Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers

Bart Elen; Chang Sun; Gokhan Ertaylan; Mahmoud Ibrahim; Michel Dumontier

REVIEW 4 major objections 4 minor 23 references

Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Synthetic chest CT images conditioned on sex and age can both correct and reveal subgroup bias in disease classifiers, matching full-real-data performance from 1% of real labels.

desk verdict A useful question and a thorough experimental sweep, but the generator was fine-tuned on the same hold-out that later becomes the audit oracle, so both headline results are likely leakage artifacts. read the letter →

arxiv 2607.14984 v1 pith:7IU64LYA submitted 2026-07-16 cs.AI

Mahmoud Ibrahim , Bart Elen , Chang Sun , Gokhan Ertaylan , Michel Dumontier This is my paper · ORCID

classification cs.AI

keywords syntheticmedicaldatafairnessauditsubpopulationevaluationdiffusionmodelsCOVID-19chestCTbiasmitigationdetectionefficiency

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 argues that a generative model conditioned on demographic attributes (sex, age) and disease status can be used on both sides of the subgroup-bias problem in medical image classification. On the training side, using a demographically balanced synthetic cohort as a pretraining prior and then fine-tuning on a tiny slice of real data produces a classifier that beats full-real-data training on both average and worst-subgroup performance, at roughly 100x real-data efficiency. On the evaluation side, a large synthetic minority cohort reproduces the per-subgroup ranking of a well-powered real test set exactly, and gives tighter estimates in exactly the minority cells where the real test lacks samples. The two uses are shown to share a single root: the generator can populate demographic cells that the real data under-represents. A reader should care because it suggests a practical recipe for making medical-image classifiers fairer and for auditing fairness when real minority labels are scarce.

What carries the argument

The load-bearing object is a fine-tuned Stable Diffusion 2.1 latent diffusion generator, conditioned by natural-language prompts of the form '{age} year old {SEX}. {COVID-19 findings on / Normal} chest CT scan' through a CLIP text encoder. It is fine-tuned end-to-end on a demographically mixed set of 62,829 real slices, and from it five synthetic cohorts are sampled, including a uniform 40,000-slice balanced cohort (2,000 per sex-age-label cell) and size-matched mirrors of the real test partitions. This generator is the shared substrate for both contributions: as a pretraining prior it pre-loads uniform per-cell exposure that a small real fine-tune preserves, and as a test proxy it provides

What would settle it

Re-train the generator on a strictly disjoint training set (e.g., only the train/val partitions, excluding every patient in the oracle test set) and repeat the audit; if Spearman rho between the synthetic proxy and oracle drops substantially from 1.00, the perfect rank agreement is an artifact of training-data overlap. Alternatively, measure nearest-neighbor feature distances between synthetic test slices and oracle real slices; near-duplicate matches would support memorization.

Watch

Extended reading notes

Core claim

Fine-tuning a latent diffusion model end-to-end on COVID-19 chest CT slices with text prompts specifying sex, age, and COVID status yields a generator whose synthetic cohorts can stand in for real data in two distinct roles. Used sequentially as a pretraining prior followed by a small real fine-tune, a uniform (demographically balanced) synthetic cohort yields a DenseNet-121 classifier reaching mean MCC 0.83 and worst-cell MCC 0.52 from only 585 real slices, surpassing both full-real-data baselines (0.78/0.17 and 0.73/0.20); the same data combined as joint augmentation gives only 0.70/0.16, isolating the training schedule as the source of the fairness lift. Used as a synthetic test set at th

Load-bearing premise

The central premise is that the synthetic cohort is generated by a model whose training data is independent of the real test sets used as the oracle; the paper violates this because its generator is fine-tuned on the very partition that is later pooled and re-split into the oracle test set, so the perfect rank agreement could reflect memorization rather than a general method.

Editorial extensions

If this is right

A hospital with only a few hundred labeled minority slices can train a classifier that outperforms one trained on a large, demographically biased real dataset, if a balanced-synthetic pretraining prior is available.
Fairness audits can use a synthetic minority cohort to rank subgroups reliably when real test sets have too few minority samples; the crossover at ~1500 real slices per cell indicates when to prefer each estimator.
The schedule of data use matters more than the data composition: sequential pretraining-plus-fine-tuning should replace joint augmentation when synthetic data is used for bias mitigation.
Threshold-sensitive metrics (F1, MCC) on synthetic test sets must be evaluated at a threshold tuned on real validation data, or the apparent drift is inflated by the operating point.
The two contributions are not independent: the same generator property (populating under-represented cells) drives both mitigation and detection, so improvements in conditioning accuracy should improve both sides.

Reading between the lines

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

If the observed effect is robust to the leakage concern, the same two-stage recipe could generalize to other medical imaging modalities where demographic subgroups are imbalanced, such as chest X-ray or retinal imaging, provided the generator's conditioning signal is at least as accurate as here.
The paper's size-based crossover implies a practical audit protocol: use the synthetic cohort to flag problem subgroups and the real test to confirm magnitudes once enough real samples exist; a decision rule of this kind could be automated and calibrated per generator.
A testable extension: compute per-subgroup FID or nearest-neighbor distances between synthetic and real test cells; if synthetic cells that are far from real cells show worse rank agreement, then generator fidelity per cell, not aggregate FID, will be the limiting factor for audit validity.
The threshold protocol suggests that synthetic test sets may be better used for ranking and relative comparisons than for absolute performance numbers, unless a deployment-realistic threshold is fixed in advance; this distinction should be made explicit in regulatory guidance.

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

Two things you should know. First, the paper tests a genuinely useful question: whether demographically balanced synthetic data works better as a pretraining prior or as joint augmentation, and whether a synthetic cohort can stand in for a real minority test set in a subgroup fairness audit. The empirical effort is substantial: 22 training regimes, five classifier seeds, and released code and model weights. Second, the headline numbers are almost certainly inflated by a data-split error. The generator is fine-tuned on the hold-out partition (Section 3.2), and that same partition is later pooled with hold-out-minority and re-split 80/20 to create the oracle real-test-large (Sections 3.1, 4.4). So roughly 70% of the oracle's slices come from the generator's training distribution. The synthetic proxy is sampled from that same generator. Perfect rank agreement with the oracle, and synthetic pretraining that beats the full-real baseline at 1% of the real data, are exactly what you would expect from a model that had already seen the test distribution. This is not a minor leak; it is the load-bearing wall under both headline claims.

What the paper does well. The schedule comparison — sequential pretraining versus joint augmentation at fixed data — is a clean counterfactual, and the 3.3x worst-cell MCC advantage is striking, though it still relies on the contaminated generator. The threshold analysis in Section 4.4 is a careful piece of measurement, separating an operating-point cliff from genuine synthetic drift; that part could be reused. The size-gated complementarity (synthetic estimates more accurate in under-sampled cells, real estimates better in well-sampled cells) is a plausible and interesting observation, but its quantitative support inherits the same split problem.

Soft spots. Beyond the circularity, the age-conditioning accuracy is 0.33 versus 0.20 chance — weak for the demographic axis that most needs fixing. The FID numbers are computed against data the generator trained on, which flatters them. The paper is candid that privacy-side analyses are missing, but does not supply them. All of this is fixable in principle: retrain the generator on train only and regenerate every result. But as submitted, the central claims are not supported.

Who this is for. Anyone working on synthetic data for medical fairness will want to know this paper exists, mostly as a warning about evaluation design. The idea has legs; the current numbers do not.

Recommendation. Send it to peer review rather than desk-reject: the error is identifiable and a competent referee should catch it, and the underlying question is important enough to justify a careful re-run with a clean split. Just be prepared that the revision will require redoing every experiment.

Referee Report

4 major / 4 minor

Summary. The paper proposes using a demographically-conditioned Stable Diffusion 2.1 generator, fine-tuned on COVIDx-CT-3A chest CT slices, for two purposes: (i) bias mitigation, by pretraining a DenseNet-121 classifier on balanced synthetic slices and then fine-tuning on a tiny real set, and (ii) bias detection, by using a large synthetic minority cohort as a proxy for a well-powered real test set in subgroup fairness audits. The central claims are that FT-Balanced@1% real data (585 slices) surpasses full-real-data baselines on both mean and worst-cell MCC, and that the synthetic test cohort achieves perfect Spearman rank agreement (ρ=1.00) with the real oracle on MCC and Recall, while also giving more reliable per-cell estimates in under-sampled minority cells.

Significance. If valid, the results would be practically important: a released fine-tuned generator, open code and weights, a paired comparison isolating training schedule from data composition, a size-gated crossover between synthetic and real evaluation, and a threshold protocol for synthetic test evaluation. The 100× data-efficiency claim and the perfect rank agreement claim are the load-bearing results. However, both are threatened by a structural data-split circularity: the generator is fine-tuned on the same hold-out partition that, after pooling and re-splitting, forms the oracle against which the synthetic proxy is judged. This makes the headline numbers potentially artifacts of leakage rather than evidence of a general method.

major comments (4)

[§3.1–3.2, Table 1] The evaluation protocol is circular. The generator is fine-tuned end-to-end on the hold-out partition (62,829 slices, §3.2). The audit oracle real-test-large is then constructed by pooling hold-out with hold-out-minority and re-splitting 80/20 at the patient level (§3.1, §4.4), so roughly 70% of real-test-large's slices come from the generator's training set. Synth-Test-Large is sampled from that same generator. The perfect rank agreement (ρ=1.00) and the low per-cell synthetic bias in §4.5 therefore reflect the generator having seen most of the oracle distribution, not an independent ability to reproduce subgroup rankings. This is a load-bearing flaw for the bias-detection contribution.
[§4.3, §3.2] The bias-mitigation claim is contaminated by the same leakage. FT-Balanced@1% pretrains on Synth-Balanced, which is sampled from the generator fine-tuned on hold-out, and the resulting classifier is evaluated on hold-out (the real test set). The synthetic pretraining distribution thus includes information about the evaluation distribution. Beating the full-real TRTR baselines could stem from the generator having memorized test-distribution content rather than from balanced-synthetic pretraining as a general strategy. A proper experiment would require training the generator on a partition entirely disjoint from the evaluation set, or at least a memorization check (e.g., nearest-neighbour distance between synthetic slices and real test slices), which the paper explicitly lists as missing in §5.
[§3.2, Table 3] Age conditioning accuracy is only 0.33 against a chance level of 0.20 for the five-class age problem. This is barely above chance and undermines the claim of a 'demographically-conditioned' generator, particularly for the age-axis subgroup results in §4.5. The paper does not report per-age confusion matrices or conditioning accuracy on held-out real data, so it is unclear whether the synthetic cohorts actually provide meaningful age-specific variation. Given that the audit's subgroups are defined jointly by sex and age, this weakens the interpretability of the claimed perfect rank correlation across age cells.
[§4.5, Figure 5B] The rank-agreement result is reported as ρ=1.00 with 95% bootstrap CIs [0.93,1.00] and [0.86,1.00], but the figure caption states this is over '7 evaluable subgroups'. With n=7, a perfect Spearman correlation is far less impressive than the abstract implies, and the reported bootstrap CIs appear too tight for such a small sample. Please report the actual number of subgroups used for each ρ, the full ranking, and the effect of excluding any subgroup. This is a presentation issue, but it directly affects the strength of the headline claim.

minor comments (4)

[Abstract / full text] There is a typo in the full-text abstract: 'F or bias mitigation' should read 'For bias mitigation'.
[§4.5, Figure 5 caption] The caption says '7evaluable subgroups' without a space, and the text mentions 'ten joint sex–age cells'. Clarify how many subgroups are used in each analysis and why some are excluded.
[§4.4, Table 4] The table reports 95% CIs for each metric, but the caption says 'All 95% CIs exclude zero.' Please specify whether these are bootstrap CIs or over-seed CIs, and how many bootstrap resamples were used.
[§5, Limitations] The limitations section mentions that privacy-side analyses would strengthen the claims, but does not discuss the more serious data-split issue. Please add a discussion of generator–evaluation disjointness as a required validation step for synthetic test-set proxies.

Circularity Check

2 steps flagged · score 8.0 of 10

Generator fine-tuned on the audit oracle makes both headline results non-independent: synthetic-oracle agreement and 100× efficiency are partly forced by construction.

fitted input called prediction [§3.1, §3.2, §4.5]
"hold-out (62,829 slices) matches the joint (sex, age, label) distribution of train/val and is the held-back real reference; the name marks its dual role as the generator’s training data (Section 3.2). ... The per-subgroup audit (Section 4.4) pools hold-out with hold-out-minority and re-splits the union 80/20 at the patient level ... yielding the well-powered minority oracle real-test-large (71,037 slices) ... Synth-Test-Large mirrors the well-powered real-test-large oracle and is the audit proxy in Section 4.4."

The generator is fine-tuned end-to-end on the hold-out partition, which is then pooled with hold-out-minority and re-split to form the oracle real-test-large. The synthetic proxy synth-test-large is sampled from that same generator. Therefore the claimed perfect rank agreement between synth-test-large and the oracle (Spearman ρ = 1.00 on MCC and Recall) is not a prediction against an unseen distribution; it is a resampling of a model trained on most of the oracle's slices. The paper's own label 'dual role' acknowledges that hold-out is both generator training data and held-back reference, making the evaluation circular by construction.
fitted input called prediction [§3.1, §3.2, §4.3]
"Aggregate experiments (Sections 4.1–4.3) use the standard three-way train/val/hold-out split. ... We fine-tune the U-Net denoiser and the CLIP text encoder [8] end-to-end on the hold-out partition for 20,000 optimization steps."

The bias-mitigation headline (FT-Balanced@1% reaching mean MCC 0.83 and worst-cell 0.52, surpassing full-real baselines) is evaluated on the hold-out real test set. But the same hold-out set was used to train the generator that produces Synth-Balanced, the pretraining data for FT-Balanced. Thus the classifier's apparent 100× real-data efficiency may reflect leakage of test-set information through the synthetic pretraining cohort, not a general property of balanced-synthetic pretraining. The evaluation and the generator's training data are the same partition.

full rationale

The dominant circularity is a data-split issue, not a self-citation or ansatz issue. The paper explicitly defines hold-out as having a 'dual role' (generator training data and held-back reference), then merges hold-out into the audit oracle. Because the synthetic proxy is sampled from the generator fine-tuned on that oracle's constituent data, the perfect rank agreement and the synthetic pretraining gains are at least partly forced by construction. The paper itself concedes that privacy-side analyses (nearest-neighbour, membership inference) 'would strengthen claims about synthetic-data deployment,' indicating no memorization check was performed. The paired comparison between AUG and FT at fixed synthetic data retains some independent content (schedule effect), but the two main claims as stated—synthetic cohort as a valid auditor and 100× efficiency over full-real training—are contaminated by this structural overlap. No self-citation circularity or definitional renaming was found. Score 8 reflects central, partial circularity of both headline results.

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

The main burden is the assumed independence between the generator's training data and the evaluation oracle; the paper's split design violates it. The generator's own fidelity (especially age conditioning) and the slice-level sampling are additional domain assumptions that are not independently validated.

free parameters (3)

size-gated evaluation crossover threshold = 1500 real-test slices
In §4.5/Fig. 6, the claim that synthetic estimates are more reliable in under-sampled cells depends on a threshold at 1500 real-test slices; this threshold is read off the same experimental data, not pre-specified.
balanced synthetic cohort size = 40,000 slices (2,000 per cell)
All training-side regimes use Synth-Balanced fixed at 40,000 slices; the headline 100× efficiency and the schedule comparison depend on this budget (§3.3, §4.2–4.3).
real-val decision threshold T* = 0.24
Used in the threshold-robustness analysis to shrink the synthetic-real gap; tuned on real-val, so not a prediction (§4.4, Fig. 4).

assumptions (4)

domain assumption Diffusion models can synthesize usable chest CT images after fine-tuning on COVIDx-CT-3A
The entire approach assumes the SD 2.1 fine-tune produces realistic, label/demographic-faithful CT slices; Table 3 age conditioning accuracy (0.33 vs 0.20 chance) weakens this for age.
ad hoc to paper hold-out can be used both as generator training data and as part of the evaluation oracle without leakage
Load-bearing and violated: hold-out (62,829 slices) is the generator's fine-tuning set and is pooled/re-split into real-test-large and real-test-small (§3.1–3.2), so the synthetic-vs-real comparisons are not independent.
domain assumption Slice-level sampling from the generator preserves the intended per-cell subgroup distribution
Synthetic cohorts are sampled per-slice, not per-patient, and the generator sees only prompts; whether this yields faithful per-(sex, age) cell estimates is assumed (§3.3, §4.5).
domain assumption COVIDx-CT-3A sex/age labels and COVID status are correct for patient-disjoint stratification
All fairness cells and generator conditioning rely on the dataset's demographic labels.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers." pith.science (2026). https://pith.science/paper/7IU64LYA

@misc{pith2026260714984,
  author       = {Pith},
  title        = {Pith review of: Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IU64LYA}},
  note         = {Machine review of arXiv:2607.14984}
}

abstract

Per-subgroup fairness audits of medical image classifiers face a sample-size problem: minority subgroups in held-out test sets have so few samples that the resulting confidence intervals on per-subgroup performance are wider than the bias the audit is meant to detect. We argue that a demographically-conditioned synthetic generator can do both: mitigate bias on the training side and detect bias on the evaluation side. Working on COVID-19 chest CT classification with an end-to-end fine-tuned Stable Diffusion 2.1 generator, we make two findings. For bias mitigation (training), a demographically-balanced synthetic cohort is most useful as a pretraining prior, not as joint augmentation: with the same fixed data, sequential pretraining followed by fine-tuning substantially outperforms joint augmentation, and the resulting classifier surpasses the full-real baseline at $\sim$$100\times$ real-data efficiency. For bias detection (evaluation), across five synthetic minority cohorts and five classifier seeds, the synthetic estimator reproduces the subgroup ranking of a well-powered real oracle (Spearman $\rho = 1.00$ on MCC and Recall) and gives the more reliable per-cell estimate where the small real test set runs out of samples. The synthetic cohort is therefore most useful in exactly the cells that fairness audits care about, as both a fix for and a measure of subgroup bias.

Figures

Figures reproduced from arXiv: 2607.14984 by the authors.

**Figure 1.** Data-provenance map. The four COVIDx-CT-3A patientdisjoint partitions, the two derived per-subgroup evaluation sets (real-test-large, real-test-small) obtained by merging and re-splitting hold-out∪ hold-out-minority, the SD 2.1 fine-tuned generator, and the four synthetic cohorts sampled from it. Exact counts and the downstream use each set feeds are given in [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

**Figure 2.** Augmentation lift on Real-Test (5-seed CIs). Mean (left) and worstcell (right) MCC across 10 subgroups, as a function of real training-set size N. AUG curves: mean ± 95 % CI over five classifier seeds. Real-only baselines (TRTR-Biased 0.784 / 0.172, TRTR-Balanced 0.728 / 0.204) shown as horizontals. AUG @ N ≥40,000 exceeds both baselines on both metrics. 4.3 Bias mitigation II: Two-stage synthetic pretraining: 100… view at source ↗

**Figure 3.** Two-stage synthetic pretraining + real fine-tuning. Mean (left) and worst-cell (right) MCC on Real-Test, as a function of real fine-tuning data size (log scale). FT-Balanced and FT-Matched curves: mean ± 95 % CI over five seeds. The circled point marks FT-Balanced at 1 % real (585 slices), which clears both baselines on both metrics (100× real-data efficiency); FT-Matched stays roughly 0.12 MCC below FT-Balanced acr… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Threshold sweep on aggregate F1 and MCC. Macro-F1 (left) and MCC (right) as a function of the decision threshold T on p(covid), for the TRTR-Biased classifier on real-test (blue) and synth-test (green). Tuning T on real-val (T ≈ 0.24) shrinks the apparent [PITH_FULL_I…

**Figure 5.** Figure 5: Subgroup audit consistency. (A) Per-subgroup F1 (top) and MCC (bottom) for the Real-Test-Large oracle (gold) vs Synth-Test-Large proxy (navy, mean over 5 cohorts ×5 seeds); per-panel Pearson r over the cells shown. (B) Spearman ρ between each proxy’s per-subgroup ranki…

**Figure 6.** Figure 6: Calibration error binned by real-test cell size. Mean per-cell |mean(proxy) − mean(oracle)| on MCC (left) and F1 (right), split at 1500 realtest slices. The order flips between the two proxies: SYN is ∼2.3× less biased than Test on under-sampled cells, while Test is ∼…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 7 linked inside Pith

[1]

In: Pacific Symposium on Biocomputing 2021

Seyyed-Kalantari, L., Liu, G., McDermott, M., Chen, I.Y., Ghassemi, M.: CheX- clusion: Fairness gaps in deep chest X-ray classifiers. In: Pacific Symposium on Biocomputing 2021. pp. 232–243 (2020)

2021
[2]

Nature Medicine27(12), 2176–2182 (2021)

Seyyed-Kalantari, L., Zhang, H., McDermott, M.B.A., Chen, I.Y., Ghassemi, M.: Underdiagnosisbiasofartificialintelligencealgorithmsappliedtochestradiographs in under-served patient populations. Nature Medicine27(12), 2176–2182 (2021)

2021
[3]

arXiv preprint arXiv:2312.10083 (2023)

Yang, Y., Zhang, H., Gichoya, J.W., Katabi, D., Ghassemi, M.: The limits of fair medical imaging AI in the wild. arXiv preprint arXiv:2312.10083 (2023)

arXiv 2023
[4]

In: AMIA Summits on Translational Science Proceedings

Lin, M., Xiao, Y., Hou, B., et al.: Evaluate underdiagnosis and overdiagnosis bias of deep learning model on primary open-angle glaucoma diagnosis in under-served populations. In: AMIA Summits on Translational Science Proceedings. p. 370 (2023)

2023
[5]

In: CVPR

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684–10695 (2022)

2022
[6]

In: NeurIPS

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840–6851 (2020)

2020
[7]

arXiv preprint arXiv:2207.12598 (2022)

Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

arXiv 2022
[8]

arXiv preprint arXiv:2103.00020 (2021)

Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020 (2021)

arXiv 2021

Show all 23 references

[9]

arXiv preprint arXiv:2211.12737 (2022)

Chambon, P., Bluethgen, C., Delbrouck, J.B., et al.: RoentGen: Vision-language foundation model for chest X-ray generation. arXiv preprint arXiv:2211.12737 (2022)

2022 arXiv
[10]

In: ISBI

Packhäuser, K., Folle, L., Thamm, F., Maier, A.: Generation of anonymous chest radiographs using latent diffusion models for training thoracic abnormality classi- fication systems. In: ISBI. pp. 1–5 (2023)

2023
[11]

arXiv preprint arXiv:2405.09806 (2024)

Cho, J., Mathur, R., Zakka, C., et al.: MediSyn: A generalist text-guided latent dif- fusion model for diverse medical image synthesis. arXiv preprint arXiv:2405.09806 (2024)

2024 arXiv
[12]

Jiang, Y., Chen, H., Loew, M., Ko, H.: COVID-19 CT image synthesis with a conditionalgenerativeadversarialnetwork.IEEEJournalofBiomedicalandHealth Informatics25(2), 441–452 (2020)

2020
[13]

arXiv preprint arXiv:1706.02633 (2017)

Esteban, C., Hyland, S.L., Rätsch, G.: Real-valued (medical) time series generation with recurrent conditional GANs. arXiv preprint arXiv:1706.02633 (2017)

2017 arXiv
[14]

Radiology312(3), e232471 (2024)

Koetzier, L.R., Wu, J., Mastrodicasa, D., et al.: Generating synthetic data for medical imaging. Radiology312(3), e232471 (2024)

2024
[15]

In: NeurIPS

van Breugel, B., Seedat, N., Imrie, F., van der Schaar, M.: Can you rely on your model evaluation? Improving model evaluation with synthetic test data. In: NeurIPS. vol. 36, pp. 1889–1904 (2023)

1904
[16]

OpenReview preprint (2023) Demographically-Conditioned Synthetic Chest CT 17

van Breugel, B., Seedat, N., Imrie, F., van der Schaar, M.: SYNG4ME: Model evaluation using synthetic test data. OpenReview preprint (2023) Demographically-Conditioned Synthetic Chest CT 17

2023
[17]

arXiv preprint arXiv:2301.12334 (2023)

Um, S., Lee, S., Ye, J.C.: Don’t play favorites: Minority guidance for diffusion models. arXiv preprint arXiv:2301.12334 (2023)

2023 arXiv
[18]

In: CVPR

Sehwag, V., Hazirbas, C., Gordo, A., Ozgenel, F., Canton, C.: Generating high fidelity data from low-density regions using diffusion models. In: CVPR. pp. 11492– 11501 (2022)

2022
[19]

In: CVPR

Qin, Y., Zheng, H., Yao, J., Zhou, M., Zhang, Y.: Class-balancing diffusion models. In: CVPR. pp. 18434–18443 (2023)

2023
[20]

Frontiers in Medicine8, 729287 (2022)

Gunraj,H.,Sabri,A.,Koff,D.,Wong,A.:COVID-NetCT-2:Enhanceddeepneural networks for detection of COVID-19 from chest CT images through bigger, more diverse learning. Frontiers in Medicine8, 729287 (2022)

2022
[21]

Scientific Data6(1), 317 (2019)

Johnson, A.E.W., Pollard, T.J., Berkowitz, S.J., Greenbaum, N.R., Lungren, M.P., Deng, C., Mark, R.G., Horng, S.: MIMIC-CXR, a de-identified publicly avail- able database of chest radiographs with free-text reports. Scientific Data6(1), 317 (2019)

2019
[22]

GitHub repository,https://github.com/Project-MONAI/GenerativeModels/tree/ 7428fce/model-zoo/models/cxr_image_synthesis_latent_diffusion_model (2023), last accessed 2026-05-30

MONAI Generative Models: Chest X-ray with Latent Diffusion Models. GitHub repository,https://github.com/Project-MONAI/GenerativeModels/tree/ 7428fce/model-zoo/models/cxr_image_synthesis_latent_diffusion_model (2023), last accessed 2026-05-30

2023
[23]

GitHub repository,https://github.com/Warvito/ generative_chestxray(2023), last accessed 2026-05-30

Pinaya,W.H.L.(Warvito):LatentdiffusionmodelsforchestX-raygenerationusing MONAI Generative Models. GitHub repository,https://github.com/Warvito/ generative_chestxray(2023), last accessed 2026-05-30

2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

In: Pacific Symposium on Biocomputing 2021

Seyyed-Kalantari, L., Liu, G., McDermott, M., Chen, I.Y., Ghassemi, M.: CheX- clusion: Fairness gaps in deep chest X-ray classifiers. In: Pacific Symposium on Biocomputing 2021. pp. 232–243 (2020)

2021

[2] [2]

Nature Medicine27(12), 2176–2182 (2021)

Seyyed-Kalantari, L., Zhang, H., McDermott, M.B.A., Chen, I.Y., Ghassemi, M.: Underdiagnosisbiasofartificialintelligencealgorithmsappliedtochestradiographs in under-served patient populations. Nature Medicine27(12), 2176–2182 (2021)

2021

[3] [3]

arXiv preprint arXiv:2312.10083 (2023)

Yang, Y., Zhang, H., Gichoya, J.W., Katabi, D., Ghassemi, M.: The limits of fair medical imaging AI in the wild. arXiv preprint arXiv:2312.10083 (2023)

arXiv 2023

[4] [4]

In: AMIA Summits on Translational Science Proceedings

Lin, M., Xiao, Y., Hou, B., et al.: Evaluate underdiagnosis and overdiagnosis bias of deep learning model on primary open-angle glaucoma diagnosis in under-served populations. In: AMIA Summits on Translational Science Proceedings. p. 370 (2023)

2023

[5] [5]

In: CVPR

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684–10695 (2022)

2022

[6] [6]

In: NeurIPS

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS. vol. 33, pp. 6840–6851 (2020)

2020

[7] [7]

arXiv preprint arXiv:2207.12598 (2022)

Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

arXiv 2022

[8] [8]

arXiv preprint arXiv:2103.00020 (2021)

Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020 (2021)

arXiv 2021