.3 lists these as evaluation metrics; no perceptual or clinical reader evaluation is performed, so the 'best model' claim is metric-conditional.
domain assumption A shared U-Net backbone with the same channel counts is a fair capacity match across GAN, diffusion, and flow matching families.
Section 4.3 states 'Each model employs a 2D U-Net encoder-decoder backbone with skip connections.' If one family needs a different capacity or conditioning architecture to perform well, the comparison favors the family best matched to this backbone.
domain assumption Early stopping on validation loss selects models fairly for all families.
Section 4.3 says 'Training proceeds for up to 300 epochs, with early stopping based on validation loss.' Different model families may have different optimal stopping rules and validation metrics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Benchmarking GANs, Diffusion Models, and Flow Matching for T1w-to-T2w MRI Translation." pith.science (2026). https://pith.science/paper/RUZU3IZ5

@misc{pith2026250714575,
  author       = {Pith},
  title        = {Pith review of: Benchmarking GANs, Diffusion Models, and Flow Matching for T1w-to-T2w MRI Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RUZU3IZ5}},
  note         = {Machine review of arXiv:2507.14575}
}

abstract

Magnetic Resonance Imaging (MRI) enables the acquisition of multiple image contrasts, such as T1-weighted (T1w) and T2-weighted (T2w) scans, each offering distinct diagnostic insights. However, acquiring all desired modalities increases scan time and cost, motivating research into computational methods for cross-modal synthesis. To address this, recent approaches aim to synthesize missing MRI contrasts from those already acquired, reducing acquisition time while preserving diagnostic quality. Image-to-image (I2I) translation provides a promising framework for this task. In this paper, we present a comprehensive benchmark of generative models$\unicode{x2013}$specifically, Generative Adversarial Networks (GANs), diffusion models, and flow matching (FM) techniques$\unicode{x2013}$for T1w-to-T2w 2D MRI I2I translation. All frameworks are implemented with comparable settings and evaluated on three publicly available MRI datasets of healthy adults. Our quantitative and qualitative analyses show that the GAN-based Pix2Pix model outperforms diffusion and FM-based methods in terms of structural fidelity, image quality, and computational efficiency. Consistent with existing literature, these results suggest that flow-based models are prone to overfitting on small datasets and simpler tasks, and may require more data to match or surpass GAN performance. These findings offer practical guidance for deploying I2I translation techniques in real-world MRI workflows and highlight promising directions for future research in cross-modal medical image synthesis. Code and models are publicly available at https://github.com/AndreaMoschetto/medical-I2I-benchmark.

Figures

Figures reproduced from arXiv: 2507.14575 by the authors.

**Figure 1.** Visual comparison of predictions from different methods across two randomly selected test subjects. The first and second columns display the input T1w and groundtruth T2w slices, respectively. The remaining columns show the predictions from the evaluated generative models. Rows one and two correspond to the two different subjects, while row three displays a zoomed-in region from the second subject, highlighting an… view at source ↗

**Figure 2.** Visual comparison of predictions from different methods across two test subjects that show white matter lesions. The first and second columns display the input T1w and ground-truth T2w slices, respectively. The remaining columns show the predictions from the evaluated generative models. Rows one and two correspond to the two different subjects, while row three displays a zoomed-in region from the second subject, h… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 13 canonical work pages

[1]

Machine Learning: Science and Technology6(1), 015022 (2025)

Akbar, M.U., Wang, W., Eklund, A.: Beware of diffusion models for synthesizing medical images—a comparison with gans in terms of memorizing brain mri and chest x-ray images. Machine Learning: Science and Technology6(1), 015022 (2025)

work page 2025
[2]

Google Scholar (2016)

Alkan, C., Cocjin, J., Weitz, A.: Magnetic resonance contrast prediction using deep learning. Google Scholar (2016)

work page 2016
[3]

Medical image analysis 12(1), 26–41 (2008)

Avants, B.B., Epstein, C.L., Grossman, M., Gee, J.C.: Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Medical image analysis 12(1), 26–41 (2008)

work page 2008
[4]

arXiv preprint arXiv:2506.03719 (2025)

Bertrand, Q., Gagneux, A., Massias, M., Emonet, R.: On the closed-form of flow matching: Generalization does not arise from target stochasticity. arXiv preprint arXiv:2506.03719 (2025)

arXiv 2025
[5]

arXiv preprint arXiv:2211.02701 (2022)

Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

arXiv 2022
[6]

arXiv preprint arXiv:2503.07535 (2025) 12 A

Chadebec, C., Tasar, O., Sreetharan, S., Aubin, B.: Lbm: Latent bridge matching for fast image-to-image translation. arXiv preprint arXiv:2503.07535 (2025) 12 A. Moschetto et al

arXiv 2025
[7]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Choo, K., Jun, Y., Yun, M., Hwang, S.J.: Slice-consistent 3d volumetric brain ct-to-mri translation with 2d brownian bridge diffusion model. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 657–667. Springer (2024)

2024
[8]

In: Medical Imaging with Deep Learning

Dey, A., Ebrahimi, M.: Mtsr-mri: Combined modality translation and super- resolution of magnetic resonance images. In: Medical Imaging with Deep Learning. pp. 743–757. PMLR (2024)

work page 2024

Show all 23 references

[9]

Advances in neural information processing systems 34, 8780–8794 (2021)

Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780–8794 (2021)

2021
[10]

Advances in neural information processing systems 33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020)

2020
[11]

NeuroImage 260, 119474 (2022)

Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: Synthstrip: skull- stripping for any brain image. NeuroImage 260, 119474 (2022)

2022
[12]

In: ECCV

Huang, Y., Zheng, F., Sun, X., Li, Y., Shao, L., Zheng, Y.: Generalized brain image synthesis with transferable convolutional sparse coding networks. In: ECCV. pp. 183–199. Springer (2022)

2022
[13]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with condi- tional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125–1134 (2017)

2017
[14]

reports of practical Oncology and radiotherapy 26(1), 35–42 (2021)

Kawahara, D., Nagata, Y.: T1-weighted and t2-weighted mri image synthesis with convolutional generative adversarial networks. reports of practical Oncology and radiotherapy 26(1), 35–42 (2021)

2021
[15]

arXiv preprint arXiv:2210.02747 (2022)

Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

2022 arXiv
[16]

arXiv preprint arXiv:2410.02423 (2024)

Martin, S., Gagneux, A., Hagemann, P., Steidl, G.: Pnp-flow: Plug-and-play image restoration with flow matching. arXiv preprint arXiv:2410.02423 (2024)

2024 arXiv
[17]

Medical Physics 51(4), 2538–2548 (2024)

Pan, S., Abouei, E., Wynne, J., Chang, C.W., Wang, T., Qiu, R.L., Li, Y., Peng, J., Roper, J., Patel, P., et al.: Synthetic ct generation from mri using 3d transformer- based denoising diffusion model. Medical Physics 51(4), 2538–2548 (2024)

2024
[18]

In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

2015
[19]

NeuroImage: Clinical 6, 9–19 (2014)

Shinohara, R.T., Sweeney, E.M., Goldsmith, J., Shiee, N., Mateen, F.J., Calabresi, P.A., Jarso, S., Pham, D.L., Reich, D.S., Crainiceanu, C.M., et al.: Statistical normalization techniques for magnetic resonance imaging. NeuroImage: Clinical 6, 9–19 (2014)

2014
[20]

IEEE transactions on medical imag- ing 29(6), 1310–1320 (2010)

Tustison, N.J., Avants, B.B., Cook, P.A., Zheng, Y., Egan, A., Yushkevich, P.A., Gee, J.C.: N4itk: improved n3 bias correction. IEEE transactions on medical imag- ing 29(6), 1310–1320 (2010)

2010
[21]

In: Medical Imaging 2022: Image Processing

Vaidya, A., Stough, J.V., Patel, A.A.: Perceptually improved t1-t2 mri translations using conditional generative adversarial networks. In: Medical Imaging 2022: Image Processing. vol. 12032, pp. 505–511. SPIE (2022)

2022
[22]

arXiv preprint arXiv:2503.00266 (2025)

Yazdani, M., Medghalchi, Y., Ashrafian, P., Hacihaliloglu, I., Shahriari, D.: Flow matching for medical image synthesis: Bridging the gap between speed and quality. arXiv preprint arXiv:2503.00266 (2025)

2025 arXiv
[23]

In: Proceedings of the IEEE/CVF international conference on computer vision

Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3836–3847 (2023)

2023

Pith tools

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

[1] [1]

Machine Learning: Science and Technology6(1), 015022 (2025)

Akbar, M.U., Wang, W., Eklund, A.: Beware of diffusion models for synthesizing medical images—a comparison with gans in terms of memorizing brain mri and chest x-ray images. Machine Learning: Science and Technology6(1), 015022 (2025)

work page 2025

[2] [2]

Google Scholar (2016)

Alkan, C., Cocjin, J., Weitz, A.: Magnetic resonance contrast prediction using deep learning. Google Scholar (2016)

work page 2016

[3] [3]

Medical image analysis 12(1), 26–41 (2008)

Avants, B.B., Epstein, C.L., Grossman, M., Gee, J.C.: Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Medical image analysis 12(1), 26–41 (2008)

work page 2008

[4] [4]

arXiv preprint arXiv:2506.03719 (2025)

Bertrand, Q., Gagneux, A., Massias, M., Emonet, R.: On the closed-form of flow matching: Generalization does not arise from target stochasticity. arXiv preprint arXiv:2506.03719 (2025)

arXiv 2025

[5] [5]

arXiv preprint arXiv:2211.02701 (2022)

Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

arXiv 2022

[6] [6]

arXiv preprint arXiv:2503.07535 (2025) 12 A

Chadebec, C., Tasar, O., Sreetharan, S., Aubin, B.: Lbm: Latent bridge matching for fast image-to-image translation. arXiv preprint arXiv:2503.07535 (2025) 12 A. Moschetto et al

arXiv 2025

[7] [7]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Choo, K., Jun, Y., Yun, M., Hwang, S.J.: Slice-consistent 3d volumetric brain ct-to-mri translation with 2d brownian bridge diffusion model. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 657–667. Springer (2024)

2024

[8] [8]

In: Medical Imaging with Deep Learning

Dey, A., Ebrahimi, M.: Mtsr-mri: Combined modality translation and super- resolution of magnetic resonance images. In: Medical Imaging with Deep Learning. pp. 743–757. PMLR (2024)

work page 2024