Toward Robust Medical Fairness: Debiased Dual-Modal Alignment via Text-Guided Attribute-Disentangled Prompt Learning for Vision-Language Models

Benteng Ma; Jiang He; Qi Dou; Yong Xia; Yuexuan Xia; Zhiyong Wang

REVIEW 3 major objections 4 minor 59 references

Toward Robust Medical Fairness: Debiased Dual-Modal Alignment via Text-Guided Attribute-Disentangled Prompt Learning for Vision-Language Models

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DualFairVL claims that jointly debiasing text and visual branches in a frozen CLIP, guided by text anchors, yields fairness and accuracy that survive distribution shifts, outperforming methods that debias only one modality.

desk verdict A well-built but under-reported fairness method: the architecture is genuinely new, and the empirical scope is strong, but the headline numbers need code, error bars, and a direct test of the debiasing mechanism before I would trust them. read the letter →

arxiv 2508.18886 v1 pith:TI2F6GZS submitted 2025-08-26 cs.CV

Yuexuan Xia , Benteng Ma , Jiang He , Zhiyong Wang , Qi Dou , Yong Xia This is my paper

classification cs.CV

keywords medicalfairnessvision-languagemodelspromptlearningattributedisentanglementdomaingeneralizationdistributionshiftCLIPprotectedattributes

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

Medical AI systems can be less accurate or fair for some demographic groups, and the gap grows when a model trained on one hospital's images is used at another. This paper argues that a major cause is that prior fairness methods remove bias from only the text side or only the image side of vision-language models, leaving the other side contaminated and misaligned. DualFairVL is a prompt-learning framework that debiases both sides at once: it extracts a sensitive-attribute direction from text embeddings, subtracts it from the disease embedding to form a debiased text anchor, and lets that anchor steer visual features through cross-attention and a hypernetwork while prototype losses separate sensitive from target information. The reported result is consistent fairness and accuracy gains over full fine-tuning and parameter-efficient baselines, with only 3.6M trainable parameters, across five in-distribution and three out-of-distribution medical imaging benchmarks. If this holds, fairness and robustness can be improved without large compute or full model access.

What carries the argument

The load-bearing mechanism is the text-guided attribute disentanglement loop. First, ridge-regularized linear projection (Eq. 5-6) estimates the component of the target text embedding aligned with the sensitive-attribute embedding and subtracts it, yielding text anchors that are approximately orthogonal. These anchors then serve as queries in a cross-attention layer that fuses text and visual features, and a hypernetwork converts branch- and layer-indexed attribute embeddings into local adapters that rewrite the visual prompts. Finally, a prototype-based loss models branch features as von Mises-Fisher distributions and enforces compactness within class-attribute groups while penalizing simil

What would settle it

After training DualFairVL, train a linear probe to predict the sensitive attribute from the final target-branch visual features. If the probe's accuracy remains high while DEOdds and DPD are low, the debiasing is not happening where claimed. A second test: use a sensitive attribute with no meaningful CLIP text token, such as race in chest X-ray; if fairness gains vanish or degrade relative to vision-only baselines, the text-anchor mechanism is not the cause of the improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that residual bias in CLIP-based medical diagnosis persists because prior methods debias a single modality, and that the fix is to debias both modalities while keeping them aligned. Concretely, DualFairVL constructs two text anchors by prompt tuning: one captures the sensitive attribute, the other the disease/utility attribute; the sensitive direction is removed from the disease embedding by ridge-regularized projection (Eq. 5-6). These anchors then condition a cross-attention module that fuses text with visual features, and a hypernetwork disentangles the fused features and emits instance-aware visual prompts that are injected into the image encoder. A prototype

Load-bearing premise

The load-bearing premise is that the sensitive attribute is linearly encoded in CLIP's text space by the same shared learnable prompts, so subtracting that direction from text anchors also removes sensitive information from medical image features even after a domain shift.

Editorial extensions

If this is right

Unimodal debiasing, whether text-only or vision-only, is predicted to leave residual bias that grows under domain shift, so future fairness methods should evaluate both modalities jointly.
A frozen CLIP with trainable prompts can match or beat full fine-tuning on fairness and accuracy in medical imaging, making large-model debiasing feasible with roughly 3.6M trainable parameters.
Fairness obtained by text-anchor alignment should transfer across imaging domains because the anchor carries domain-agnostic text semantics rather than vendor- or site-specific image statistics.
The same dual-branch recipe should apply to other protected attributes and medical modalities, with per-dataset tuning of a small set of hyperparameters.
If the central claim is correct, fairness under distribution shift does not require adversarial training or strong domain-pair assumptions; explicit disentanglement plus cross-modal alignment is sufficient.

Reading between the lines

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

The transfer assumption could be tested directly: for attributes with no clean textual counterpart in CLIP, such as race in chest X-ray, the method's advantage over vision-only baselines should shrink or vanish if the text anchor is doing the debiasing work.
The KDE visualization shows alignment qualitatively; an editorial next step would be to quantify residual sensitive information in the final target-branch features with a linear probe and to report calibration alongside AUC and parity metrics.
Because the framework is built on CLIP, it may not transfer unchanged to vision-language models with different text encoders or embedding geometries; replacing the backbone would clarify whether the text-anchor mechanism or the generic dual-branch losses carry the gains.
The ridge coefficient alpha=60 and prototype temperature phi=0.1 are tuned for a CLIP ViT-B/16 backbone; on new domains these hyperparameters may need re-estimation, and an annealing schedule for alpha could trade semantic preservation against debiasing more flexibly.

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

First thing to know: this is a real attempt, not a toy. DualFairVL combines text-anchor ridge projection, dual-branch prompts, cross-attention, a hypernetwork, and prototype losses into one CLIP fine-tuning pipeline. That combination is new in the medical-fairness literature as far as I can tell, and the evaluation is broad: eight datasets, four modalities, ID and OOD, against 13+ baselines. If the numbers are right, this is a useful parameter-efficient advance.

What the paper does well: it treats debiasing as a cross-modal alignment problem rather than a unimodal patch; the architecture is coherent on paper; the ablations are structured to isolate components. The consistency of the reported gains—DualFairVL wins or nearly wins on AUC and fairness across most rows—is more convincing than cherry-picked single-dataset results.

Soft spots, in increasing seriousness. First, reporting discipline: three-run averages with no standard deviations, especially for DEOdds and DPD on small binarized groups, makes the exact margins untrustworthy. Ablations in Table 4 are checkmarks with no numbers. The alpha=60 and phi=0.1 choices come from sweeps; the paper doesn't say explicitly these were validation-only, and prompt length 4 is selected the same way. None of this is fatal, but it blocks verification.

Second, the mechanism. The stress-test is on the right track: Eq. 5–6 removes a linear direction in text space, and the paper assumes that direction also carries the sensitive signal in image features. That is an assumption, not demonstrated. Fig. 6 is KDE only, not a quantitative residual-attribute measurement. I would not call this fatal—the method works or doesn't regardless—but it is a gap in the paper's explanation of why it works.

The self-stated limitation section only mentions CLIP-only scope. It omits these reporting and mechanism gaps. That is in-scope evidence and worth saying in review.

Who is this for? Anyone working on fairness in medical VLMs or parameter-efficient debiasing. It deserves a serious referee. I would send it out, with a request for code, per-dataset splits/hyperparameters, error bars, and a direct residual-attribute measurement—for example, a linear probe accuracy for the sensitive attribute on the debiased features. My own verdict is conditional: potential, but not citable until the empirical claims are verifiable. Reading group? Maybe, if the group enjoys discussing reporting standards as much as method novelty.

Referee Report

3 major / 4 minor

Summary. The paper proposes DualFairVL, a prompt-learning framework for vision-language models (CLIP) that jointly debiases text and image branches for medical image classification under in-distribution and out-of-distribution settings. The method constructs sensitive-attribute (SA) and target-attribute (TA) text anchors from shared learnable prompts, removes the SA-aligned component from the TA anchor via a ridge-regularized projection (Eqs. 5–6), uses these anchors to condition cross-attention on visual features (Eqs. 8–10), and adds a hypernetwork plus prototype-based losses (Eqs. 15–18) to disentangle and align visual features. Experiments cover five in-distribution datasets and three OOD scenarios, comparing with vision-only and vision-language prompt-tuning baselines. The central claim is that dual-modal debiasing—rather than unimodal debiasing—yields state-of-the-art fairness and accuracy with only 3.6M trainable parameters.

Significance. If the central mechanism is valid, the paper makes a useful contribution: it is one of the first attempts to combine text-side projection debiasing, visual disentanglement, and cross-modal alignment in a parameter-efficient prompt-tuning framework. The benchmark coverage is broad (eight datasets, four modalities, sex/race/skin-type attributes), the parameter count is modest, and the ablation structure explicitly isolates textual projection, cross-modal interaction, hypernetwork, and prototype regularization. The authors commit to releasing code, and the sensitivity analyses for α, φ, prompt length, and data ratios provide useful practical information. The main weakness is evidentiary: the paper does not directly measure residual sensitive information in the final debiased features, and the reported performance numbers lack variance estimates, which weakens the strength of the state-of-the-art claim.

major comments (3)

[§3.3–§3.4, Eqs. (5)–(10), Fig. 6] The load-bearing assumption is that the sensitive direction estimated from CLIP text prompts is also the direction carrying sensitive information in medical image features, and that removing it from the TA text anchor transfers to the visual branch through cross-attention. The citations [9,47] support text-space linearity for CLIP prompts, and [54] shows race can be predicted from images, but neither establishes that the image-side sensitive signal is linearly aligned with the text-prompt direction. Fig. 6 is only qualitative KDE evidence; it does not quantify residual attribute information in the final TA visual features. Please add a direct measurement, e.g., the accuracy of a linear probe for the sensitive attribute on the final TA features before/after the projection, or a comparison of the text SA direction with the sensitive direction extracted from image features. Without this, th
[§4.2.1, Tables 1–3, 5–7; §5.5–§5.6] All reported numbers are averages over three runs, but no standard deviations or confidence intervals are given, and no significance tests are reported. Given that some fairness differences between methods are small (e.g., Table 1, Harvard-GF3300 race: DPD 4.18 vs. 5.20–6.95), the state-of-the-art claim is not statistically grounded. In addition, the headline hyperparameters α=60, φ=0.1, and prompt length 4 appear to be selected by sweeps on the ID/OOD metrics reported in §5.5–§5.6 and Table 8. If these sweeps were performed on the same test sets used for the comparisons, the comparisons against baselines with default hyperparameters are optimistic. Please report mean±std, use a validation-based selection protocol, and state explicitly whether the hyperparameter sweeps were performed on held-out data.
[Eq. (5), §3.3.2] Equation (5) has a dimensional inconsistency: (z^SA_t)^T z^SA_t is a scalar, so adding αI (if I is the identity matrix) is not well-defined. If α is intended to be a scalar ridge coefficient, the equation should be written with a scalar α in the denominator. More substantively, for α>0 the operation subtracts only a scaled projection, so the residual vector is not exactly orthogonal to z^SA_t; the paper calls the anchors 'approximately orthogonal' but does not define in what sense. Since the orthogonality of the SA/TA anchors is one of the design principles advertised in the abstract and introduction, please clarify the exact mathematical operation and state whether the anchors are normalized, how orthogonality is measured, and why the SA anchor itself is not also projected.

minor comments (4)

[Abstract and Section 1] The method name is inconsistently typeset as 'DualF airVL' in the abstract and Section 1; there are also spacing artifacts such as 'F ull' in the abstract. Please proofread the text.
[§5.5] The text says 'The left panel of Table 7 examines the influence of φ' and 'The right panel of Table 7 analyzes α', but Table 7 contains dermatology domain-generalization results; the hyperparameter curves appear in Fig. 7. Please correct the cross-reference.
[Table 4 and Fig. 3] The ablation table lists configurations and parameter counts but no metric values; the radar charts in Fig. 3 are difficult to read quantitatively. Reporting the numeric AUC/DEOdds/DPD for each ablation setting would make the contribution of each component (Ldis, H, Attn, Proj) transparent and would strengthen the ablation claims in §4.4.
[Fig. 6] The KDE plots are described as showing 'substantially improved alignment', but no quantitative alignment score (e.g., mean cosine similarity between debiased visual features and TA text anchors) is reported. Adding such a number would make the visualization less subjective.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported fairness/accuracy gains are external empirical outcomes, not consequences of Eq. 5–6 by construction.

full rationale

The paper's central claims are empirical: DualFairVL achieves SOTA AUC/DEOdds/DPD on ID and OOD benchmarks (Tables 1–3, 5–7). The text-anchor projection in Eq. 5–6 indeed defines the debiased TA anchor by subtracting the SA-aligned component from the TA embedding, so the debiased text anchor is self-referential by construction. However, none of the reported fairness metrics is read off from that projection; DEOdds and DPD are computed on downstream visual predictions. The cross-attention, hypernetwork, and prototype losses add parameters and objectives not determined by Eq. 5–6. The paper cites prior work [9,47] for the text-space linearity premise, and includes one self-citation to [11] (Qi Dou co-author) in a positioning contrast, but that citation is not load-bearing for the main claims. The untested cross-modal linear-transfer assumption is an evidential/correctness limitation, not a circularity: no equation in the paper makes the visual fairness metric equal to the text projection. The limitations section does not acknowledge the absence of a direct residual-attribute measure in final TA features, but that omission concerns support, not circular reduction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central mechanism rests on three families of imported assumptions: (i) CLIP's text space linearly encodes sensitive attribute directions from prompts, imported from [9, 47] with no verification on these medical datasets; (ii) text-side subtraction transfers to visual debiasing through anchor-guided cross-attention; (iii) demographic labels, binarization, and fairness metrics follow the MedFair [1] protocol, including race labels transcribed from [54, 57]. The method introduces no new physical entities. The main results depend on tuned constants: alpha=60 (ridge projection), phi=0.1 (vMF temperature), lambda=0.1, delta=1, beta=0.5, prompt length 4, and per-dataset learning rates in [1e-3, 1e-2]; of these, alpha, phi, and prompt length were explicitly swept to optimize the reported fairness/accuracy trade-off (Sections 5.2.1, 5.5, 5.6).

free parameters (7)

alpha (ridge coefficient in text projection) = 60
Eq. 5 regularizes the projection of the TA text embedding onto the SA direction; Section 5.6 selects alpha=60 as 'optimal' for fairness, and all main results use it.
phi (temperature in vMF prototype losses) = 0.1
Eqs. 15-16; Section 5.5 reports the best fairness/AUC trade-off at phi=0.1, but Section 4.2.1 omits phi from the stated implementation details of the main runs.
lambda (weight on separability loss) = 0.1
Eq. 18; fixed in Section 3.5 as 'showing stable performance' without a reported sweep.
delta (weight on disentanglement loss) = 1
Eq. 19; fixed in Section 4.2.1 without a reported sweep.
beta (prototype EMA momentum) = 0.5
Eq. 17; fixed in Section 3.5 without a reported sweep.
Prompt length K = 4
Selected via sweep in Section 5.2.1 (Table 8) for the best balance of AUC and fairness; applies to both text and visual prompts.
Per-dataset learning rate = 1e-3 to 1e-2 (dataset-specific)
Section 4.2.1 reports a range rather than fixed values, implying per-dataset tuning that is not reported per table.

assumptions (4)

domain assumption CLIP text embeddings of prompts with sensitive tokens encode linearly separable, transferable sensitive directions.
Invoked in Section 3.3.1 citing [9, 47]; the whole debiasing pipeline, Eq. 5-6 and the anchor-guided attention in Section 3.4.1, assumes this holds for medical attributes such as sex, race, and skin type.
domain assumption Subtracting the SA component from text anchors removes sensitive information from visual medical features under domain shift.
The cross-attention in Eq. 8-10 steers visual features toward the debiased TA anchor; the paper offers only qualitative KDE evidence (Fig. 6) that this text-side subtraction actually debiases image features.
domain assumption Demographic labels and binarizations (sex, race, skin type) are correct and meaningful for the fairness metrics.
Section 4.1 follows MedFair [1]; race for CheXpert and MIMIC-CXR is transcribed from [54, 57], and label noise propagates into DEOdds and DPD.
standard math vMF prototypes with EMA updates (Eq. 17) provide stable branch prototypes on small binarized subgroups.
Directional statistics and EMA are standard [52], but the paper does not analyze prototype instability under small group sizes, which affects L_dis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Toward Robust Medical Fairness: Debiased Dual-Modal Alignment via Text-Guided Attribute-Disentangled Prompt Learning for Vision-Language Models." pith.science (2026). https://pith.science/paper/TI2F6GZS

@misc{pith2026250818886,
  author       = {Pith},
  title        = {Pith review of: Toward Robust Medical Fairness: Debiased Dual-Modal Alignment via Text-Guided Attribute-Disentangled Prompt Learning for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TI2F6GZS}},
  note         = {Machine review of arXiv:2508.18886}
}

read the original abstract

Ensuring fairness across demographic groups in medical diagnosis is essential for equitable healthcare, particularly under distribution shifts caused by variations in imaging equipment and clinical practice. Vision-language models (VLMs) exhibit strong generalization, and text prompts encode identity attributes, enabling explicit identification and removal of sensitive directions. However, existing debiasing approaches typically address vision and text modalities independently, leaving residual cross-modal misalignment and fairness gaps. To address this challenge, we propose DualFairVL, a multimodal prompt-learning framework that jointly debiases and aligns cross-modal representations. DualFairVL employs a parallel dual-branch architecture that separates sensitive and target attributes, enabling disentangled yet aligned representations across modalities. Approximately orthogonal text anchors are constructed via linear projections, guiding cross-attention mechanisms to produce fused features. A hypernetwork further disentangles attribute-related information and generates instance-aware visual prompts, which encode dual-modal cues for fairness and robustness. Prototype-based regularization is applied in the visual branch to enforce separation of sensitive features and strengthen alignment with textual anchors. Extensive experiments on eight medical imaging datasets across four modalities show that DualFairVL achieves state-of-the-art fairness and accuracy under both in- and out-of-distribution settings, outperforming full fine-tuning and parameter-efficient baselines with only 3.6M trainable parameters. Code will be released upon publication.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 57 canonical work pages

[54]

The Lancet Digital Health 4(6), 406–414 (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), 406–414 (2022)

work page 2022
[1]

In: ICLR (2023)

Zong, Y., Yang, Y., Hospedales, T.: Medfair: benchmarking fairness for medical imaging. In: ICLR (2023)

work page 2023
[2]

Transactions on machine learning research (2023)

Wang, H., Hong, J., Zhou, J., Wang, Z.: How robust is your fairness? evaluating and sustaining fairness under unseen distribu- tion shifts. Transactions on machine learning research (2023)

work page 2023
[3]

: An image is worth 16x16 words: Transformers for image recognition at scale

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. : An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)

work page 2021
[4]

: Learning transferable visual models from natural lan- guage supervision

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. : Learning transferable visual models from natural lan- guage supervision. In: ICML, pp. 8748–8763 (2021)

work page 2021
[5]

NeurIPS 37, 111318–111357 (2024)

Jin, R., Xu, Z., Zhong, Y., Yao, Q., QI, D., Zhou, S.K., Li, X.: Fairmedfm: fairness benchmarking for medical imaging founda- tion models. NeurIPS 37, 111318–111357 (2024)

work page 2024
[6]

In: CVPR, pp

Park, S., Byun, H.: Fair-vpt: Fair visual prompt tuning for image classification. In: CVPR, pp. 12268–12278 (2024)

work page 2024
[7]

In: ECCV, pp

Tian, Y., Wen, C., Shi, M., Afzal, M.M., Huang, H., Khan, M.O., Luo, Y., Fang, Y., Wang, M.: Fairdomain: Achieving fairness in cross-domain medical image segmentation and classification. In: ECCV, pp. 251–271 (2024)

work page 2024

Show all 59 references

[8]

In: ICLR (2024)

Dutt, R., Bohdal, O., Tsaftaris, S.A., 17 Hospedales, T.: Fairtune: Optimizing param- eter efficient fine tuning for fairness in medi- cal image analysis. In: ICLR (2024)

2024
[9]

arXiv preprint arXiv:2302.00070 (2023)

Chuang, C.-Y., Jampani, V., Li, Y., Torralba, A., Jegelka, S.: Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070 (2023)

2023 arXiv
[10]

In: ACL, pp

Liang, P.P., Li, I.M., Zheng, E., Lim, Y.C., Salakhutdinov, R., Morency, L.: Towards debiasing sentence representations. In: ACL, pp. 5502–5515 (2020)

2020
[11]

In: IPMI, pp

Deng, W., Zhong, Y., Dou, Q., Li, X.: On fair- ness of medical image classification with mul- tiple sensitive attributes via learning orthog- onal representations. In: IPMI, pp. 158–169 (2023)

2023
[12]

In: ICCV (2025)

Molahasani, M., Motamedi, A., Greenspan, M., Kim, I.-M., Etemad, A.: Prism: Reduc- ing spurious implicit biases in vision-language models with llm-guided embedding projec- tion. In: ICCV (2025)

2025
[13]

: Fairclip: Har- nessing fairness in vision-language learning

Luo, Y., Shi, M., Khan, M.O., Afzal, M.M., Huang, H., Yuan, S., Tian, Y., Song, L., Kouhana, A., Elze, T., et al. : Fairclip: Har- nessing fairness in vision-language learning. In: CVPR, pp. 12289–12301 (2024)

2024
[14]

In: AACL-IJCNLP, pp

Berg, H., Hall, S.M., Bhalgat, Y., Yang, W., Kirk, H.R., Shtedritski, A., Bain, M.: A prompt array keeps the bias away: Debias- ing vision-language models with adversarial learning. In: AACL-IJCNLP, pp. 806–822 (2022)

2022
[15]

In: CVPR, pp

Seth, A., Hemani, M., Agarwal, C.: Dear: Debiasing vision-language models with addi- tive residuals. In: CVPR, pp. 6820–6829 (2023)

2023
[16]

NeurIPS 35, 21682–21697 (2022)

Zhang, M., R´ e, C.: Contrastive adapters for foundation model group robustness. NeurIPS 35, 21682–21697 (2022)

2022
[17]

In: ICML (2024)

Zhang, J., Ma, X., Guo, S., Li, P., Xu, W., Tang, X., Hong, Z.: Amend to alignment: decoupled prompt tuning for mitigating spu- rious correlation in vision-language models. In: ICML (2024)

2024
[18]

In: CVPR, pp

You, C., Mint, Y., Dai, W., Sekhon, J.S., Staib, L., Duncan, J.S.: Calibrating multi- modal representations: A pursuit of group robustness without annotations. In: CVPR, pp. 26140–26150 (2024)

2024
[19]

In: ICML, pp

Yang, Y., Nushi, B., Palangi, H., Mirza- soleiman, B.: Mitigating spurious correlations in multi-modal models during fine-tuning. In: ICML, pp. 39365–39379 (2023)

2023
[20]

In: AMACL, pp

Jin, W., Cheng, Y., Shen, Y., Chen, W., Ren, X.: A good prompt is worth mil- lions of parameters: Low-resource prompt- based learning for vision-language models. In: AMACL, pp. 2763–2775 (2022)

2022
[21]

In: CVPR, pp

Wasim, S.T., Naseer, M., Khan, S., Khan, F.S., Shah, M.: Vita-clip: Video and text adaptive clip via multimodal prompting. In: CVPR, pp. 23034–23044 (2023)

2023
[22]

In: CVPR, pp

Huang, S., Gong, B., Pan, Y., Jiang, J., Lv, Y., Li, Y., Wang, D.: Vop: Text-video co-operative prompt tuning for cross-modal retrieval. In: CVPR, pp. 6565–6574 (2023)

2023
[23]

In: CVPR, pp

Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR, pp. 19113–19122 (2023)

2023
[24]

In: AAAI, vol

Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al.: Chexpert: A large chest radiograph dataset with uncertainty labels and expert compari- son. In: AAAI, vol. 33, pp. 590–597 (2019)

2019
[25]

Scientific data 6(1), 317 (2019)

Johnson, A.E., Pollard, T.J., Berkowitz, S.J., Greenbaum, N.R., Lungren, M.P., Deng, C.- y., Mark, R.G., Horng, S.: Mimic-cxr, a de- identified publicly available database of chest radiographs with free-text reports. Scientific data 6(1), 317 (2019)

2019
[26]

In: CVPR, pp

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 18 fitzpatrick 17k dataset. In: CVPR, pp. 1820– 1828 (2021)

2021
[27]

Scientific Data 9(1), 291 (2022)

Kovalyk, O., Morales-S´ anchez, J., Verd´ u- Monedero, R., Sell´ es-Navarro, I., Palaz´ on- Cabanes, A., Sancho-G´ omez, J.-L.: Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Scientific Data 9(1), 291 (2022)

2022
[28]

IEEE Transac- tions on Medical Imaging (2024)

Luo, Y., Tian, Y., Shi, M., Pasquale, L.R., Shen, L.Q., Zebardast, N., Elze, T., Wang, M.: Harvard glaucoma fairness: a retinal nerve disease dataset for fairness learning and fair identity normalization. IEEE Transac- tions on Medical Imaging (2024)

2024
[29]

Scientific data 5(1), 1–9 (2018)

Tschandl, P., Rosendahl, C., Kittler, H.: The ham10000 dataset, a large collection of multi- source dermatoscopic images of common pig- mented skin lesions. Scientific data 5(1), 1–9 (2018)

2018
[30]

In: MICCAI, pp

Puyol-Ant´ on, E., Ruijsink, B., Piechnik, S.K., Neubauer, S., Petersen, S.E., Razavi, R., King, A.P.: Fairness in cardiac mr image anal- ysis: an investigation of bias due to data imbalance in deep learning based segmenta- tion. In: MICCAI, pp. 413–423 (2021)

2021
[31]

In: CVPR, pp

Park, S., Lee, J., Lee, P., Hwang, S., Kim, D., Byun, H.: Fair contrastive learning for facial attribute classification. In: CVPR, pp. 10389– 10398 (2022)

2022
[32]

In: ICLR, pp

Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural net- works for group shifts: On the importance of regularization for worst-case generalization. In: ICLR, pp. 10389–10398 (2020)

2020
[33]

In: DSAA, pp

Raff, E., Sylvester, J.: Gradient reversal against discrimination: A fair neural network learning approach. In: DSAA, pp. 189–198 (2018)

2018
[34]

In: Machine Learning for Healthcare Conference, pp

Marcinkevics, R., Ozkan, E., Vogt, J.E.: Debiasing deep chest x-ray classifiers using intra-and post-processing methods. In: Machine Learning for Healthcare Conference, pp. 504–536 (2022)

2022
[35]

In: MICCAI, pp

Wu, Y., Zeng, D., Xu, X., Shi, Y., Hu, J.: Fairprune: Achieving fairness through prun- ing for dermatological disease diagnosis. In: MICCAI, pp. 743–753 (2022)

2022
[36]

In: AAAI, vol

Zhu, Y., Zhuang, F., Wang, D.: Align- ing domain-specific distribution and classifier for cross-domain classification from multiple sources. In: AAAI, vol. 33, pp. 5989–5996 (2019)

2019
[37]

In: CVPR, pp

Truong, T.-D., Le, N., Raj, B., Cothren, J., Luu, K.: Fredom: Fairness domain adaptation approach to semantic scene understanding. In: CVPR, pp. 19988–19997 (2023)

2023
[38]

In: W ACV, pp

Xiao, J., Bai, Y., Yuille, A., Zhou, Z.: Delv- ing into masked autoencoders for multi-label thorax disease classification. In: W ACV, pp. 3588–3600 (2023)

2023
[39]

Transactions on Machine Learning Research, 2835–8856 (2024)

Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features with- out supervision. Transactions on Machine Learning Research, 2835–8856 (2024)

2024
[40]

: Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching

MH Nguyen, D., Nguyen, H., Diep, N., Pham, T.N., Cao, T., Nguyen, B., Swoboda, P., Ho, N., Albarqouni, S., Xie, P., et al. : Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. NeurIPS 36, 27922–27950 (2023)

2023
[41]

In: EMNLP, vol

Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from unpaired medical images and text. In: EMNLP, vol. 2022, pp. 3876–3887 (2022)

2022
[42]

arXiv preprint arXiv:2303.00915 (2023)

Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., et al.: Biomedclip: a mul- timodal biomedical foundation model pre- trained from fifteen million scientific image- text pairs. arXiv preprint arXiv:2303.00915 (2023)

2023 arXiv
[43]

In: ECCV, pp

Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.-N.: 19 Visual prompt tuning. In: ECCV, pp. 709– 727 (2022)

2022
[44]

International Journal of Computer Vision 130(9), 2337–2348 (2022)

Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learn- ing to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337–2348 (2022)

2022
[45]

In: CVPR, pp

Zhou, Kaiyang, Yang, Jingkang, Loy, Chen Change, Liu, Ziwei: Conditional prompt learning for vision-language models. In: CVPR, pp. 16816–16825 (2022)

2022
[46]

IEEE Transac- tions on Medical Imaging (2024)

Wang, P., Zhang, H., Yuan, Y.: Mcpl: Multi- modal collaborative prompt learning for med- ical vision-language model. IEEE Transac- tions on Medical Imaging (2024)

2024
[47]

In: ICLR (2024)

Dehdashtian, S., Wang, L., Boddeti, V.N.: Fairerclip: Debiasing clip’s zero-shot predic- tions using functions in rkhss. In: ICLR (2024)

2024
[48]

In: ICML (2024)

Phan, H., Wilson, A.G., Lei, Q.: Controllable prompt tuning for balancing group distribu- tional robustness. In: ICML (2024)

2024
[49]

In: CVPR, pp

Zhang, H., Guo, Y., Kankanhalli, M.: Joint vision-language social bias removal for clip. In: CVPR, pp. 4246–4255 (2025)

2025
[50]

PNAS 113(27), 7391–7398 (2016)

Sch¨ olkopf, B., Hogg, D.W., Wang, D., Foreman-Mackey, D., Janzing, D., Simon- Gabriel, C.-J., Peters, J.: Modeling confound- ing by half-sibling regression. PNAS 113(27), 7391–7398 (2016)

2016
[51]

In: ECCV, pp

Sarhan, M.H., Navab, N., Eslami, A., Albar- qouni, S.: Fairness by learning orthogonal disentangled representations. In: ECCV, pp. 746–761 (2020). Springer

2020
[52]

Wiley Online Library, Hoboken, NJ, USA (2000)

Mardia, K.V., Jupp, P.E.: Directional Statis- tics. Wiley Online Library, Hoboken, NJ, USA (2000)

2000
[53]

European Journal of Cancer 119, 57–65 (2019)

Maron, R.C., Weichenthal, M., Utikal, J.S., Hekler, A., Berking, C., Hauschild, A., Enk, A.H., Haferkamp, S., Klode, J., Schaden- dorf, D., et al.: Systematic outperformance of 112 dermatologists in multiclass skin cancer image classification by convolutional neural networks. ...

2019
[55]

: Bcn20000: Der- moscopic lesions in the wild

Hern´ andez-P´ erez, C., Combalia, M., Podlip- nik, S., Codella, N.C., Rotemberg, V., Halpern, A.C., Reiter, O., Carrera, C., Bar- reiro, A., Helba, B., et al. : Bcn20000: Der- moscopic lesions in the wild. Scientific data 11(1), 641 (2024)

2024
[56]

Codella, N.C., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kit- tler, H., et al. : Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imag- ing (isbi), hoste...

2017
[57]

PhysioNet (2021)

Johnson, A., Bulgarelli, L., Pollard, T., Celi, L.A., Mark, R., Horng IV, S.: Mimic-iv-ed. PhysioNet (2021)

2021
[58]

In: ICML, pp

Agarwal, A., Beygelzimer, A., Dud ´ ık, M., Langford, J., Wallach, H.: A reductions approach to fair classification. In: ICML, pp. 60–69 (2018)

2018
[59]

In: ICML, pp

Agarwal, A., Dud ´ ık, M., Wu, Z.S.: Fair regression: Quantitative definitions and reduction-based algorithms. In: ICML, pp. 120–129 (2019) 20

2019

Pith tools

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

[1] [54]

The Lancet Digital Health 4(6), 406–414 (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), 406–414 (2022)

work page 2022

[2] [1]

In: ICLR (2023)

Zong, Y., Yang, Y., Hospedales, T.: Medfair: benchmarking fairness for medical imaging. In: ICLR (2023)

work page 2023

[3] [2]

Transactions on machine learning research (2023)

Wang, H., Hong, J., Zhou, J., Wang, Z.: How robust is your fairness? evaluating and sustaining fairness under unseen distribu- tion shifts. Transactions on machine learning research (2023)

work page 2023

[4] [3]

: An image is worth 16x16 words: Transformers for image recognition at scale

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. : An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)

work page 2021

[5] [4]

: Learning transferable visual models from natural lan- guage supervision

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. : Learning transferable visual models from natural lan- guage supervision. In: ICML, pp. 8748–8763 (2021)

work page 2021

[6] [5]

NeurIPS 37, 111318–111357 (2024)

Jin, R., Xu, Z., Zhong, Y., Yao, Q., QI, D., Zhou, S.K., Li, X.: Fairmedfm: fairness benchmarking for medical imaging founda- tion models. NeurIPS 37, 111318–111357 (2024)

work page 2024

[7] [6]

In: CVPR, pp

Park, S., Byun, H.: Fair-vpt: Fair visual prompt tuning for image classification. In: CVPR, pp. 12268–12278 (2024)

work page 2024

[8] [7]

In: ECCV, pp

Tian, Y., Wen, C., Shi, M., Afzal, M.M., Huang, H., Khan, M.O., Luo, Y., Fang, Y., Wang, M.: Fairdomain: Achieving fairness in cross-domain medical image segmentation and classification. In: ECCV, pp. 251–271 (2024)

work page 2024