Pith. sign in

REVIEW 2 major objections 5 minor 25 references

Are generative models fair? A study of racial bias in dermatological image generation

T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A VAE trained on clinical dermatology images reconstructs lighter skin tones better than darker ones even when training data is balanced 50/50, and its built-in uncertainty estimates do not flag the gap.

desk verdict A solid controlled study of skin-tone bias in a dermatology VAE, but the 'independent of representation' claim is confounded by condition mix and needs softening. read the letter →

arxiv 2501.11752 v2 pith:DHVZNMPZ submitted 2025-01-20 cs.CV

Miguel López-Pérez , S{o}ren Hauberg , Aasa Feragen This is my paper · ORCID
classification cs.CV
keywords racialbiasdermatologygenerativemodelsvariationalautoencoderFitzpatrickskintypeuncertaintyquantificationrepresentationfairness
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 asks whether a deep generative model trained on clinical dermatology images is fair across skin tones, and answers with a controlled VAE study on the Fitzpatrick17k dataset. It establishes that reconstruction quality tracks subgroup representation: more dark-skin images in training lowers dark-skin reconstruction error. It then reports a second, representation-independent effect: even at a 50/50 training mix the VAE reconstructs lighter skin (FST 1-2) better than darker skin (FST 5-6). It also finds that the VAE's averaged latent standard deviation, its built-in uncertainty estimate, does not distinguish the two subgroups and so cannot serve as a warning sign for this bias. The importance is that balancing datasets is necessary but apparently not sufficient for fair generative models, and uncertainty quantification needs rethinking.

What carries the argument

The central object is a VAE with a Gaussian latent ($q_\phi(z|x) = \mathcal{N}(\mu_\phi(x), \mathrm{diag}(\sigma_\phi^2(x)))$), a Gaussian decoder, and a perceptual loss based on VGG19 feature maps. The argument is carried by comparing two quantities across three training-set configurations: the reconstruction likelihood (MSE) and the averaged latent standard deviation. The three configurations -- Dataset A (100% light), Dataset B (50/50), Dataset C (100% dark) -- instantiate the representation axis, while separate light and dark test sets isolate subgroup performance.

What would settle it

Train the same VAE on a balanced 50/50 dataset, then test on light and dark sets that are matched exactly on fine-grained skin-condition labels (for example, only common conditions with nearly equal counts); if the light-minus-dark MSE gap collapses to zero, the representation-independent bias claim fails, while a gap on matched sets would confirm the bias as skin-tone-related.

Watch

Extended reading notes

Core claim

The authors train a VAE with a VGG19 perceptual loss on Fitzpatrick17k images under three training configurations -- all light (FST 1-2), 50/50 mixed, and all dark (FST 5-6) -- and test on fixed light and dark test sets. Their central discovery is that darker skin is reconstructed with systematically higher MSE (lower likelihood) than lighter skin across all configurations, including the balanced one, so the performance gap is not explained by representation alone. A second discovery is that the averaged latent standard deviation of the stochastic encoder looks nearly identical for light and dark test sets even where reconstruction performance differs, making the VAE's natural uncertainty estimate useless as a fairness diagnostic. The paper leaves open whether the residual gap is caused by skin tone itself or by differences in the fine-grained skin conditions present in the two groups.

Load-bearing premise

The paper's key assumption is that the light and dark test groups are otherwise comparable, but the fine-grained skin-condition distributions differ substantially between them, so the residual performance gap could be driven by condition difficulty rather than skin tone itself.

Editorial extensions

If this is right

  • A balanced 50/50 training set does not equalize VAE reconstruction quality between lighter and darker skin tones.
  • Latent standard deviation cannot be relied on as a fairness alarm when deploying such a VAE in clinical settings.
  • The common real-world scenario of light-dominated training data produces the largest dark-skin performance gap, so skin-tone bias should be expected in generative models trained on existing dermatology collections.
  • Improving dataset diversity may be necessary but is not sufficient; the source of the representation-independent gap needs separate investigation.

Reading between the lines

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

  • If the fine-grained condition mismatch is the true driver, then a matched test set that holds skin-condition distributions equal across skin tones should erase or shrink the residual MSE gap; this is a testable extension the paper does not run.
  • The failure of latent-variance UQ suggests exploring decoder-side heteroscedastic noise or predictive intervals on reconstructed pixels, which may better reflect subgroup difficulty.
  • Because the Fitzpatrick scale itself conflates skin tone with other attributes, the 'darker skin' group may be a proxy for correlated image properties such as texture, lighting, and lesion morphology, implying the bias could persist even with perfect skin-tone balance.
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

2 major / 5 minor

Summary. The paper studies racial bias in a VAE trained on the Fitzpatrick17k dermatology dataset. It varies the proportion of light (FST 1-2) and dark (FST 5-6) skin tones in training sets across three configurations (light-only, 50/50 mixed, dark-only), and measures reconstruction MSE and latent posterior standard deviation on held-out light and dark test sets. The authors report that reconstruction performance tracks representation, that a performance gap favoring light skin persists even under a 50/50 skin-tone split, and that latent standard deviation does not reflect this gap. They conclude that generative models are generally biased against darker skin tones and that uncertainty estimates are not useful fairness indicators.

Significance. If the residual gap under balanced skin-tone representation is confirmed, the paper makes a valuable contribution: it extends fairness analysis from discriminative to generative models, provides a reproducible protocol based on repeated independent runs, and raises a practical warning about uncertainty-based monitoring. The authors are also transparent about the condition-confounding limitation in §5.2. However, because the main abstract claim depends on ruling out condition-mix differences, the current evidence is conditional rather than definitive.

major comments (2)
  1. [§5.2, Figures 7-8] The claim "even independently of representation, the VAE performs better for lighter skin tones" (abstract) is not established because the light and dark test sets differ in fine-grained condition distribution. The authors' own Figure 7c shows substantial differences at the fine-grained level, and Figure 8 shows that condition-level MSE is uneven; §5.2 explicitly leaves open the disentanglement of skin tone and condition. The Dataset B comparison therefore confounds skin tone with condition difficulty. To substantiate the claim, please report the gap within condition-matched subgroups or after adjusting for condition prevalence, or restrict the conclusion to "with balanced skin-tone sampling."
  2. [§3.4, abstract, §5.1] The phrase "independent of representation" overstates the experimental control: only the coarse skin-tone label is balanced, not the joint distribution of skin tone and diagnosis. Since the training sets for Dataset B are 50/50 in skin tone but not in disease mix, the model may have encountered different diseases for each group; this leaves open an alternative explanation for the residual gap. Please either redesign the sampling to balance fine-grained conditions or soften the language throughout.
minor comments (5)
  1. [§3.4] The sentence "one with darker skin, comprising FST 4-6" contradicts the test-set definition in the same paragraph (FST 5-6) and the statement that "FST 3-4" images are discarded; please clarify which Fitzpatrick types are actually used.
  2. [Figure 8] The subcaptions "(a) Coarse level... (b) Medium level... (c) Fine-grained level..." do not match the main caption "Averaged MSE of skin conditions based on fine-grained labels" or the actual subplot labels (Dataset A/B/C); please correct the mismatch.
  3. [§3.1] In the description of the encoder, "qϕ(x|z)" should be "qϕ(z|x)" in the sentence "combining an stochastic encoder qϕ(x|z)."
  4. [Eq. (2)] There is a typo "widht" for "width," and the normalization factor would be clearer written as 1/(C^l W^l H^l).
  5. [Title and abstract] The paper evaluates reconstruction, not generation of new samples; the title and abstract's "image generation" should be aligned with the actual experiments or note that generation is not evaluated.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the VAE performance gap is measured on held-out test sets, not derived from the model's own fitted quantities.

full rationale

This paper is an empirical evaluation, not a derivation. The authors train a VAE on subsampled Fitzpatrick17k training configurations and evaluate reconstruction MSE and latent standard deviation on held-out test sets sampled separately (Section 3.4, Figure 3). The central claim that the VAE performs better for lighter skin tones even under a 50/50 training split is an observed outcome of that held-out evaluation, not an assumption built into the training objective or the subsampling procedure. No fitted parameter is renamed as a prediction: the latent standard deviation is measured and found to be uninformative, which is a falsifiable empirical finding rather than a consequence of construction. The experimental design is borrowed from Larrazabal et al. [14], but that reference is external, not self-citational, and it supplies only the subsampling protocol, not the conclusion. The only substantive weakness is a potential confound: Section 5.2 and Figure 7c show that fine-grained condition prevalence differs between light and dark skin groups, so the 'independent of representation' claim may be confounded by condition difficulty. However, confounding is a validity threat, not circularity; the paper explicitly flags this open question ('Can we disentangle the influence of skin tone versus specific skin condition on the results during image generation?'). There is no equation in which the output is defined in terms of the input, no self-citation chain carrying a load-bearing premise, and no ansatz smuggled in via prior work. The paper is self-contained against external benchmarks and honestly reports its limitations, so the circularity score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical study, so it introduces no fitted constants in the sense of a derivation. The central causal claim rests on three assumptions: the Fitzpatrick labels are meaningful, the skin-tone groups are otherwise comparable, and the ImageNet-pretrained perceptual features are skin-tone neutral. The second is acknowledged to be questionable, and the third is unexamined.

free parameters (1)
  • Skin-tone group definitions = Light: Fitzpatrick 1-2; Dark: Fitzpatrick 5-6; Fitzpatrick 3-4 excluded
    The binary grouping and exclusion of middle skin types is chosen by hand and directly defines the subgroups compared in every result; different boundaries could change the measured gap.
assumptions (3)
  • domain assumption Fitzpatrick skin type labels are accurate and meaningful for grouping racial and skin-tone subgroups.
    All subgroup comparisons rely on these labels from the Fitzpatrick17k dataset; the paper itself cites work questioning the Fitzpatrick scale's adequacy (Section 5.3).
  • ad hoc to paper The light and dark test groups are comparable except for skin tone, so a residual performance gap reflects skin-tone bias rather than condition or image-difficulty differences.
    This is the assumption behind the "independent of representation" claim; Section 5.2 shows it is violated at the fine-grained diagnostic level and leaves the disentangling as an open question.
  • ad hoc to paper The VGG19 perceptual loss pretrained on ImageNet provides a skin-tone-neutral measure of reconstruction quality.
    The loss and evaluation metric use features from an ImageNet-pretrained model (Section 3.2), and the paper does not test whether those features are biased by skin tone.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are generative models fair? A study of racial bias in dermatological image generation." pith.science (2026). https://pith.science/paper/DHVZNMPZ

@misc{pith2026250111752,
  author       = {Pith},
  title        = {Pith review of: Are generative models fair? A study of racial bias in dermatological image generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHVZNMPZ}},
  note         = {Machine review of arXiv:2501.11752}
}
read the original abstract

Racial bias in medicine, such as in dermatology, presents significant ethical and clinical challenges. This is likely to happen because there is a significant underrepresentation of darker skin tones in training datasets for machine learning models. While efforts to address bias in dermatology have focused on improving dataset diversity and mitigating disparities in discriminative models, the impact of racial bias on generative models remains underexplored. Generative models, such as Variational Autoencoders (VAEs), are increasingly used in healthcare applications, yet their fairness across diverse skin tones is currently not well understood. In this study, we evaluate the fairness of generative models in clinical dermatology with respect to racial bias. For this purpose, we first train a VAE with a perceptual loss to generate and reconstruct high-quality skin images across different skin tones. We utilize the Fitzpatrick17k dataset to examine how racial bias influences the representation and performance of these models. Our findings indicate that VAE performance is, as expected, influenced by representation, i.e. increased skin tone representation comes with increased performance on the given skin tone. However, we also observe, even independently of representation, that the VAE performs better for lighter skin tones. Additionally, the uncertainty estimates produced by the VAE are ineffective in assessing the model's fairness. These results highlight the need for more representative dermatological datasets, but also a need for better understanding the sources of bias in such model, as well as improved uncertainty quantification mechanisms to detect and address racial bias in generative models for trustworthy healthcare technologies.

Figures

Figures reproduced from arXiv: 2501.11752 by the authors.

Figure 1
Figure 1. Example predictions from our VAE model trained on a balanced subset of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Distribution of samples according to the FST in the Fitzpatrick17k [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Likelihood or Mean Square Error (MSE) of the VAEs in the test sets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Example reconstruction of lighter skin tones. Reconstructions are pro [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Example reconstruction of darker skin tones. Reconstructions are pro [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Averaged standard deviation of the latent variable of the VAEs in the test [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 8
Figure 8. Figure 8: In general, we observe that the behavior is uneven. ‘Dataset B – Mixed’ [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 7
Figure 7. Figure 7: Normalized count of skin conditions categorized across varying levels of [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Averaged MSE of skin conditions based on fine-grained labels. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    In: 2023 IEEE 6th International Conference on Multimedia Information Processing and Retrieval (MIPR)

    Almuzaini, A.A., Dendukuri, S.K., Singh, V.K.: Toward fairness across skin tones in dermatological image processing. In: 2023 IEEE 6th International Conference on Multimedia Information Processing and Retrieval (MIPR). pp. 1–7 (2023).https: //doi.org/10.1109/MIPR59079.2023.00030

    work page arXiv 2023

  2. [2]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2024)

    Benmalek, A., Cintas, C., Tadesse, G.A.: Impact of skin tone diversity on out- of-distribution detection methods in dermatology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2024)

    work page 2024

  3. [3]

    Science advances 8(31), eabq6147 (2022)

    Daneshjou, R., Vodrahalli, K., Novoa, R.A., Jenkins, M., Liang, W., Rotemberg, V., Ko, J., Swetter, S.M., Bailey, E.E., Gevaert, O., et al.: Disparities in derma- tology ai performance on a diverse, curated clinical image set. Science advances 8(31), eabq6147 (2022)

    work page 2022

  4. [4]

    Academic Emergency Medicine24(8), 895–904 (2017)

    Dehon, E., Weiss, N., Jones, J., Faulconer, W., Hinton, E., Sterling, S.: A sys- tematic review of the impact of physician implicit racial bias on clinical decision making. Academic Emergency Medicine24(8), 895–904 (2017)

    work page 2017

  5. [5]

    The Lancet Digital Health 4(6), e406–e414 (2022)

    Gichoya, J.W., Banerjee, I., Bhimireddy, A.R., Burns, J.L., Celi, L.A., Chen, L.C., Correa, R., Dullerud, N., Ghassemi, M., Huang, S.C., et al.: Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health 4(6), e406–e414 (2022)

    work page 2022

  6. [6]

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

    Gottfrois, P., Gröger, F., Andriambololoniaina, F.H., Amruthalingam, L., Gonzalez-Jimenez, A., Hsu, C., Kessy, A., Lionetti, S., Mavura, D., Ng’ambi, W., et al.: Passion for dermatology: Bridging the diversity gap with pigmented skin images from sub-saharan africa. In: International Conference on Medical Image Computing and Computer-Assisted Intervention....

    work page 2024

  7. [7]

    Nature Medicine30(2), 573–583 (2024)

    Groh, M., Badri, O., Daneshjou, R., Koochek, A., Harris, C., Soenksen, L.R., Doraiswamy, P.M., Picard, R.: Deep learning-aided decision support for diagnosis of skin disease across skin tones. Nature Medicine30(2), 573–583 (2024)

    work page 2024

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Groh, M., Harris, C., Soenksen, L., Lau, F., Han, R., Kim, A., Koochek, A., Badri, O.: Evaluating deep neural networks trained on clinical images in dermatology with the fitzpatrick 17k dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1820–1828 (2021)

    work page 2021

Show all 25 references
  1. [9]

    In: 2017 IEEE winter conference on applications of computer vision (WACV)

    Hou, X., Shen, L., Sun, K., Qiu, G.: Deep feature consistent variational au- toencoder. In: 2017 IEEE winter conference on applications of computer vision (WACV). pp. 1133–1141. IEEE (2017)

    2017

  2. [10]

    Advances in neural information processing systems 31 (2018)

    Huang, H., He, R., Sun, Z., Tan, T., et al.: Introvae: Introspective variational autoencoders for photographic image synthesis. Advances in neural information processing systems 31 (2018)

    2018

  3. [11]

    In: Workshop on Clinical Image-Based Procedures

    Kalb, T., Kushibar, K., Cintas, C., Lekadir, K., Diaz, O., Osuala, R.: Revisiting skin tone fairness in dermatological lesion classification. In: Workshop on Clinical Image-Based Procedures. pp. 246–255. Springer (2023)

    2023

  4. [12]

    López-Pérez et al

    Kingma,D.P.,Welling,M.:Auto-encodingvariationalbayes.In:InternationalCon- ference on Learning Representations (ICLR) (2014) 14 M. López-Pérez et al

    2014

  5. [13]

    In: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention

    Kinyanjui, N.M., Odonga, T., Cintas, C., Codella, N.C., Panda, R., Sattigeri, P., Varshney, K.R.: Fairness of classifiers across skin tones in dermatology. In: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention. pp. 320–329. Springer (2020)

    2020

  6. [14]

    Proceedings of the National Academy of Sciences117(23), 12592–12594 (2020)

    Larrazabal, A.J., Nieto, N., Peterson, V., Milone, D.H., Ferrante, E.: Gender im- balance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences117(23), 12592–12594 (2020)

    2020

  7. [15]

    ACM Computing Surveys55(9), 1–46 (2023)

    Li, B., Qi, P., Liu, B., Di, S., Liu, J., Pei, J., Yi, J., Zhou, B.: Trustworthy ai: From principles to practices. ACM Computing Surveys55(9), 1–46 (2023)

    2023

  8. [16]

    In: Proceedings of the ACM conference on health, inference, and learning

    Oakden-Rayner, L., Dunnmon, J., Carneiro, G., Ré, C.: Hidden stratification causes clinically meaningful failures in machine learning for medical imaging. In: Proceedings of the ACM conference on health, inference, and learning. pp. 151–159 (2020)

    2020

  9. [17]

    In: MICCAI Workshop onFairnessof AIin Medical Imaging.pp

    Olesen, V., Weng, N., Feragen, A., Petersen, E.: Slicing through bias: Explaining performance gaps in medical image analysis using slice discovery methods. In: MICCAI Workshop onFairnessof AIin Medical Imaging.pp. 3–13.Springer(2024)

    2024

  10. [18]

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

    Petersen, E., Feragen, A., da Costa Zemsch, M.L., Henriksen, A., Wiese Chris- tensen, O.E., Ganz, M., Initiative, A.D.N.: Feature robustness and sex differences in medical imaging: a case study in mri-based alzheimer’s disease detection. In: International Conference on Medical...

    2022

  11. [19]

    Patterns4(7) (2023)

    Petersen, E., Holm, S., Ganz, M., Feragen, A.: The path toward equal performance in medical machine learning. Patterns4(7) (2023)

    2023

  12. [20]

    In: Xing, E.P., Jebara, T

    Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and ap- proximate inference in deep generative models. In: Xing, E.P., Jebara, T. (eds.) Proceedings of the 31st International Conference on Machine Learning. Proceed- ings of Machine Learning Research, vol. ...

    2014

  13. [21]

    In: NeurIPS 2022 Workshop on Synthetic Data for Empowering ML Research (2022)

    Sagers, L.W., Diao, J.A., Groh, M., Rajpurkar, P., Adamson, A., Manrai, A.K.: Im- proving dermatology classifiers across populations using images generated by large diffusion models. In: NeurIPS 2022 Workshop on Synthetic Data for Empowering ML Research (2022)

    2022

  14. [22]

    Advances in Neural Information Processing Systems36 (2024)

    Schumann, C., Olanubi, F., Wright, A., Monk, E., Heldreth, C., Ricco, S.: Consen- sus and subjectivity of skin tone annotation for ml fairness. Advances in Neural Information Processing Systems36 (2024)

    2024

  15. [23]

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015),https://arxiv.org/abs/1409.1556

    2015 arXiv

  16. [24]

    Jama 314(6), 555–556 (2015)

    Williams, D.R., Wyatt, R.: Racial bias in health care and health: challenges and opportunities. Jama 314(6), 555–556 (2015)

    2015

  17. [25]

    arXiv preprint arXiv:2111.11665 (2021)

    Zhou, Y., Huang, S.C., Fries, J.A., Youssef, A., Amrhein, T.J., Chang, M., Baner- jee, I., Rubin, D., Xing, L., Shah, N., et al.: Radfusion: Benchmarking performance and fairness for multimodal pulmonary embolism detection from ct and ehr. arXiv preprint arXiv:2111.11665 (2021)

    2021 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.