Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Metrics that matter: Evaluating image quality metrics for medical image generation

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Many no-reference image quality metrics are blind to localised anatomical errors in synthetic MRI and can rank worse models as better.

desk verdict A well-designed stress-test of NRIQMs for medical image generation, with a real negative result, but the downstream proxy (Evanyseg) is unvalidated for synthetic inputs, making the central comparison conditional. read the letter →

arxiv 2505.07175 v1 pith:5M3RHUH2 submitted 2025-05-12 eess.IV cs.CV

classification eess.IVcs.CV
keywords no-referenceimagequalitymetricsgenerativemedicalimagingFIDKIDdownstreamsegmentationevaluationbrainMRIdatamemorisationmodecollapse
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

This paper tries to establish that widely used no-reference image quality metrics are not reliable for judging synthetic medical images, especially when clinical safety is at stake. It compares upstream metric scores with downstream segmentation performance on brain MRI data, under controlled perturbations and across three generative architectures. The paper finds that nearly all tested metrics are insensitive to clinically relevant local changes, such as tumour boundary blurring and internal intensity gradients, that do degrade segmentation quality. It also finds that the metrics can rank a GAN above a VAE for vessel images even though the GAN's output is much worse for the downstream vessel segmentation task.

What carries the argument

The key machinery is a two-phase comparative evaluation framework. Phase 1 applies controlled perturbations, including noise, morphological manipulations, and distribution shifts, to real BraTS and IXI images and measures both upstream NRIQM responses and downstream segmentation quality. Phase 2 applies the same suite of upstream metrics and the same downstream task to synthetic images from a VAE, a GAN, and a DDPM. The downstream yardstick is the Evanyseg quality predictor, a regression model described in Section 2.2.1 that predicts a Dice score from an input image and a segmentation mask without ground truth, and it is the instrument that exposes the divergence between metric scores and task utility.

What would settle it

Take the perturbed real BraTS images used in the morphological experiments, for which ground-truth tumour masks exist, and compute the true Dice between the segmentor's output and the ground truth; if Evanyseg's predicted Dice diverges from these true Dice as boundary blur or intensity gradient strength grows, the paper's downstream yardstick is miscalibrated.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that no-reference image quality metrics can detect global distributional differences but fail to reflect whether a generated medical image is usable for a downstream clinical task, and are largely blind to localised anatomical alterations. Concretely, tumour boundary blurring and internal intensity gradients left upstream NRIQMs essentially unchanged while the predicted tumour segmentation score dropped from 0.948 at baseline to 0.845 at the strongest blur, and on IXI the GAN scored better than the VAE on several distance metrics yet scored 0.67 versus 0.86 on vessel segmentation. This means upstream metric rankings can be inverted relative to task suitability, and a model whose images contain clinically significant structural flaws can pass common no-reference quality checks.

Load-bearing premise

The downstream yardstick, the Evanyseg quality predictor, is assumed to give valid predicted Dice scores when applied to perturbed real images and to synthetic images, even though it was trained only on real images with synthetically perturbed segmentation masks.

Editorial extensions

If this is right

  • Distance metrics like FID and KID can appear to improve when the generated set duplicates training data or drops rare anatomical classes, so a model that memorises or collapses modes may look better rather than worse.
  • No-reference image quality metrics cannot serve as a safety check for local anatomical fidelity: a synthetic tumour with blurred margins or altered internal texture can pass all tested no-reference checks.
  • Rankings produced by common distance metrics on medical image generations may invert the true task-based ranking, as with the GAN versus VAE comparison on IXI vessel segmentation.
  • Downstream segmentation alone is also insufficient, since it stayed near baseline under density shifts, mode collapse, and mode invention, so upstream and downstream evaluations must be combined.
  • Clinical deployment decisions based only on upstream metrics risk selecting models whose outputs mislead downstream decision-support tools.

Reading between the lines

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

  • A practical remedy implied by the findings is to report an authenticity-oriented metric such as AuthPct alongside distance metrics, because the paper shows the two respond in opposite directions to data duplication and can together flag memorisation.
  • The same local insensitivity is likely to appear in other imaging modalities such as CT or X-ray, since global feature statistics average away small regional changes; this is a testable extension rather than a claim of the paper.
  • The morphological perturbation setup could be turned into a calibration tool: for any candidate metric, define the smallest boundary blur or intensity gradient it can detect, and require that detection floor to sit below clinically significant changes.
  • The Evanyseg proxy could be validated against true Dice on the perturbed real images where ground truth masks exist; if its predictions diverge there, the upstream-versus-downstream comparisons in the paper would need to be re-based on a validated estimator.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper evaluates a suite of no-reference image quality metrics (NRIQMs) for assessing generative medical images, using brain MRI data (BraTS tumour images and IXI vascular MRA). In Phase 1, the authors apply controlled perturbations (noise, morphological manipulations, distribution shifts) to real images and measure (i) upstream metric responses via Z-score heatmaps and (ii) downstream segmentation quality via the Evanyseg predicted-Dice model. In Phase 2, they apply the same metrics and the Evanyseg-based downstream evaluation to outputs from pretrained VAE, GAN, and DDPM models. The central claims are that many upstream NRIQMs correlate poorly with downstream task suitability, are profoundly insensitive to localised morphological alterations, and can yield misleadingly optimistic scores under data memorisation and mode collapse; the authors recommend a multifaceted evaluation framework. The manuscript includes a literature review, detailed methodology, tables of Evanyseg scores, and an open-source GitHub repository, but the downstream yardstick is a regression model whose predictions are not validated for the perturbed and synthetic inputs to which it is applied.

Significance. If the findings are correct, the paper makes a useful empirical contribution by cataloguing the behaviour of a broad set of NRIQMs under clinically motivated perturbations and by demonstrating concrete ranking divergences between upstream metrics and a downstream segmentation proxy. The observation that distance metrics such as FID/KID can appear to improve under data duplication or mode collapse is a valuable caution that aligns with prior literature, and the open-source implementation is a reproducible asset. The clinical-safety framing is appropriate. However, the significance is conditional on the validity of the Evanyseg proxy, because every upstream-versus-downstream comparison in Sections 4.1 and 4.2 rests on it, and the paper does not provide the calibration evidence needed to establish that proxy as a faithful measure of true segmentation quality in the regimes used. The absence of confidence intervals, significance tests, and replicated model instances further limits the strength of the conclusions as stated.

major comments (4)
  1. [§2.2.1, §4.1.2, §4.2] The Evanyseg quality predictor is trained only on pairs of real images with synthetically perturbed segmentation masks, yet it is applied to (a) real images whose image content itself is perturbed (boundary blurring, intensity gradients, noise) and (b) fully synthetic images from VAE, GAN, and DDPM models. The paper provides no validation of Evanyseg's predicted Dice against true Dice in either regime, even though true Dice is computable for all Phase 1 perturbed real images because ground-truth masks exist. Since every upstream-versus-downstream comparison in Sections 4.1 and 4.2 depends on this proxy, a systematic bias in Evanyseg—for example, a learned association between boundary blur and low score from the mask-perturbation training data—could produce the observed drops in Table 3 (0.948 to 0.845) even if real segmentation accuracy were unchanged. The central claim that upstream metrics 'correlate poorly with downstream task suitability' is therefore not adequately supported until the proxy is calibrated.
  2. [§4.1, Tables 2–5, Figs. 4–7] All quantitative comparisons in Phase 1 are based on single runs with no confidence intervals, significance tests, or multiple random seeds. Statements such as 'profound insensitivity' and 'negligible changes' in Section 4.1.2 are not backed by any uncertainty quantification, and several reported differences are very small (e.g., Evanyseg scores of 0.907 versus 0.904 in Table 5, or the near-constant scores in Table 4). Without error bars or a hypothesis test, it is unclear whether these differences are meaningful effects or noise, and the strength of the paper's conclusions exceeds what the evidence supports.
  3. [§3.2, §4.2] The Phase 2 comparison uses a single pretrained instance of each generative architecture (VQ-VAE, WGAN, and latent diffusion model), with weights sourced from prior work or the Medigan library. There is no replication across seeds, training runs, or model variants. Consequently, the claimed VAE/GAN contradiction on IXI (Table 6: 0.86 versus 0.67) and the DDPM superiority could reflect model-specific hyperparameters, training data, or checkpoint quality rather than architecture-level properties. The conclusion that 'upstream metric rankings of generative models do not align with downstream segmentation performance' is stated as a general result but is supported only by an anecdotal comparison of single checkpoints.
  4. [§3.1.2, §4.1.2] In the morphological manipulation experiments on real BraTS images, the paper reports only Evanyseg predicted scores and not the actual Dice between the segmentor's output and the ground-truth masks, although those ground-truth masks are available. Reporting true Dice in this setting would (i) provide a direct, validated measure of downstream degradation under boundary blurring and intensity gradients, and (ii) serve as an in-phase calibration check for the Evanyseg proxy. Without this information, the claim that the downstream task 'registered a response' rests entirely on the unvalidated predictor.
minor comments (5)
  1. [§2.1] The metric is introduced as 'Ct Score (Clusterability Score)' but later referred to as 'CT Score' in Section 2 and Table 1; consistent notation would avoid confusion.
  2. [Table 1] Several entries contain artifact-like editorial notes, e.g., '(Note: Original formula matched FID, likely incorrect for likelihood)' for FLS and '(Note: Original formula seemed incomplete/unclear)' for Ct Score; these should be cleaned up and replaced with accurate formulas or removed.
  3. [Figure 2] The diagram contains the typo 'Upsteam' for 'Upstream' in the 'Upsteam' box.
  4. [Abstract and §5] The abstract claims that metrics 'correlate poorly' with downstream suitability, but no correlation coefficient is computed anywhere; the evidence is a qualitative divergence in rankings across two datasets and three models. Consider qualifying the language to 'rankings diverge' or reporting an actual correlation metric.
  5. [References] Reference [32] (Han et al., GAN-based synthetic brain MR image generation) is cited as the source for the Wasserstein GAN architecture, but the WGAN was originally proposed by Arjovsky et al.; the citation should be corrected or the discrepancy clarified.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; empirical metric comparison with external downstream proxy. Minor self-citations supply pre-trained models but do not encode the conclusions.

full rationale

The paper contains no derivation chain of the kind that can be circular: it is an empirical stress-test of standard NRIQMs against a downstream segmentation proxy. The upstream metrics (FID, KID, LPIPS, PRDC, etc.) are external, pre-existing tools with fixed definitions, and their scores are computed on held-out perturbed/synthetic images rather than fitted to the paper's conclusions. The downstream yardstick, the Evanyseg score, is adopted from Senbi et al. [71] and trained on real images with synthetically perturbed masks against ground-truth Dice; it is not trained on the paper's own experimental conditions or optimised to produce the reported upstream/downstream discrepancy. The reported ranking of generative models (DDPM > VAE > GAN on IXI; DDPM > GAN on BraTS) is therefore an empirical output of independent model weights and segmentors, not a quantity forced by construction. The paper's self-citations ([18], [19], [51]) provide pre-trained segmentors and generative model weights; these are legitimate experimental dependencies rather than load-bearing circular justifications, since the conclusions do not reduce to the existence or definitions of those models. The manuscript explicitly acknowledges the dependency in its Limitations: 'the downstream evaluation relied on a specific pre-trained segmentor and the Evanyseg method, introducing dependencies on these components.' Whether the Evanyseg proxy is well calibrated on synthetic or morphologically altered inputs is a validity/correctness concern, not a circularity concern. No equation in the paper is shown to be equivalent to its own input, and no fitted parameter is renamed as a prediction. Score 2 reflects only minor, non-load-bearing self-citation in the choice of experimental artifacts.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new theoretical entities or free parameters. It relies on standard pre-trained feature extractors, pre-trained generative models, pre-trained segmentors, and an adopted ground-truth-free quality predictor. The main assumptions are domain-level: that the datasets, perturbations, and the Evanyseg proxy are valid instruments for measuring what the paper claims to measure.

assumptions (6)
  • domain assumption BraTS and IXI brain MRI datasets are representative exemplars for medical image generation evaluation.
    The study generalizes from two brain MRI datasets; the authors acknowledge in Limitations that results may differ across modalities and tasks.
  • domain assumption Controlled perturbations faithfully model the clinical failure modes of interest (noise, morphological inaccuracy, memorisation, mode collapse, density shift, mode invention).
    The validity of sensitivity conclusions depends on whether Gaussian blurring of tumour boundaries and contrast remapping actually resemble generative model errors; this is asserted, not empirically verified.
  • domain assumption The Evanyseg regression model transfers to perturbed and synthetic images without loss of calibration.
    Section 2.2.1 trains the quality predictor on real images with perturbed masks; using it on synthetic images assumes distributional robustness that is not tested.
  • domain assumption InceptionV3 features, possibly augmented by the findings of Woodland et al. [78], are an appropriate feature space for medical images in these metrics.
    The paper does not run its own feature-extractor comparison; it adopts InceptionV3 based on prior work.
  • domain assumption The selected pre-trained generative models (VQ-VAE, WGAN, LDM) are representative of their architecture families.
    Only one checkpoint per architecture is used for each dataset; architecture-level conclusions rest on a single instance.
  • domain assumption Pre-trained segmentation models for vessels [18] and tumours [51] are accurate enough to serve as downstream tasks.
    The downstream evaluation assumes these segmentors produce clinically meaningful masks when applied to perturbed and synthetic images.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Metrics that matter: Evaluating image quality metrics for medical image generation." pith.science (2026). https://pith.science/paper/5M3RHUH2

@misc{pith2026250507175,
  author       = {Pith},
  title        = {Pith review of: Metrics that matter: Evaluating image quality metrics for medical image generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5M3RHUH2}},
  note         = {Machine review of arXiv:2505.07175}
}
read the original abstract

Evaluating generative models for synthetic medical imaging is crucial yet challenging, especially given the high standards of fidelity, anatomical accuracy, and safety required for clinical applications. Standard evaluation of generated images often relies on no-reference image quality metrics when ground truth images are unavailable, but their reliability in this complex domain is not well established. This study comprehensively assesses commonly used no-reference image quality metrics using brain MRI data, including tumour and vascular images, providing a representative exemplar for the field. We systematically evaluate metric sensitivity to a range of challenges, including noise, distribution shifts, and, critically, localised morphological alterations designed to mimic clinically relevant inaccuracies. We then compare these metric scores against model performance on a relevant downstream segmentation task, analysing results across both controlled image perturbations and outputs from different generative model architectures. Our findings reveal significant limitations: many widely-used no-reference image quality metrics correlate poorly with downstream task suitability and exhibit a profound insensitivity to localised anatomical details crucial for clinical validity. Furthermore, these metrics can yield misleading scores regarding distribution shifts, e.g. data memorisation. This reveals the risk of misjudging model readiness, potentially leading to the deployment of flawed tools that could compromise patient safety. We conclude that ensuring generative models are truly fit for clinical purpose requires a multifaceted validation framework, integrating performance on relevant downstream tasks with the cautious interpretation of carefully selected no-reference image quality metrics.

Figures

Figures reproduced from arXiv: 2505.07175 by the authors.

Figure 1
Figure 1. A graph of different types of metrics used in generative medical image evaluation, categorised by modality. Many metrics developed on natural image datasets do not adequately address the unique complexities of medical images, such as subtle pathological variations, noise sensitivity, and domain shifts arising from diverse scanner configurations. While specialised metrics tai￾lored to medical imaging have been propos… view at source ↗
Figure 1
Figure 1. These metrics evaluate image quality based [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of our methodology. datasets as our base image sets. We then create mul￾tiple distinct perturbed image sets by applying con￾trolled manipulations to copies of these base images. These manipulations, detailed in the following sec￾tions, simulate specific challenges including noise in￾jection, morphological alterations, and various forms of distribution shift (e.g., data memorisation, mode changes). For metri… view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Overview of the controlled perturbation experiments in Phase 1. We apply noise, mor [PITH_FULL_IMAGE:figures/full_fig_p013_3.png]
Figure 4
Figure 4. Figure 4: Heatmap illustrating the sensitivity of each evaluation metric to various forms of noise [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Heatmap illustrating the sensitivity of upstream evaluation metrics to morphological pertur [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Heatmap illustrating the sensitivity of upstream evaluation metrics to increasing levels of [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Heatmap illustrating the sensitivity of upstream evaluation metrics to density shift, mode [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Heatmap illustrating the sensitivity of upstream evaluation metrics across different gener [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Content Generation Models in Computational Pathology: A Comprehensive Survey on Methods, Applications, and Challenges

    eess.IV 2025-05 conditional novelty 3.0 of 10

    A systematic review that taxonomizes over 150 generative-model papers in computational pathology into four generation-task domains and discusses datasets, evaluation, and clinical barriers.

Reference graph

Works this paper leans on

81 extracted references · 76 canonical work pages · cited by 1 Pith paper

  1. [1]

    Virtual clinical trials in medical imaging: a review.J Med Imag, 7(4):042805–042805, 2020

    Ehsan Abadi, William P Segars, Benjamin MW Tsui, Paul E Kinahan, Nick Bottenus, Alejandro F Frangi, Andrew Maidment, Joseph Lo, and Ehsan Samei. Virtual clinical trials in medical imaging: a review.J Med Imag, 7(4):042805–042805, 2020

  2. [2]

    Bobomirzaevich Abdusalomov, Rakhmatillo Rakhmatullaevich Nasimov, Nodirabegim Rakhmatullaevna Nasimova, Bobir Mirzayevich Muminov, and Taeg Keun Whangbo

    A. Bobomirzaevich Abdusalomov, Rakhmatillo Rakhmatullaevich Nasimov, Nodirabegim Rakhmatullaevna Nasimova, Bobir Mirzayevich Muminov, and Taeg Keun Whangbo. Evaluating synthetic medical images using artificial intelligence with the gan algorithm. Sensors, 23(7):3440, 2023

  3. [3]

    A new generative adversarial net- work for medical images super resolution.Sci Rep, 12(1):9533, 2022

    Waqar Ahmad, Hazrat Ali, Zubair Shah, and Shoaib Azmat. A new generative adversarial net- work for medical images super resolution.Sci Rep, 12(1):9533, 2022

  4. [4]

    Beware of diffusion models for synthesizing medical images—a comparison with GANs in terms of memorizing brain MRI and chest x-ray images.Mach Learn Sci Technol, 6(1):015022, 2025

    Muhammad Usman Akbar, Wuhao Wang, and Anders Eklund. Beware of diffusion models for synthesizing medical images—a comparison with GANs in terms of memorizing brain MRI and chest x-ray images.Mach Learn Sci Technol, 6(1):015022, 2025

  5. [5]

    How faith- ful is your synthetic data? Sample-level metrics for evaluating and auditing generative models

    Ahmed Alaa, Boris Van Breugel, Evgeny S Saveliev, and Mihaela Van Der Schaar. How faith- ful is your synthetic data? Sample-level metrics for evaluating and auditing generative models. In International Conference on Machine Learning, pages 290–306. PMLR, 2022

  6. [6]

    Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features.Scien- tific data, 4(1):1–13, 2017

    Spyridon Bakas, Hamed Akbari, Aristeidis Soti- ras, Michel Bilello, Martin Rozycki, Justin S Kirby, John B Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features.Scien- tific data, 4(1):1–13, 2017

  7. [7]

    Identifying the best machine learning algo- rithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge, 2019

    Spyridon Bakas, Mauricio Reyes, Andras Jakab, et al. Identifying the best machine learning algo- rithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge, 2019

  8. [8]

    Sutherland, Michael Arbel, and Arthur Gretton

    Miko/suppress laj Bi´ nkowski, Danica J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs, 2021

Show all 81 references
  1. [9]

    Data models, representation, and adequacy-for-purpose.Euro- pean Journal for Philosophy of Science, 11(1):31, 2021

    Alisa Bokulich and Wendy S Parker. Data models, representation, and adequacy-for-purpose.Euro- pean Journal for Philosophy of Science, 11(1):31, 2021

  2. [10]

    Med3d: Transfer learning for 3d medical image analysis, 2019

    Sihong Chen, Kai Ma, and Yefeng Zheng. Med3d: Transfer learning for 3d medical image analysis, 2019

  3. [11]

    Synthesising 3d cardiac cine-mr images and corre- sponding segmentation masks using a latent dif- fusion model

    Nina Cheng, Zhengji Liu, Yash Deo, Haoran Dou, Ning Bi, Kun Wu, Fengming Lin, Zeike A Tay- lor, Nishant Ravikumar, and Alejandro F Frangi. Synthesising 3d cardiac cine-mr images and corre- sponding segmentation masks using a latent dif- fusion model. In2024 IEEE International ...

  4. [12]

    Effectively un- biased FID and Inception score and where to find them

    Min Jin Chong and David Forsyth. Effectively un- biased FID and Inception score and where to find them. In2020 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 6069–6078. IEEE, 2020

  5. [13]

    Review of medical image quality assessment.Biomed Sig- nal Process Control, 27:145–154, 2016

    Li Sze Chow and Raveendran Paramesran. Review of medical image quality assessment.Biomed Sig- nal Process Control, 27:145–154, 2016

  6. [14]

    Correlation between subjective and objective assessment of magnetic resonance (MR) images.Magn Reson Imaging, 34(6):820– 831, 2016

    Li Sze Chow, Heshalini Rajagopal, Raveendran Paramesran, Alzheimer’s Disease Neuroimaging Initiative, et al. Correlation between subjective and objective assessment of magnetic resonance (MR) images.Magn Reson Imaging, 34(6):820– 831, 2016

  7. [15]

    Score-based diffusion models for accelerated MRI.Med Image Anal, 80:102479, 2022

    Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated MRI.Med Image Anal, 80:102479, 2022

  8. [16]

    Automatic tumor segmentation using knowledge-based techniques.IEEE Trans Med Imag, 17(2):187–201, 2006

    Matthew C Clark, Lawrence O Hall, Dmitry B Goldgof, Richard F Murtagh, and Michael L Silbiger. Automatic tumor segmentation using knowledge-based techniques.IEEE Trans Med Imag, 17(2):187–201, 2006

  9. [17]

    Generative adversarial networks to synthe- size missing t1 and flair mri sequences for use in a multisequence brain tumor segmentation model

    Gian Marco Conte, Alexander D Weston, David C Vogelsang, Kenneth A Philbrick, Jason C Cai, Maurizio Barbera, Francesco Sanvito, Daniel H Lachance, Robert B Jenkins, W Oliver Tobin, et al. Generative adversarial networks to synthe- size missing t1 and flair mri sequences for us...

  10. [18]

    Learned local 21 attention maps for synthesising vessel segmenta- tions from T2 MRI

    Yash Deo, Rodrigo Bonazzola, Haoran Dou, Yan Xia, Tianyou Wei, Nishant Ravikumar, Alejan- dro F Frangi, and Toni Lassila. Learned local 21 attention maps for synthesising vessel segmenta- tions from T2 MRI. InInternational Workshop on Simulation and Synthesis in Medical Imagin...

  11. [19]

    Shape-guided conditional latent diffusion models for synthesising brain vasculature

    Yash Deo, Haoran Dou, Nishant Ravikumar, Ale- jandro F Frangi, and Toni Lassila. Shape-guided conditional latent diffusion models for synthesising brain vasculature. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 164–173. Springer, 2023

  12. [20]

    Few-shot learning in diffusion models for generating cerebral aneurysm geome- tries

    Yash Deo, Fengming Lin, Haoran Dou, Nina Cheng, Nishant Ravikumar, Alejandro F Frangi, and Toni Lassila. Few-shot learning in diffusion models for generating cerebral aneurysm geome- tries. In2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2024

  13. [21]

    Dohmen, M.A

    M. Dohmen, M.A. Klemens, I.M. Baltruschat, T. Truong, and M. Lenga. Similarity and quality metrics for MR image-to-image translation.Sci Rep, 15(1):3853, 2025

  14. [22]

    Three-dimensional medical im- age synthesis with denoising diffusion probabilistic models

    Zolnamar Dorjsembe, Sodtavilan Odonchimed, and Furen Xiao. Three-dimensional medical im- age synthesis with denoising diffusion probabilistic models. InMedical Imaging with Deep Learning, 2022

  15. [23]

    Combating mode col- lapse in GAN training: An empirical analysis us- ing hessian eigenvalues, 2020

    Ricard Durall, Avraam Chatzimichailidis, Peter Labus, and Janis Keuper. Combating mode col- lapse in GAN training: An empirical analysis us- ing hessian eigenvalues, 2020

  16. [24]

    From CNNs to GANs for cross-modality medical image estimation.Comput Biol Med, 146:105556, 2022

    Azin Shokraei Fard, David C Reutens, and Viktor Vegh. From CNNs to GANs for cross-modality medical image estimation.Comput Biol Med, 146:105556, 2022

  17. [25]

    Graham, Petru-Daniel Tudosiu, Tom Vercauteren, and M

    Virginia Fernandez, Walter Hugo Lopez Pinaya, Pedro Borges, Mark S. Graham, Petru-Daniel Tudosiu, Tom Vercauteren, and M. Jorge Car- doso. Generating multi-pathological and multi- modal images and labels for brain mri.Medical Image Analysis, 97:103278, 2024

  18. [26]

    Graham, Tom Vercauteren, and M

    Virginia Fernandez, Walter Hugo Lopez Pinaya, Pedro Borges, Petru-Daniel Tudosiu, Mark S. Graham, Tom Vercauteren, and M. Jorge Car- doso.Can Segmentation Models Be Trained with Fully Synthetically Generated Data?, page 79–90. Springer International Publishing, 2022

  19. [27]

    The Vendi score: A diversity evaluation metric for machine learning, 2023

    Dan Friedman and Adji Bousso Dieng. The Vendi score: A diversity evaluation metric for machine learning, 2023

  20. [28]

    Gudbjartsson and S

    H. Gudbjartsson and S. Patz. The Rician dis- tribution of noisy MRI data.Magn Reson Med, 34(6):910–4, December 1995. Erratum in: Magn Reson Med 1996 Aug;36(2):332

  21. [29]

    Adaptive diffu- sion priors for accelerated mri reconstruction.Med Image Anal, 88:102872, 2023

    Alper G¨ ung¨ or, Salman UH Dar, S ¸aban¨Ozt¨ urk, Yilmaz Korkmaz, Hasan A Bedel, Gokberk Elmas, Muzaffer Ozbey, and Tolga C ¸ ukur. Adaptive diffu- sion priors for accelerated mri reconstruction.Med Image Anal, 88:102872, 2023

  22. [30]

    The big argument for ai safety cases

    Ibrahim Habli, Richard Hawkins, Colin Paterson, Philippa Ryan, Yan Jia, Mark Sujan, and John McDermid. The big argument for ai safety cases. arXiv preprint arXiv:2503.11705, 2025

  23. [31]

    GAN-based synthetic brain MR image generation

    Changhee Han, Hideaki Hayashi, Leonardo Rundo, Ryosuke Araki, Wataru Shimoda, Shinichi Muramatsu, Yujiro Furukawa, Giancarlo Mauri, and Hideki Nakayama. GAN-based synthetic brain MR image generation. In2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), p...

  24. [32]

    GAN-based synthetic brain MR image generation

    Changhee Han, Hideaki Hayashi, Leonardo Rundo, Ryosuke Araki, Wataru Shimoda, Shinichi Muramatsu, Yujiro Furukawa, Giancarlo Mauri, and Hideki Nakayama. GAN-based synthetic brain MR image generation. In2018 IEEE 15th international Symposium on Biomedical Imaging (ISBI 2018), p...

  25. [33]

    GANs trained by a two time-scale update rule con- verge to a local Nash equilibrium.Adv Neural Inf Process Syst, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Un- terthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule con- verge to a local Nash equilibrium.Adv Neural Inf Process Syst, 30, 2017

  26. [34]

    Denois- ing diffusion medical models

    Pham Ngoc Huy and Tran Minh Quan. Denois- ing diffusion medical models. In2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2023

  27. [35]

    IXI dataset – brain development.https: //brain-development.org/ixi-dataset/

    Information eXtraction from Images Consortium. IXI dataset – brain development.https: //brain-development.org/ixi-dataset/. Ac- cessed: 2023-02-14

  28. [36]

    Abhinav K Jha, Kyle J Myers, Nancy A Obu- chowski, Ziping Liu, Md Ashequr Rahman, Babak Saboury, Arman Rahmim, Barry A Siegel, Eric C Frey, and Ehsan Samei. Objective task-based 22 evaluation of artificial intelligence-based medical imaging methods: Framework, strategies, and ...

  29. [37]

    Cola-diff: Conditional latent dif- fusion model for multi-modal MRI synthesis

    Lan Jiang, Ye Mao, Xiangfeng Wang, Xi Chen, and Chao Li. Cola-diff: Conditional latent dif- fusion model for multi-modal MRI synthesis. In International Conference on Medical Image Com- puting and Computer-Assisted Intervention, pages 398–408. Springer, 2023

  30. [38]

    COVID-19 CT image synthesis with a conditional generative adversarial network

    Yifan Jiang, Han Chen, Murray Loew, and Hanseok Ko. COVID-19 CT image synthesis with a conditional generative adversarial network. IEEE J Biomed Health Inform, 25(2):441–452, 2020

  31. [39]

    Feature likelihood divergence: Evalu- ating the generalization of generative models using samples, 2024

    Marco Jiralerspong, Avishek Joey Bose, Ian Gemp, Chongli Qin, Yoram Bachrach, and Gau- thier Gidel. Feature likelihood divergence: Evalu- ating the generalization of generative models using samples, 2024

  32. [40]

    Image quality assessment for magnetic resonance imaging.IEEE Access, 11:14154–14168, 2023

    Sergey Kastryulin, Jamil Zakirov, Nicola Pezzotti, and Dmitry V Dylov. Image quality assessment for magnetic resonance imaging.IEEE Access, 11:14154–14168, 2023

  33. [41]

    Diffusion models in medical imaging: A comprehensive sur- vey.Med Image Anal, 88:102846, 2023

    Amirhossein Kazerouni, Ehsan Khodapanah Agh- dam, Moein Heidari, Reza Azad, Mohsen Fayyaz, Ilker Hacihaliloglu, and Dorit Merhof. Diffusion models in medical imaging: A comprehensive sur- vey.Med Image Anal, 88:102846, 2023

  34. [42]

    Deep learning approaches for data aug- mentation in medical imaging: a review.J Imag- ing, 9(4):81, 2023

    Aghiles Kebaili, J´ erˆ ome Lapuyade-Lahorgue, and Su Ruan. Deep learning approaches for data aug- mentation in medical imaging: a review.J Imag- ing, 9(4):81, 2023

  35. [43]

    Denoising diffusion probabilistic mod- els for 3D medical image generation.Sci Rep, 13(1):7303, 2023

    Firas Khader, Gustav M¨ uller-Franzes, Soroosh Tayebi Arasteh, Tianyu Han, Christoph Haar- burger, Maximilian Schulze-Hagen, Philipp Schad, Sandy Engelhardt, Bettina Baeßler, Sebastian Fo- ersch, et al. Denoising diffusion probabilistic mod- els for 3D medical image generation...

  36. [44]

    Kingma and M

    D.P. Kingma and M. Welling. Auto-encoding vari- ational Bayes.arXiv.org, abs/1312.6114, 2013

  37. [45]

    Improved precision and recall metric for assessing generative models.Adv Neural Inf Process Syst, 32, 2019

    Tuomas Kynk¨ a¨ anniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models.Adv Neural Inf Process Syst, 32, 2019

  38. [46]

    Improved precision and recall metric for assessing generative models, 2019

    Tuomas Kynk¨ a¨ anniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models, 2019

  39. [47]

    What distinguishes data from models?European Journal for Philosophy of Sci- ence, 9(2):22, 2019

    Sabina Leonelli. What distinguishes data from models?European Journal for Philosophy of Sci- ence, 9(2):22, 2019

  40. [48]

    Comparative study of the methodolo- gies used for subjective medical image quality as- sessment.Phys Med Biol, 66(15):15TR02, 2021

    Lucie L´ evˆ eque, Meriem Outtas, Hantao Liu, and Lu Zhang. Comparative study of the methodolo- gies used for subjective medical image quality as- sessment.Phys Med Biol, 66(15):15TR02, 2021

  41. [49]

    Chaofeng Li and Alan C. Bovik. Content- partitioned structural similarity index for image quality assessment.Signal Processing: Image Communication, 25(7):517–526, 2010. Special Is- sue on Image and Video Quality Assessment

  42. [50]

    Fddm: Unsupervised medical image translation with a frequency-decoupled diffusion model, 2024

    Yunxiang Li, Hua-Chieh Shao, Xiaoxue Qian, and You Zhang. Fddm: Unsupervised medical image translation with a frequency-decoupled diffusion model, 2024

  43. [51]

    High- throughput 3DRA segmentation of brain vascula- ture and aneurysms using deep learning.Comput Methods Programs Biomed, 230:107355, 2023

    Fengming Lin, Yan Xia, Shuang Song, Nishant Ravikumar, and Alejandro F Frangi. High- throughput 3DRA segmentation of brain vascula- ture and aneurysms using deep learning.Comput Methods Programs Biomed, 230:107355, 2023

  44. [52]

    Preserving privacy in healthcare: A system- atic review of deep learning approaches for syn- thetic data generation.Comput Methods Programs Biomed, page 108571, 2024

    Yintong Liu, U Rajendra Acharya, and Jen Hong Tan. Preserving privacy in healthcare: A system- atic review of deep learning approaches for syn- thetic data generation.Comput Methods Programs Biomed, page 108571, 2024

  45. [53]

    Comparison of objective image quality metrics to expert radiolo- gists’ scoring of diagnostic quality of MR images

    Allister Mason, James Rioux, Sharon E Clarke, Andreu Costa, Matthias Schmidt, Valerie Keough, Thien Huynh, and Steven Beyea. Comparison of objective image quality metrics to expert radiolo- gists’ scoring of diagnostic quality of MR images. IEEE Trans Med Imag, 39(4):1064–1072, 2019

  46. [54]

    A non-parametric test to detect data- copying in generative models, 2020

    Casey Meehan, Kamalika Chaudhuri, and Sanjoy Dasgupta. A non-parametric test to detect data- copying in generative models, 2020

  47. [55]

    The mul- timodal brain tumor image segmentation bench- mark (brats).IEEE transactions on medical imag- ing, 34(10):1993–2024, 2014

    Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Jo- hannes Slotboom, Roland Wiest, et al. The mul- timodal brain tumor image segmentation bench- mark (brats).IEEE transactions on medical imag- in...

  48. [56]

    Fast approx- imation of the sliced-Wasserstein distance using concentration of random projections.Adv Neural Inf Process Syst, 34:12411–12424, 2021

    Kimia Nadjahi, Alain Durmus, Pierre E Jacob, Roland Badeau, and Umut Simsekli. Fast approx- imation of the sliced-Wasserstein distance using concentration of random projections.Adv Neural Inf Process Syst, 34:12411–12424, 2021

  49. [57]

    Re- liable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Re- liable fidelity and diversity metrics for generative models. InInternational Conference on Machine Learning, pages 7176–7185. PMLR, 2020

  50. [58]

    Generating images with sparse representations

    Charlie Nash, Jacob Menick, Sander Dieleman, and Peter Battaglia. Generating images with sparse representations. InInternational Con- ference on Machine Learning, pages 7958–7968. PMLR, 2021

  51. [59]

    medigan: a python li- brary of pretrained generative models for medi- cal image synthesis.Journal of Medical Imaging, 10(6):061403, 2023

    Richard Osuala, Grzegorz Skorupko, Noussair Lazrak, Lidia Garrucho, Eloy Garc´ ıa, Smriti Joshi, Socayna Jouide, Michael Rutherford, Fred Prior, Kaisar Kushibar, et al. medigan: a python li- brary of pretrained generative models for medi- cal image synthesis.Journal of Medical...

  52. [60]

    Unsupervised medical image trans- lation with adversarial diffusion models.IEEE Transactions on Medical Imaging, 42(12):3524– 3539, 2023

    Muzaffer ¨Ozbey, Onat Dalmaz, Salman UH Dar, Hasan A Bedel, S ¸aban¨Ozturk, Alper G¨ ung¨ or, and Tolga C ¸ ukur. Unsupervised medical image trans- lation with adversarial diffusion models.IEEE Transactions on Medical Imaging, 42(12):3524– 3539, 2023

  53. [61]

    Pak, Yuyin Yong, Hakjoo Lee, Letian Sun, Semir Beraki, Linda Wang, Roger Chen, Mihela van der Schaar, and Ahmed M Alaa

    Neha S. Pak, Yuyin Yong, Hakjoo Lee, Letian Sun, Semir Beraki, Linda Wang, Roger Chen, Mihela van der Schaar, and Ahmed M Alaa. Aligning synthetic medical images with clinical knowledge using human feedback. InAdvances in Neural In- formation Processing Systems, volume 36

  54. [62]

    Virtually a measurement.Nature Physics, 17(2):146, 2021

    Wendy S Parker. Virtually a measurement.Nature Physics, 17(2):146, 2021

  55. [63]

    Evidence and knowledge from computer simulation.Erkenntnis, 87(4):1521– 1538, 2022

    Wendy S Parker. Evidence and knowledge from computer simulation.Erkenntnis, 87(4):1521– 1538, 2022

  56. [64]

    Model evaluation

    Wendy S Parker. Model evaluation. InThe Rout- ledge Handbook of Philosophy of Scientific Model- ing, pages 208–219. Routledge, 2024

  57. [65]

    Towards performant and reliable undersampled mr recon- struction via diffusion model sampling

    Cheng Peng, Pengfei Guo, S Kevin Zhou, Vishal M Patel, and Rama Chellappa. Towards performant and reliable undersampled mr recon- struction via diffusion model sampling. InInter- national Conference on Medical Image Computing and Computer-Assisted Intervention, pages 623–

  58. [66]

    Walter H. L. Pinaya, Petru-Daniel Tudosiu, Jes- sica Dafflon, Pedro F da Costa, Virginia Fernan- dez, Parashkev Nachev, Sebastien Ourselin, and M. Jorge Cardoso. Brain imaging generation with latent diffusion models, 2022

  59. [67]

    Brain imaging generation with latent diffusion models

    Walter HL Pinaya, Petru-Daniel Tudosiu, Jes- sica Dafflon, Pedro F Da Costa, Virginia Fernan- dez, Parashkev Nachev, Sebastien Ourselin, and M Jorge Cardoso. Brain imaging generation with latent diffusion models. InMICCAI Workshop on Deep Generative Models, pages 117–126. Spri...

  60. [68]

    Controllable medical image generation via gener- ative adversarial networks

    Zhihang Ren, X Yu Stella, and David Whitney. Controllable medical image generation via gener- ative adversarial networks. InIS&T International Symposium on Electronic Imaging, volume 33. NIH Public Access, 2021

  61. [69]

    The epistemic superiority of ex- periment to simulation.Synthese, 195(11):4883– 4906, 2018

    Sherrilyn Roush. The epistemic superiority of ex- periment to simulation.Synthese, 195(11):4883– 4906, 2018

  62. [70]

    Improved techniques for training GANs

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. Adv Neural Inf Process Syst, 29, 2016

  63. [71]

    Towards ground-truth-free evaluation of any seg- mentation in medical images, 2024

    Ahjol Senbi, Tianyu Huang, Fei Lyu, Qing Li, Yuhui Tao, Wei Shao, Qiang Chen, Chengyan Wang, Shuo Wang, Tao Zhou, and Yizhe Zhang. Towards ground-truth-free evaluation of any seg- mentation in medical images, 2024

  64. [72]

    Medi- cal image generation using generative adversarial networks: A review.Health informatics: A com- putational perspective in healthcare, pages 77–96, 2021

    Nripendra Kumar Singh and Khalid Raza. Medi- cal image generation using generative adversarial networks: A review.Health informatics: A com- putational perspective in healthcare, pages 77–96, 2021

  65. [73]

    Solving inverse problems in medical imag- ing with score-based generative models.arXiv preprint arXiv:2111.08005, 2021

    Yang Song, Liyue Shen, Lei Xing, and Stefano Er- mon. Solving inverse problems in medical imag- ing with score-based generative models.arXiv preprint arXiv:2111.08005, 2021

  66. [74]

    Exposing flaws of gen- erative model evaluation metrics and their unfair treatment of diffusion models.Adv Neural Inf Pro- cess Syst, 36, 2024

    George Stein, Jesse Cresswell, Rasa Hossein- zadeh, Yi Sui, Brendan Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of gen- erative model evaluation metrics and their unfair treatment of diffusion models.Adv Neu...

  67. [75]

    Neural discrete representation learning.Adv Neural Inf Process Syst, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.Adv Neural Inf Process Syst, 30, 2017

  68. [76]

    Wang, E.P

    Z. Wang, E.P. Simoncelli, and A.C. Bovik. Mul- tiscale structural similarity for image quality as- sessment. InThe Thirty-Seventh Asilomar Con- ference on Signals, Systems & Computers, 2003, volume 2, pages 1398–1402 Vol.2, 2003

  69. [77]

    Krishna Pandu Wicaksono, Koji Fujimoto, Ya- sutaka Fushimi, Akihiko Sakata, Sachi Okuchi, Takuya Hinoda, Satoshi Nakajima, Yukihiro Ya- mao, Kazumichi Yoshida, Kanae Kawai Miyake, et al. Super-resolution application of generative adversarial network on brain time-of-flight mr ...

  70. [78]

    Yung, Ankit B

    McKell Woodland, Austin Castelo, Mais Al Taie, Jessica Albuquerque Marques Silva, Mohamed Eltaher, Frank Mohn, Alexander Shieh, Supra- teek Kundu, Joshua P. Yung, Ankit B. Patel, and Kristy K. Brock. Feature extraction for generative medical imaging evaluation: New evidence ag...

  71. [79]

    A novel hybrid generative adver- sarial network for ct and mri super-resolution re- construction.Physics in Medicine & Biology, 68(13):135007, 2023

    Yueyue Xiao, Chunxiao Chen, Liang Wang, Jie Yu, Xue Fu, Yuan Zou, Zhe Lin, and Kun- peng Wang. A novel hybrid generative adver- sarial network for ct and mri super-resolution re- construction.Physics in Medicine & Biology, 68(13):135007, 2023

  72. [80]

    On mode collapse in genera- tive adversarial networks

    Kaifeng Zhang. On mode collapse in genera- tive adversarial networks. InArtificial Neural Networks and Machine Learning–ICANN 2021: 30th International Conference on Artificial Neu- ral Networks, Bratislava, Slovakia, September 14– 17, 2021, Proceedings, Part II 30, pages 563–5...

  73. [81]

    The unreasonable effectiveness of deep features as a perceptual met- ric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual met- ric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018. 25

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.