Can Diffusion Models Bridge the Domain Gap in Cardiac MR Imaging?

Duygu Sarikaya; Kieran Zucker; Marc de Kamps; Nishant Ravikumar; Xin Ci Wong

REVIEW 3 major objections 6 minor 36 references

Can Diffusion Models Bridge the Domain Gap in Cardiac MR Imaging?

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A source-domain diffusion model can convert cardiac MR images from any other centre into anatomy-preserving synthetic images, and using them for training or test-time adaptation improves surface-based segmentation on unseen domains.

desk verdict The central claim is contradicted by Table 1; the DG results are worse than baseline on every metric, and DA is only shown on synthetic target data. read the letter →

arxiv 2508.06327 v1 pith:ZC2HPICD submitted 2025-08-08 cs.CV

Xin Ci Wong , Duygu Sarikaya , Kieran Zucker , Marc De Kamps , Nishant Ravikumar This is my paper

classification cs.CV

keywords domainshiftcardiacmagneticresonancediffusionmodelimagesegmentationadaptationgeneralisationsyntheticdatatest-time

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 show that a diffusion model trained on a single source domain can act as a bridge across cardiac MR domains. Given a reference image from any other centre, the model synthesises a version that matches the source domain's appearance while keeping the anatomy intact, so the reference's segmentation mask stays valid. The authors then train segmentation networks on synthetic source-domain data alone (domain generalisation) and also project target-domain test images into the source domain before inference (domain adaptation). They report that both strategies significantly improve surface-based segmentation metrics on an unseen target dataset compared with models trained on real data alone, without data augmentation or transfer learning. If correct, the contribution is a practical route to domain-shift robustness in settings where only one centre's data is available.

What carries the argument

The load-bearing object is the iterative refinement sampling loop in Algorithm 1. It starts from a heavily noised version of the reference image, runs the DDPM reverse process, and at each timestep computes the predicted clean image via the standard reparameterisation formula. It then updates the latent by taking a gradient step that minimises the $\ell^2$ distance between a linear low-pass filter $\varphi_D$ applied to the reference and to the generated image. The filter scale $D$ controls how much structural detail is inherited from the reference; the guidance strength $\tau$ controls the fidelity-versus-randomness trade-off. This guidance is what ties the synthetic image's gross anatomy t

What would settle it

Run Algorithm 1 on target images with $D=2$ and compare the low-pass-filtered versions of the reference and the synthetic output: measure the surface distance between the RV and myocardium contours of $\varphi_2(y_0)$ and $\varphi_2(\hat{y}_0)$. If these distances are non-negligible, or if a segmenter trained on synthetic source data predicts a materially different RV mask on the synthetic image than the reference's ground truth, the no-concept-shift assumption fails and the reported surface-metric gains would not transfer to real unseen data.

Watch

Extended reading notes

Core claim

A source-domain diffusion model can generate cardiac MR images structurally faithful to a reference but visually in the source domain. The mechanism is reference-guided reverse diffusion with a gradient constraint on low-pass-filtered features. Using M&Ms as source and ACDC as target, the paper shows this helps domain generalisation (training on synthetic source data) and test-time domain adaptation. Across all tested architectures, both strategies significantly improve surface-based metrics (p<0.01) over real-data baselines, with modest declines in overlap metrics. The authors interpret this as the model discarding domain-specific appearance while preserving anatomy.

Load-bearing premise

The whole approach rests on the assumption that the reference image's segmentation mask stays valid for the synthetic image—that is, that coarse, low-pass-filtered structural fidelity is enough to preserve the boundaries (left ventricle, right ventricle, myocardium) that the downstream segmenter needs, even for structures at or below the filter scale.

Editorial extensions

If this is right

A segmentation model trained exclusively on synthetic source-domain images performs better on unseen target images than the same architecture trained on real source images, for surface-based metrics.
Target-domain images can be adapted at inference time to the source domain, giving smoother, less variable predicted masks without retraining the segmenter.
The approach works without data augmentation or transfer learning, indicating that a single diffusion model can substitute for expensive multi-centre data collection in data-scarce settings.
Combining both strategies (mixed adaptation) yields the most consistent surface-based improvements, with modest declines in overlap metrics attributed to synthetic image instability.
The t-SNE analysis suggests the SD-DM aligns feature distributions between source and target, making representations from different vendors more indistinguishable.

Reading between the lines

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

The low-pass filter scale $D$ and guidance strength $\tau$ jointly control how much target-specific detail is preserved; choosing $D$ too large could erase small but clinically important structures (e.g., the right ventricle boundary) while keeping the reference mask valid, silently degrading real-world use.
The paper's domain-alignment evidence is qualitative (t-SNE); a quantitative distribution-distance metric (e.g., FID in image space) could be used to choose $D$ and $\tau$ per dataset and to verify whether alignment holds for more extreme vendor differences than M&Ms-to-ACDC.
The no-concept-shift assumption may break when source and target differ in label protocol or cardiac phase; the method's applicability to datasets with different label spaces would require re-validation, not just re-training.
Because the diffusion model is trained only on source images, it can only repaint appearance within the source distribution's expressive range; for target domains featuring new tissue contrasts (e.g., oedema, scar), the synthetic images might not represent the target's actual appearance, so the reported gains may be specific to the tested cohort.

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

Short version: the paper's advertised result — that both DG and DA beat training on real data alone — is contradicted by its own Table 1. The DG rows are worse than the real-data baseline on every metric, and the DA improvements are only measured against synthetic ACDC generated by the same diffusion model. So the headline claim should not be trusted.

What's actually new: applying Choi et al.'s ILVR / Gao et al.'s DDA to a pixel-wise segmentation task, with a mixed-adaptation variant, and evaluating on cardiac MR with three segmentation backbones. That's a legitimate incremental step, and the code is available. The qualitative figures show the synthetic images preserve anatomy reasonably well, and the t-SNE overlap is suggestive, though it's not a substitute for downstream segmentation numbers.

The problems are structural. First, the abstract says both strategies significantly improved surface metrics on unseen target data. Table 1 shows the DG setup (synthetic M&Ms trained, real ACDC tested) is worse than baseline for all six metrics for all three models. No Welch test turns that into p<0.01. The wording might have meant DA and mixed adaptation, but the text explicitly defines DG earlier, so it's misleading. Second, the DA evaluation is circular in an important sense: the test set is synthetic ACDC produced by the same SD-DM. A reviewer would want to see real ACDC results after test-time adaptation. Third, the hyperparameters D=2, N=25 were chosen using ablations on real ACDC, so the target data has already been seen. That leaks test information. Fourth, there is no comparison to simple denoising or intensity matching, which would be the natural baseline for a claim that the diffusion model's alignment, rather than generic smoothing, is what helps.

The method section is clear and the authors are honest about the overlap-metric declines, but the central claim is not established. The paper is probably useful as a building block for someone working on reference-guided diffusion for segmentation, and the code repo is a plus. But as a standalone result, I wouldn't cite it without re-running the evaluation on real target data.

For peer review: I'd desk reject in current form. It needs a major revision with an honest abstract, a real-target DA evaluation, and a non-generative baseline. A serious referee could help after that, but the current version doesn't deserve referee time because the primary claim is contradicted by the paper's own table.

Referee Report

3 major / 6 minor

Summary. The paper proposes a reference-guided diffusion model (SD-DM) trained on a source cardiac MR dataset (M&Ms) and uses it to synthesize images that resemble a given reference while shifting appearance toward the source domain. The authors evaluate three segmentation architectures (2D nnU-Net, 3D nnU-Net, vanilla U-Net) under two strategies: domain generalisation (DG), where segmentation models are trained on synthetic source-domain images and tested on real target-domain images (ACDC), and domain adaptation (DA), where target-domain images are first adapted with the diffusion model before segmentation. The abstract and Section 5 claim that both strategies significantly improved surface-based segmentation metrics compared with training on real data alone (Welch's t-test, p < 0.01).

Significance. If the claims were sound, the method would be practically valuable for data-scarce cardiac MR settings because it promises test-time domain alignment without transfer learning or augmentation. The paper provides code, uses standard baselines, reports multiple segmentation metrics, and includes qualitative and t-SNE evidence. However, the central claim is not supported by the paper's own Table 1: the DG rows show uniformly worse performance than the real-data baseline on every metric and architecture, and the DA evaluation is conducted on synthetic ACDC images produced by the same diffusion model rather than on the original real ACDC test set. The contribution, as currently evidenced, is therefore unsubstantiated.

major comments (3)

[Abstract; Section 5, Table 1] The central claim that both strategies significantly improved surface-based metrics is contradicted by Table 1 for the DG setting. For 2D nnU-Net, the DG row (synthetic M&Ms-trained model on real ACDC) is worse than the baseline on every metric: DSC 0.656 vs 0.839, IoU 0.520 vs 0.736, ASD 3.0 vs 1.6, ASSD 3.6 vs 1.6, HD 33.9 vs 24.0, HD95 12.1 vs 6.7. The same pattern holds for 3D nnU-Net and vanilla U-Net. A Welch's t-test cannot transform these results into p < 0.01 improvements. If the intended comparison in the abstract is a different subset of rows, the text must be rewritten and supported with the corresponding statistical details.
[Table 1; Sections 4 and 5] The DA and Mixed Adaptation results are evaluated on 'Synthetic ACDC' images generated by the same SD-DM, not on the original real ACDC test images. Thus the reported improvements in surface metrics may reflect the generative model smoothing or simplifying the test images rather than genuinely adapting them. No segmentation results on original real ACDC are reported for DA-adapted models, and no baseline using a simple denoising or intensity-matching preprocessing is given. This makes the DA claim circular: the model is being measured on data it produced, not on the unseen target domain.
[Section 3, Algorithm 1] The validity of the synthetic labels rests on the assumption stated in Section 3 that no concept shift occurs, i.e., that p(y|f(xs)) = p(y|xt). However, the low-pass-filter gradient guidance in Algorithm 1 line 10 can preserve only coarse structures, and the paper provides no quantitative evaluation of anatomical fidelity (e.g., Dice between the reference mask and the synthesized image's true anatomy). The DG results in Table 1, where synthetic-trained models degrade sharply on real ACDC, suggest that the synthetic data does not fully preserve all task-relevant structures. This load-bearing assumption is therefore unsupported.

minor comments (6)

[Author affiliation line] There is a typo: 'Leedsscxcw@leeds.ac.uk' should be separated or reformatted.
[Section 3] The notation p(f(xs)) = p(xs) is confusing; f is described as the diffusion model but the equation suggests a distributional equality that is not clearly defined.
[Algorithm 1] The symbols y_t^g, y_t, and the update rule are introduced abruptly. A short derivation or reference for the DDPM reparameterisation in line 9 would improve readability.
[Table 1] The table caption mentions 'blue underlined values', but these may not be visible in grayscale printing; consider using a more accessible formatting.
[Section 4] The number of images for synthetic ACDC and synthetic M&Ms is not stated, nor are the exact hyperparameters for the segmentation models (e.g., loss function, optimizer) beyond architecture names.
[Section 5] The claim that Welch's t-test gives p < 0.01 is not accompanied by the number of samples, the variance structure, or per-structure statistics; add this information to Appendix B.

Circularity Check

1 steps flagged · score 6.0 of 10

The DA 'improvement' is measured on synthetic ACDC produced by the proposed DM, making the central claim self-referential; the independent DG experiment contradicts it.

self definitional [Abstract; Section 5, Table 1 (DA with Synthetic ACDC vs Baseline with Real ACDC)]
"Both strategies significantly improved segmentation performance on data from an unseen target domain... compared to training segmentation models on real data alone. Table 1: 'Baseline with Real ACDC' vs 'DA with Synthetic ACDC'."

The DA rows are evaluated on 'Synthetic ACDC'—images generated by the proposed SD-DM (Algorithm 1) from real ACDC references—not on original real ACDC. The only experiment on real ACDC (DG rows) shows synthetic-trained models are worse than baseline on every metric (e.g., 2D nnU-Net DSC 0.656 vs 0.839). Thus the headline 'both strategies' claim is supported only by treating the method's own generated images as the unseen target domain. Surface-metric gains in DA rows can arise from the generator's low-pass-filter smoothing of boundaries, so the 'prediction' of improved segmentation on the unseen target is not an independent measurement; it is, by construction, an evaluation on the model's own output.

full rationale

The paper's DG experiment on real ACDC is an independent external check, but it contradicts the abstract's claim of significant improvement for both strategies: all DG rows are worse than baselines. The DA and Mixed Adaptation experiments, which show surface-metric improvements, are evaluated on synthetic ACDC generated by the very DM under test, so the 'unseen target domain' is redefined as the generator's output. This is a self-referential evaluation: the method produces the test data and then is credited for improving segmentation on that data, without any DA result on original real ACDC. The assumption 'no concept shift' (Section 3) is stated but not validated; if structural fidelity fails, the synthetic labels are invalid. No baseline with simple denoising or intensity matching is provided, so the surface-metric gains could be due to smoothing. These issues make the central claim partially circular and unsupported, though the underlying method is not a pure tautology.

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

The paper relies on the validity of reference-guided diffusion to preserve anatomy, the assumption of semantic homogeneity between cardiac MR datasets, and the absence of concept shift. Its main empirical parameters (D, N, tau) are hand-selected via ablations on the target test domain, which is a source of potential leakage.

free parameters (4)

guidance strength tau = 6
Fixed for all reported results; ablations shown for tau=1,18,24 in Fig. 5 but main experiments use tau=6 without justification.
scale factor D = 2
Chosen for main results; ablation across D values in Fig. 3 and used to set D=2 for reported tables.
diffusion range N = 25
Chosen for main results; ablations with N values in Fig. 4. Selection appears to be based on ablation on the target ACDC domain.
diffusion steps and noise schedule = 1000/2000/3000, linear/cosine
Experimented but not specified which configuration produced the final synthetic data.

assumptions (4)

domain assumption Source and target cardiac MR domains are semantically homogeneous (anatomy consistent across domains)
Stated in Section 3 to justify translating images without changing labels.
domain assumption No concept shift: p(y|f(xs)) = p(y|xt) because same annotation protocol used
Section 3; the segmentation labels for adapted images are assumed valid despite appearance changes.
ad hoc to paper The reverse diffusion process with low-pass-filter gradient guidance (Algorithm 1) maps target images to source distribution while preserving reference structure
This is the core mechanism; it is not proven but assumed from ILVR (ref [9]).
standard math DDPM training and sampling equations (Ho et al., Nichol et al.) are correct and applicable
Background method used without modification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Can Diffusion Models Bridge the Domain Gap in Cardiac MR Imaging?." pith.science (2026). https://pith.science/paper/ZC2HPICD

@misc{pith2026250806327,
  author       = {Pith},
  title        = {Pith review of: Can Diffusion Models Bridge the Domain Gap in Cardiac MR Imaging?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZC2HPICD}},
  note         = {Machine review of arXiv:2508.06327}
}

read the original abstract

Magnetic resonance (MR) imaging, including cardiac MR, is prone to domain shift due to variations in imaging devices and acquisition protocols. This challenge limits the deployment of trained AI models in real-world scenarios, where performance degrades on unseen domains. Traditional solutions involve increasing the size of the dataset through ad-hoc image augmentation or additional online training/transfer learning, which have several limitations. Synthetic data offers a promising alternative, but anatomical/structural consistency constraints limit the effectiveness of generative models in creating image-label pairs. To address this, we propose a diffusion model (DM) trained on a source domain that generates synthetic cardiac MR images that resemble a given reference. The synthetic data maintains spatial and structural fidelity, ensuring similarity to the source domain and compatibility with the segmentation mask. We assess the utility of our generative approach in multi-centre cardiac MR segmentation, using the 2D nnU-Net, 3D nnU-Net and vanilla U-Net segmentation networks. We explore domain generalisation, where, domain-invariant segmentation models are trained on synthetic source domain data, and domain adaptation, where, we shift target domain data towards the source domain using the DM. Both strategies significantly improved segmentation performance on data from an unseen target domain, in terms of surface-based metrics (Welch's t-test, p < 0.01), compared to training segmentation models on real data alone. The proposed method ameliorates the need for transfer learning or online training to address domain shift challenges in cardiac MR image analysis, especially useful in data-scarce settings.

Figures

Figures reproduced from arXiv: 2508.06327 by the authors.

**Figure 1.** The proposed reference-based cardiac MR synthesis using source domain-trained DM (SD-DM) can project target cardiac MR images to the source domain. The process involves a forward diffusion (blue arrow) process that transforms the reference image through noise injection, followed by iterative denoising of the transformed input via a reverse diffusion (red arrow) process, while conditioning on the reference with a de… view at source ↗

**Figure 2.** Overview of the Experimental Design. [A] In the sampling process, the real image from either the source domain (M&Ms) or the target domain (ACDC) is selected as reference. The trained DM has hyperparameters: scale factor D, diffusion range N, and controlling guidance strength τ , to refine the generative process and enhance similarity to the reference image. [B] For segmentation, we evaluate three models: 2D nnU-Ne… view at source ↗

**Figure 3.** Image synthesis using various scaling factor D and diffusion range N with fixed τ = 6. As D increases, more semantic information lost (yellow → green region) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 5.** Figure 5: Controlling structural consistency by extremal settings of D and τ . All images are synthesised using the same reference (indicated with a yellow arrow in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: t-SNE visualisation of (a) original datasets and (b) synthetic datasets. Each point represents a cardiac MR scan. (a) Distinct clustering patterns are observed for different manufacturers, indicating domain shifts. (b) The synthetic data exhibits improved overlap, for…

**Figure 7.** Figure 7: Performance metrics with 95% confidence intervals across anatomical structures. Each bar chart summarizes segmentation performance per structure—Left Ventricle (LV), Myocardium (MYO), and Right Ventricle (RV)—measured using: (a) Dice Similarity Coefficient (DSC), (b) I…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages

[1]

Abuduweili, A., Yuan, C., et al.: Enhancing sample generation of diffusion models using noise level correction (2025), https://arxiv.org/abs/2412.05488

work page Pith review arXiv 2025
[2]

Ahmad,F.,Hou,W.,Xiong,J.,Xia,Z.:FullyautomatedcardiacMRIsegmentation using dilated residual network. Med. Phys.50(4), 2162–2175 (Apr 2023)

work page 2023
[3]

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 (2024), https://arxiv.org/abs/2305.07644

work page Pith review arXiv 2024
[4]

Computers in Biology and Medicine 161, 106973 (2023)

Al Khalil, Y., Amirrajab, S., et al.: Reducing segmentation failures in cardiac mri via late feature fusion and gan-based augmentation. Computers in Biology and Medicine 161, 106973 (2023)

work page 2023
[5]

Ammar, A., Bouattane, O., Youssfi, M.: Automatic cardiac cine MRI segmentation and heart disease classification. Comput. Med. Imaging Graph.88(101864) (2021)

work page 2021
[6]

Bernard, O., Lalande, A., Zotti, et al.: Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: Is the problem solved? IEEE Trans. Med. Imaging37(11), 2514–2525 (Nov 2018)

work page 2018
[7]

IEEE Trans

Campello, V.M., Gkontra, P., Izquierdo, C., et al.: Multi-Centre, Multi-Vendor and Multi-Disease cardiac segmentation: The M&Ms challenge. IEEE Trans. Med. Imaging 40(12), 3543–3554 (Dec 2021)

work page 2021
[8]

Chen, C., Qin, C., Qiu, H., et al.: Deep learning for cardiac image segmentation: A review. Front. Cardiovasc. Med.7, 25 (Mar 2020)

work page 2020

Show all 36 references

[9]

2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp

Choi, J., Kim, S., et al.: Ilvr: Conditioning method for denois- ing diffusion probabilistic models. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) pp. 14347–14356 (2021), https://api.semanticscholar.org/CorpusID:236950721

2021
[10]

IEEE Transactions on Medi- cal Imaging43(8), 2924–2936 (2024)

Cui, H., Li, Y., et al.: Toward accurate cardiac mri segmentation with variational autoencoder-based unsupervised domain adaptation. IEEE Transactions on Medi- cal Imaging43(8), 2924–2936 (2024). https://doi.org/10.1109/TMI.2024.3382624

2024
[11]

Full, P.M., Isensee, F., Jäger, P.F., Maier-Hein, K.: Studying robust- ness of semantic segmentation under domain shift in cardiac mri (2020), https://arxiv.org/abs/2011.07592

2020 arXiv
[12]

In: 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

Gao, J., Zhang, J., et al.: Back to the Source: Diffusion-Driven Adaptation to Test- Time Corruption . In: 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR). pp. 11786–11796. IEEE Computer Society, Los Alami- tos, CA, USA (Jun 2023). https://doi.org/...

2023
[13]

In: Linguraru, M.G., Dou, Q., Feragen, A., Giannarou, S., Glocker, B., Lekadir, K., Schnabel, J.A

Gu, M., Thies, M., Mei, S., Wagner, F., et al.: Unsupervised domain adaptation usingsoft-labeledcontrastivelearningwithreversedmontecarlomethodforcardiac image segmentation. In: Linguraru, M.G., Dou, Q., Feragen, A., Giannarou, S., Glocker, B., Lekadir, K., Schnabel, J.A. (eds...

2024
[14]

IEEE Transactions on Biomedical Engineering 69(3), 1173–1185 (2022)

Guan, H., Liu, M.: Domain adaptation for medical image analysis: A sur- vey. IEEE Transactions on Biomedical Engineering 69(3), 1173–1185 (2022). https://doi.org/10.1109/TBME.2021.3117407

2022
[15]

Habijan, M., Babin, D., Galić, I., Leventić, H., Romić, K., Velicki, L., Pižurica, A.: Overviewofthewholeheartandheartchambersegmentationmethods.Cardiovasc. Eng. Technol.11(6), 725–747 (Dec 2020)

2020
[16]

In: Pro- ceedings of the 34th International Conference on Neural Information Processing Systems

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Pro- ceedings of the 34th International Conference on Neural Information Processing Systems. NIPS ’20, Curran Associates Inc., Red Hook, NY, USA (2020)

2020
[17]

Huang, J., Guan,D.,Xiao, A.,Lu,S.: Fsdr: Frequency space domain randomization for domain generalization (2021), https://arxiv.org/abs/2103.02370

2021 arXiv
[18]

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmen- tation (2024), https://arxiv.org/abs/2404.09556

2024 arXiv
[19]

Kong, Z., Ping, W.: On fast sampling of diffusion probabilistic models (2021), https://arxiv.org/abs/2106.00132

2021 arXiv
[20]

Diagnostics13(18), 2947 (2023)

Kushol, R., Wilman, A.H., Kalra, S., Yang, Y.H.: Dsmri: Domain shift analyzer for multi-center mri datasets. Diagnostics13(18), 2947 (2023)

2023
[21]

International Journal of Computer Vision133(1), 31–64 (Jul 2024)

Liang, J., He, R., Tan, T.: A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision133(1), 31–64 (Jul 2024). https://doi.org/10.1007/s11263-024-02181-w

2024 doi
[22]

Ma, J., He, Y., et al.: Segment anything in medical images. Nat. Commun.15(1), 654 (Jan 2024)

2024
[23]

Müller, D., Soto-Rey, I., Kramer, F.: Towards a guideline for evaluation metrics in medical image segmentation (2022), https://arxiv.org/abs/2202.05273

2022 arXiv
[24]

Nichol, A., Dhariwal, P.: Improved denoising diffusion probabilistic models (2021), https://arxiv.org/abs/2102.09672

2021 arXiv
[25]

Niemeijer, J., Schwonberg, M., Termöhlen, J.A., Schmidt, N.M., Fingscheidt, T.: Generalization by adaptation: Diffusion-based domain extension for domain- generalized semantic segmentation (2023), https://arxiv.org/abs/2312.01850

2023 arXiv
[26]

Patil, S.S., Ramteke, et al.: A domain-shift invariant CNN framework for cardiac MRI segmentation across unseen domains. J. Digit. Imaging36(5), 2148–2163 (Oct 2023)

2023
[27]

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation (2015)

2015
[28]

Skorupko, G., Osuala, R., et al.: Debiasing cardiac imaging with controlled latent diffusion models (2024), https://arxiv.org/abs/2403.19508

2024 arXiv
[29]

Thermos, S., Liu, X., O’Neil, A., Tsaftaris, S.A.: Controllable cardiac synthesis via disentangled anatomy arithmetic (2021), https://arxiv.org/abs/2107.01748

2021 arXiv
[30]

In: Statistical Atlases and Computational Models of the Heart

Ugurlu, D., Puyol-Antón, et al.: The impact of domain shift on left and right ventricle segmentation in short axis cardiac mr images. In: Statistical Atlases and Computational Models of the Heart. Multi-Disease, Multi-View, and Multi-Center Right Ventricular Segmentation in Ca...

2022
[31]

In: 2024 46th Annual International Confer- ence of the IEEE Engineering in Medicine and Biology Society (EMBC) (2024)

Urcia-Vázquez, J., Pérez-Pelegrí, M., Monmeneu, et al.: Generating realistic car- diac mr images using diffusion models. In: 2024 46th Annual International Confer- ence of the IEEE Engineering in Medicine and Biology Society (EMBC) (2024)

2024
[32]

Usman Akbar, M., Larsson, M., Blystad, I., Eklund, A.: Brain tumor segmentation using synthetic MR images - a comparison of GANs and diffusion models. Sci. Data 11(1), 259 (Feb 2024) Can Diffusion Models Bridge the Domain Gap in Cardiac MR Imaging? 13

2024
[33]

Wang, J., Lan, C., Liu, C., et al.: Generalizing to unseen domains: A survey on domain generalization (2022), https://arxiv.org/abs/2103.03097

2022 arXiv
[34]

BJR|Artificial Intelligence1(1) (Mar 2024)

Webber, G., Reader, A.J.: Diffusion models for medical image reconstruction. BJR|Artificial Intelligence1(1) (Mar 2024)

2024
[35]

IEEE Transactions on Medical Imaging40(12), 3555–3567 (2021)

Wu, F., Zhuang, X.: Unsupervised domain adaptation with variational approxi- mation for cardiac segmentation. IEEE Transactions on Medical Imaging40(12), 3555–3567 (2021). https://doi.org/10.1109/TMI.2021.3090412

2021
[36]

Proceedings of the IEEE 112(10), 1583–1609 (Oct 2024)

Yoon, J.S., Oh, K., Shin, Y., et al.: Domain generalization for medical image analysis: A review. Proceedings of the IEEE 112(10), 1583–1609 (Oct 2024). https://doi.org/10.1109/jproc.2024.3507831 A Appendix A: Vanilla U-Net Architecture We implemented a 2D vanilla U-Net model ...

2024

Pith tools

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

[1] [1]

Abuduweili, A., Yuan, C., et al.: Enhancing sample generation of diffusion models using noise level correction (2025), https://arxiv.org/abs/2412.05488

work page Pith review arXiv 2025

[2] [2]

Ahmad,F.,Hou,W.,Xiong,J.,Xia,Z.:FullyautomatedcardiacMRIsegmentation using dilated residual network. Med. Phys.50(4), 2162–2175 (Apr 2023)

work page 2023

[3] [3]

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 (2024), https://arxiv.org/abs/2305.07644

work page Pith review arXiv 2024

[4] [4]

Computers in Biology and Medicine 161, 106973 (2023)

Al Khalil, Y., Amirrajab, S., et al.: Reducing segmentation failures in cardiac mri via late feature fusion and gan-based augmentation. Computers in Biology and Medicine 161, 106973 (2023)

work page 2023

[5] [5]

Ammar, A., Bouattane, O., Youssfi, M.: Automatic cardiac cine MRI segmentation and heart disease classification. Comput. Med. Imaging Graph.88(101864) (2021)

work page 2021

[6] [6]

Bernard, O., Lalande, A., Zotti, et al.: Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: Is the problem solved? IEEE Trans. Med. Imaging37(11), 2514–2525 (Nov 2018)

work page 2018

[7] [7]

IEEE Trans

Campello, V.M., Gkontra, P., Izquierdo, C., et al.: Multi-Centre, Multi-Vendor and Multi-Disease cardiac segmentation: The M&Ms challenge. IEEE Trans. Med. Imaging 40(12), 3543–3554 (Dec 2021)

work page 2021

[8] [8]

Chen, C., Qin, C., Qiu, H., et al.: Deep learning for cardiac image segmentation: A review. Front. Cardiovasc. Med.7, 25 (Mar 2020)

work page 2020