Towards Fair Medical AI: Adversarial Debiasing of 3D CT Foundation Embeddings

Guangyao Zheng; Michael A. Jacobs; Vishwa S. Parekh; Vladimir Braverman

REVIEW 5 major objections 6 minor 14 references

Towards Fair Medical AI: Adversarial Debiasing of 3D CT Foundation Embeddings

T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a VAE trained against a demographic adversary can strip sex and age information out of 3D CT foundation embeddings while preserving lung-cancer risk prediction and reducing fairness gaps.

desk verdict Useful embedding-level debiasing with clear limits: linear probes show attenuation, not elimination, and the paper overstates the claim. read the letter →

arxiv 2502.04386 v1 pith:Q2KTPVWF submitted 2025-02-05 cs.CV cs.AIcs.LG

Guangyao Zheng , Michael A. Jacobs , Vladimir Braverman , Vishwa S. Parekh This is my paper · ORCID

classification cs.CVcs.AIcs.LG

keywords fairnessinmedicalAIadversarialdebiasingvariationalautoencoder3DCTfoundationembeddingslungcancerriskpredictionequalopportunitydifferencedatapoisoningdefense

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

Self-supervised embeddings of 3D CT scans carry demographic signals such as age and sex, and those signals can leak into downstream clinical predictions. This paper proposes a variational autoencoder trained in opposition to a demographic predictor: the predictor tries to recover age and sex from the latent code, and the encoder is pushed to make that difficult. On the NLST lung-cancer screening cohort, the transformed embeddings reduce sex-prediction AUC from 0.999 to 0.669, raise age-prediction error, and cut equal-opportunity differences for 1-year and 2-year cancer risk, while cancer-prediction accuracy stays at 0.986 and 0.977. The paper's central point is that fairness for black-box foundation embeddings can be achieved as a post-processing step, without retraining the foundation model.

What carries the argument

The central object is a variational autoencoder with a multi-branch demographic adversary attached to its latent code. The encoder maps the original 1408-dimensional CT embedding to mean and log-variance parameters, a reparameterization trick allows gradient flow, and a linear decoder reconstructs the input. The loss sums reconstruction mean squared error, KL divergence toward a unit Gaussian, and the adversary's prediction losses for sex (binary cross-entropy) and age (mean squared error). The encoder is trained to minimize the adversarial loss, pushing the latent space to be demographically neutral, while the reconstruction term keeps clinically useful information available. Because the transformation operates only on existing embeddings, it is downstream-task independent and compatible with black-box foundation models.

What would settle it

Train a multilayer perceptron or kernel SVM on the debiased embeddings to predict sex and age on the held-out test split; if sex AUC returns toward the original 0.999 or age MAE falls back toward 2.7, demographic information was not actually eliminated and the fairness claim is false.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that demographic encoding in a pretrained 3D CT embedding is separable from the clinical signal it carries. By routing 1408-dimension embeddings through a VAE with a 500-dimensional bottleneck and simultaneously training an adversary with separate branches for sex and age, the authors obtain transformed embeddings that preserve, and in AUC terms slightly improve, 1-year and 2-year lung-cancer risk prediction while making demographic attributes much harder to decode linearly. Reported numbers include a drop in sex-prediction AUC from 0.999 to 0.669, an increase in age-prediction mean absolute error from 2.734 to 4.169, and reductions in equal-opportunity difference for both sensitive attributes on both prediction horizons. The same debiased embeddings keep EOD close to zero when up to 100% of one demographic group's cancer labels are flipped, whereas original embeddings become severely unfair. The authors take this as evidence that the method eliminates multiple encoded demographic attributes simultaneously and independently of the downstream task.

Load-bearing premise

The claim that demographic information is eliminated is tested only with linear classifiers; if sex or age remains recoverable by a nonlinear adversary, the central fairness guarantee collapses.

Editorial extensions

If this is right

Debiasing can be applied after the fact to embeddings from a black-box 3D CT foundation model, so no retraining of the foundation model is required.
Equal-opportunity differences for sex and age drop for both 1-year and 2-year lung-cancer risk prediction when the debiased embeddings are used.
The debiased embeddings keep EOD near zero even when 100% of one demographic group's cancer labels are flipped, whereas the original embeddings show EOD approaching 1.
Multiple sensitive attributes are handled in a single transformation because the adversary has one branch per attribute.
A latent dimension of 500 marks a chosen balance: smaller bottlenecks erase more demography but also erase clinical signal, while larger ones preserve more of both.

Reading between the lines

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

A direct extension the paper leaves implicit is adding race as a third adversary branch; the NLST demographic table shows race is available, so the framework could be tested on the attribute most commonly associated with imaging bias.
Because the transformation is downstream-agnostic, one trained VAE could act as a fairness layer in front of any consumer of the same foundation embeddings, making debiasing a deployment-time operation rather than a training-time one.
The linear-probe evaluation sets a low bar for 'elimination'; measuring how demographic AUC grows as adversary capacity increases would turn the central claim into a quantifiable leakage curve.
Since age is itself a risk factor for lung cancer, the reported accuracy preservation may partly reflect the model ignoring age; a clinically deployable version would need to separate demographic leakage from legitimate age-related clinical signal.

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

The paper does something useful and slightly understated: it shows a VAE-based adversarial debiasing layer can be trained on top of black-box 3D CT foundation embeddings, stripping linear demographic predictability for sex and age while keeping lung cancer risk accuracy roughly intact. That is a genuinely convenient property—no need to retrain the foundation model, and no need to know its internals. The NLST experiments are real, the code is up, and the EOD reductions plus the poisoning robustness plots are consistent with the method working as advertised. It is an incremental combination of known components (VAE + adversarial branches), not a new theory, and the paper is honest enough to list race and other tasks as missing.

The soft spots are mostly about how far the claims outrun the evidence. "Demographic information is no longer encoded" is too strong. The evaluation only uses linear probes; sex AUC goes from 0.999 to 0.669 and age MAE from 2.73 to 4.17. That is attenuation, not elimination. A nonlinear probe could plausibly recover more, and the paper gives no adversary-strength analysis or latent-visualization sanity check. The same linear-only evaluation applies to the downstream task, so the "without compromising accuracy" claim is really "without compromising linear classifier accuracy." Also, the main text reports only accuracy for cancer prediction (0.986, 0.977) and hides the AUCs; the appendix has AUCs for debiased embeddings only, with no original-embedding AUC for comparison, which makes the trade-off hard to assess. No error bars, and the latent dimension of 500 is selected on a tune split with no discussion of variance across seeds.

None of this sinks the paper, but it does mean the headline should be "reduces linear demographic leakage" rather than "eliminates encoded demographics." The poisoning-robustness result is a downstream consequence of removing group signal, not an independent discovery, but it is still nice to see.

For a reader working on fairness in medical imaging or on embedding-space interventions, this is a solid, citable data point. It deserves a proper peer review: the method is reproducible, the evaluation is substantial, and the missing analyses (nonlinear probes, error bars, full AUC tables) are exactly what reviewers should ask for. I would send it to review, with the expectation of a major-revision request.

Referee Report

5 major / 6 minor

Summary. This manuscript proposes a variational autoencoder (VAE) with adversarial branches that maps 1408-dimensional 3D CT foundation embeddings into a lower-dimensional latent space intended to remove sex and age information while preserving lung cancer risk prediction. Using the NLST dataset, the authors report that after debiasing, linear sex-prediction AUC drops from 0.999 to 0.669, age-prediction MAE increases from 2.734 to 4.169, 1-year and 2-year cancer-prediction accuracy remain around 0.986 and 0.977, equal opportunity difference (EOD) decreases, and EOD stays low under demographic label-poisoning attacks. The paper claims that demographic information is 'eliminated' while clinical predictive accuracy is maintained.

Significance. If substantiated, the approach would be practically valuable because it provides post-hoc debiasing of closed, black-box foundation-model embeddings, handles multiple sensitive attributes simultaneously, and is downstream-task independent. The manuscript has clear strengths: it uses a public dataset with patient-wise splits, releases code, evaluates on a clinically relevant task, and includes a poisoning-robustness analysis. However, the central claims are currently supported only by linear-probe evaluations, the reported numbers do not match the 'elimination' language, and the downstream evaluation omits key metrics such as AUC and uncertainty estimates. The work is promising but needs substantial additional evidence before the main claims can be accepted.

major comments (5)

[Abstract; Section 3.1] The abstract's claim that the debiased embeddings 'effectively eliminate multiple encoded demographic information' is not supported by the results in Section 3.1: after debiasing, sex-prediction AUC is 0.669, which remains far above chance (0.5), and age-prediction MAE is 4.169, which is close to the trivial baseline of predicting the mean age given the NLST age standard deviation of 5.0 years. The paper should either weaken the claim to 'substantially reduce linearly decodable demographic information' or provide evidence of elimination, such as AUC at chance level with confidence intervals and comparison to a no-information baseline.
[Section 2.3; Section 3.1] Both the debiasing adversary and the evaluation probes appear to be restricted to linear families, so the central fairness result may not generalize to nonlinear adversaries. The adversary is described only as 'multiple branches' with no hidden-layer specification, and the evaluation uses a sigmoid classifier for sex and linear regression for age. A latent representation can be linearly unreadable while still encoding demographics nonlinearly. The authors should specify the adversary architecture, test with nonlinear probes such as an MLP or kernel SVM, and report how demographic predictability varies with adversary capacity.
[Section 3.1; Figure 3] The claim that debiasing 'does not compromise predictive accuracy' rests on accuracy values of 0.986 and 0.977 for an imbalanced cancer-prediction task, where always predicting the majority class can yield high accuracy. No AUC values from Figure 3 are reported in the main text, and Appendix B reports post-debiasing AUCs of 0.897 and 0.856 for latent dimension 500 but does not give the corresponding original-embedding AUCs or any variability estimate. The paper should report AUC with confidence intervals for both original and debiased embeddings, along with class prevalence.
[Appendix B] The latent dimension is selected using a 'tune split' that is not described in Section 2.1, and no repeated runs or random seeds are reported anywhere in the paper. This makes it impossible to assess whether the choice of 500 and the reported improvements are stable or overfit to the tuning set. The authors should define the tune split, explain how it was derived from the training data, and provide means and standard deviations over multiple runs.
[Section 3.2] The poisoning-robustness result is presented as an independent finding, but it follows directly from reduced demographic predictability: if the model cannot distinguish demographic groups, targeted label flipping for one group cannot produce large EOD. This experiment is useful as an application of the debiased embeddings, but it should be framed as a consequence rather than a separate validation of the method.

minor comments (6)

[Abstract] The term 'Variation Autoencoder' should be 'Variational Autoencoder'.
[Table A.1] The percentages appear off by a factor of 100; for example, White patients are listed as '0.93%' of the training set, but they actually constitute approximately 93%.
[Section 3.2] The phrase 'Similar to experiment 2' should refer to Experiment 1.
[Keywords] The keywords include 'Progressive Streaming' and 'Data-Efficiency', which appear unrelated to the manuscript's content and may be a copy-paste artifact.
[Figure B.1 caption] The caption 'The difference for sex is the larger, the better' should be reworded for clarity.
[Section 2.2] The manuscript states that statistical significance was set at p < 0.05, but no significance tests or p-values are reported anywhere in the results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the debiasing result is empirically evaluated on held-out test data, and the self-citation is non-load-bearing.

full rationale

Walking the derivation chain, the paper's central results are (i) original CT foundation embeddings contain demographic signal, (ii) adversarial VAE debiasing reduces that signal, and (iii) lung cancer risk prediction accuracy is retained. Step (i) is independently measured in Section 3.1 with linear classifiers (sex AUC 0.999 -> 0.669, age MAE 2.734 -> 4.169) on a held-out test split, not imported solely from the self-citation in Section 1. Step (ii) is the direct result of optimizing the adversarial loss described in Section 2.3; reporting that linear demographic probes lose accuracy on a held-out test set is a standard train/test evaluation of that objective, not a fitted parameter renamed as a prediction. The latent-dimension choice in Appendix B is model selection on a tune split, with the main numbers reported on the test split. Step (iii) is evaluated with independent classifiers on the same splits. The poisoning robustness result (Section 3.2) is an empirical consequence of reduced group information rather than a circular restatement; it is a downstream validation, and the paper does not claim it is independent of the debiasing objective. The only self-citation (Zheng et al., 2024) motivates the premise that original embeddings encode demographics, but the current paper reproduces that premise with its own baseline classifiers, so the citation is not load-bearing. The concern that nonlinear demographic leakage may remain is a correctness/robustness limitation (the paper only uses linear probes), not a circularity.

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

No new physical or mathematical entities are postulated. The free parameters are hyperparameters of the VAE/adversary; the most consequential is the latent dimension, chosen on the tune split. The key domain assumptions are the sufficiency of linear probes and the choice of EOD as the fairness metric.

free parameters (4)

Latent dimension = 500
Selected empirically on a tune split (Appendix B) to balance sex-predictability reduction against cancer-prediction preservation; the selection criterion is not formalized.
Learning rates = 0.0005 encoder/decoder, 0.002 adversary
Hand-chosen; no sensitivity analysis.
Batch size = 32
Hand-chosen; no sensitivity analysis.
Number of epochs = 100
Hand-chosen; no early stopping described.

assumptions (5)

domain assumption Linear classifiers are sufficient probes for demographic encoding.
The paper measures debiasing effectiveness only with linear models (Section 2.2); if nonlinear encoding remains, the central claim fails.
domain assumption EOD is an appropriate fairness metric for this clinical task.
Fairness is quantified via Equal Opportunity Difference (Section 3.1); other fairness definitions could yield different conclusions.
domain assumption NLST is representative for evaluating lung cancer screening fairness.
The dataset is a single public screening trial; generalization to other populations or imaging protocols is untested.
standard math The CT Foundation model embeddings are fixed and treated as black-box inputs.
The debiasing operates on provided 1408-dim features without modifying the foundation model, which is a valid but restrictive setup.
standard math Variational autoencoder reconstruction preserves task-relevant information.
The VAE loss balances reconstruction and KL divergence; the paper assumes task-relevant features survive the bottleneck.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Fair Medical AI: Adversarial Debiasing of 3D CT Foundation Embeddings." pith.science (2026). https://pith.science/paper/Q2KTPVWF

@misc{pith2026250204386,
  author       = {Pith},
  title        = {Pith review of: Towards Fair Medical AI: Adversarial Debiasing of 3D CT Foundation Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q2KTPVWF}},
  note         = {Machine review of arXiv:2502.04386}
}

read the original abstract

Self-supervised learning has revolutionized medical imaging by enabling efficient and generalizable feature extraction from large-scale unlabeled datasets. Recently, self-supervised foundation models have been extended to three-dimensional (3D) computed tomography (CT) data, generating compact, information-rich embeddings with 1408 features that achieve state-of-the-art performance on downstream tasks such as intracranial hemorrhage detection and lung cancer risk forecasting. However, these embeddings have been shown to encode demographic information, such as age, sex, and race, which poses a significant risk to the fairness of clinical applications. In this work, we propose a Variation Autoencoder (VAE) based adversarial debiasing framework to transform these embeddings into a new latent space where demographic information is no longer encoded, while maintaining the performance of critical downstream tasks. We validated our approach on the NLST lung cancer screening dataset, demonstrating that the debiased embeddings effectively eliminate multiple encoded demographic information and improve fairness without compromising predictive accuracy for lung cancer risk at 1-year and 2-year intervals. Additionally, our approach ensures the embeddings are robust against adversarial bias attacks. These results highlight the potential of adversarial debiasing techniques to ensure fairness and equity in clinical applications of self-supervised 3D CT embeddings, paving the way for their broader adoption in unbiased medical decision-making.

Figures

Figures reproduced from arXiv: 2502.04386 by the authors.

**Figure 1.** Conceptual illustration of our framework. The original 3D CT Foundation Model embedding encodes demographic information, which can lead to bias in downstream tasks. Our VAE can debias multiple demographic information while preserving downstream task performance. 2. Materials and Methods 2.1. Dataset and 3D CT Foundation Model Embeddings This retrospective study utilized the publicly available National Lung Screeni… view at source ↗

**Figure 2.** (a) ROC curve of linear classifier performance on sex prediction comparing original embedding (blue) vs. embedding after VAE reconstruction (red). (b) Scatter plot of linear classifier performance on age prediction comparing original embedding (blue) vs. embedding after VAE reconstruction (red). Equal Opportunity Difference (EOD) (Obermeyer et al., 2019). EOD measures disparities in model performance across demogr… view at source ↗

**Figure 3.** ROC curve of linear classifier performance on (a) cancer in 1 year, (b) cancer in 2 years prediction comparing original embedding (blue) vs. embedding after VAE reconstruction (red) lung cancer prediction task, the accuracy remained unchanged at 0.986 for both the original and debiased embeddings. Similarly, for the 2-year lung cancer prediction, accuracy was nearly identical at 0.977 across both embedding types. Th… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: EOD comparison for original embedding (blue) vs. VAE debiased embedding (red) for cancer in 1 year prediction when the (a) male, (b) female patient’s cancer in 1 year label is poisoned by X percentage (X = 0, 25, 50, 75, 100). 3.2. Experiment 2: Robustness to Data Pois…

**Figure 5.** Figure 5: EOD comparison for original embedding (blue) vs. VAE debiased embedding (red) for cancer in 2 years prediction when the (a) male, (b) female patient’s cancer in 2 years label is poisoned by X percentage (X = 0, 25, 50, 75, 100). feasible to eliminate demographic biases…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[1]

Debias-CLR: A Contrastive Learning Based Debiasing Method for Algorithmic Fairness in Healthcare Applications

Ankita Agarwal, Tanvi Banerjee, William Romine, and Mia Cajita. Debias-clr: A contrastive learning based debiasing method for algorithmic fairness in healthcare applications. arXiv preprint arXiv:2411.10544, 2024

work page Pith review arXiv 2024
[2]

Two-step adversarial debiasing with partial learning -- medical image case-studies

Ramon Correa, Jiwoong Jason Jeong, Bhavik Patel, Hari Trivedi, Judy W Gichoya, and Imon Banerjee. Two-step adversarial debiasing with partial learning--medical image case-studies. arXiv preprint arXiv:2111.08711, 2021

work page Pith review arXiv 2021
[3]

Efficient adversarial debiasing with concept activation vector—medical image case-studies

Ramon Correa, Khushbu Pahwa, Bhavik Patel, Celine M Vachon, Judy W Gichoya, and Imon Banerjee. Efficient adversarial debiasing with concept activation vector—medical image case-studies. Journal of Biomedical Informatics, 149: 0 104548, 2024

work page 2024
[4]

Ai recognition of patient race in medical imaging: a modelling study

Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, et al. Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4 0 (6): 0 e406--e414, 2022

work page 2022
[5]

Nlst embeddings in ct\_foundation\_demo.ipynb, Accessed: Oct 26 2024

Google Health. Nlst embeddings in ct\_foundation\_demo.ipynb, Accessed: Oct 26 2024. URL https://github.com/Google-Health/imaging-research/tree/master/ct-foundation

work page 2024
[6]

Debiased noise editing on foundation models for fair medical image classification

Ruinan Jin, Wenlong Deng, Minghui Chen, and Xiaoxiao Li. Debiased noise editing on foundation models for fair medical image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 164--174. Springer, 2024

work page 2024
[7]

Taking medical imaging embeddings 3d, Accessed: Oct 26 2024

Atilla Kiraly and Madeleine Traverse. Taking medical imaging embeddings 3d, Accessed: Oct 26 2024. URL https://research.google/blog/taking-medical-imaging-embeddings-3d/

work page 2024
[8]

Hidden in Plain Sight: Undetectable Adversarial Bias Attacks on Vulnerable Patient Populations

Pranav Kulkarni, Andrew Chan, Nithya Navarathna, Skylar Chan, Paul H Yi, and Vishwa S Parekh. Hidden in plain sight: Undetectable adversarial bias attacks on vulnerable patient populations. arXiv preprint arXiv:2402.05713, 2024

work page Pith review arXiv 2024

Show all 14 references

[9]

The national lung screening trial: overview and study design

National Lung Screening Trial Research Team NLST. The national lung screening trial: overview and study design. Radiology, 258 0 (1): 0 243--253, 2011

2011
[10]

Dissecting racial bias in an algorithm used to manage the health of populations

Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366 0 (6464): 0 447--453, 2019

2019
[11]

An adversarial training framework for mitigating algorithmic biases in clinical machine learning

Jenny Yang, Andrew AS Soltan, David W Eyre, Yang Yang, and David A Clifton. An adversarial training framework for mitigating algorithmic biases in clinical machine learning. NPJ digital medicine, 6 0 (1): 0 55, 2023

2023
[12]

Advancing multimodal medical capabilities of gemini

Lin Yang, Shawn Xu, Andrew Sellergren, Timo Kohlberger, Yuchen Zhou, Ira Ktena, Atilla Kiraly, Faruk Ahmed, Farhad Hormozdiari, Tiam Jaroensri, et al. Advancing multimodal medical capabilities of gemini. arXiv preprint arXiv:2405.03162, 2024

2024 arXiv
[13]

Mitigating unwanted biases with adversarial learning

Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. Mitigating unwanted biases with adversarial learning. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 335--340, 2018

2018
[14]

Demographic predictability in 3d ct foundation embeddings

Guangyao Zheng, Michael A Jacobs, and Vishwa S Parekh. Demographic predictability in 3d ct foundation embeddings. arXiv preprint arXiv:2412.00110, 2024

2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Debias-CLR: A Contrastive Learning Based Debiasing Method for Algorithmic Fairness in Healthcare Applications

Ankita Agarwal, Tanvi Banerjee, William Romine, and Mia Cajita. Debias-clr: A contrastive learning based debiasing method for algorithmic fairness in healthcare applications. arXiv preprint arXiv:2411.10544, 2024

work page Pith review arXiv 2024

[2] [2]

Two-step adversarial debiasing with partial learning -- medical image case-studies

Ramon Correa, Jiwoong Jason Jeong, Bhavik Patel, Hari Trivedi, Judy W Gichoya, and Imon Banerjee. Two-step adversarial debiasing with partial learning--medical image case-studies. arXiv preprint arXiv:2111.08711, 2021

work page Pith review arXiv 2021

[3] [3]

Efficient adversarial debiasing with concept activation vector—medical image case-studies

Ramon Correa, Khushbu Pahwa, Bhavik Patel, Celine M Vachon, Judy W Gichoya, and Imon Banerjee. Efficient adversarial debiasing with concept activation vector—medical image case-studies. Journal of Biomedical Informatics, 149: 0 104548, 2024

work page 2024

[4] [4]

Ai recognition of patient race in medical imaging: a modelling study

Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, et al. Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4 0 (6): 0 e406--e414, 2022

work page 2022

[5] [5]

Nlst embeddings in ct\_foundation\_demo.ipynb, Accessed: Oct 26 2024

Google Health. Nlst embeddings in ct\_foundation\_demo.ipynb, Accessed: Oct 26 2024. URL https://github.com/Google-Health/imaging-research/tree/master/ct-foundation

work page 2024

[6] [6]

Debiased noise editing on foundation models for fair medical image classification

Ruinan Jin, Wenlong Deng, Minghui Chen, and Xiaoxiao Li. Debiased noise editing on foundation models for fair medical image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 164--174. Springer, 2024

work page 2024

[7] [7]

Taking medical imaging embeddings 3d, Accessed: Oct 26 2024

Atilla Kiraly and Madeleine Traverse. Taking medical imaging embeddings 3d, Accessed: Oct 26 2024. URL https://research.google/blog/taking-medical-imaging-embeddings-3d/

work page 2024

[8] [8]

Hidden in Plain Sight: Undetectable Adversarial Bias Attacks on Vulnerable Patient Populations

Pranav Kulkarni, Andrew Chan, Nithya Navarathna, Skylar Chan, Paul H Yi, and Vishwa S Parekh. Hidden in plain sight: Undetectable adversarial bias attacks on vulnerable patient populations. arXiv preprint arXiv:2402.05713, 2024

work page Pith review arXiv 2024