Pith. sign in

REVIEW 5 major objections 5 minor 28 references

Decoupling Clinical and Class-Agnostic Features for Reliable Few-Shot Adaptation under Shift

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read DRiFt splits medical image and text features into clinical and spurious components, and this explicit decoupling lifts few-shot accuracy and robustness under distribution shift.

desk verdict A plausible few-shot medical VLM adaptation recipe whose core decoupling mechanism is underspecified and whose robustness claim rests on a single seed; worth refereeing, but only with heavy revisions. read the letter →

arxiv 2509.09397 v1 pith:2ME5JMIU submitted 2025-09-11 cs.CV

Umaima Rahman , Raza Imam , Mohammad Yaqub , Dwarikanath Mahapatra This is my paper · ORCID
classification cs.CV
keywords medicalvision-languagemodelsfew-shotlearningdistributionshiftfeaturedecouplingspuriouscorrelationslow-rankadaptationpromptout-of-distributiongeneralization
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 is trying to establish that medical vision-language models can be made more reliable in few-shot settings by explicitly separating clinically relevant, invariant information from task-agnostic spurious cues in both image and text. It proposes DRiFt, which combines low-rank adapters with learnable prompt tokens and uses three objectives: aligning invariant image-text features, neutralizing spurious features by pushing their class predictions toward uniform, and reducing statistical dependence between invariant and spurious features given the label. The headline numbers are a +11.4% Top-1 accuracy and +3.3% Macro-F1 improvement over prior prompt-based methods on in-distribution medical tasks, with competitive cross-dataset transfer. A sympathetic reader would care because this is a parameter-efficient path (no full fine-tuning) toward suppressing the spurious correlations that make clinical VLMs unsafe under domain shift.

What carries the argument

The feature-decoupling step is the machinery. A projection function phi_M acts on each raw embedding z_M (M for vision or text) and splits it into z_{M,u} (invariant, clinically relevant) and z_{M,s} (spurious). The three losses are then targeted: the contrastive loss aligns only invariant embeddings across modalities; the KL-divergence loss drives predictions from spurious embeddings toward a uniform class distribution so they carry no classification signal; and conditional independence losses reduce the mutual information between invariant and spurious embedding parts given the label. This separation is what lets the model learn clinical cues without being misled by domain-specific artifac

What would settle it

Take the same pipeline but replace the projection functions with a fixed random linear map that does not separate clinical content. If average Top-1 accuracy remains near 45.8%, the decoupling losses are not doing the claimed work; if it falls sharply, the projection choice is the critical ingredient. Alternatively, rerun the 16-shot setting with ten different random seeds and check whether the +11.4 average advantage over the baseline persists.

Watch

Extended reading notes

Core claim

The central claim is that spurious associations, such as a clothing shadow or incidental report wording, can be suppressed if each embedding is decomposed into invariant and spurious parts and the training losses act on those parts separately. DRiFt projects image and text embeddings into two subspaces, aligns only the invariant subspace with a contrastive loss, pushes the spurious subspace's predictions toward a uniform distribution, and penalizes conditional dependence between the two subspaces given the label. Trained with 16 samples per class, low-rank adapters, and learnable prompts, the framework reports an average Top-1 accuracy of 45.8% across eight medical tasks, compared with 34.4%

Load-bearing premise

The load-bearing premise is that the fixed projection functions used to split embeddings into invariant and spurious parts are meaningful; if those projections are arbitrary or poorly matched to the data, the three losses cannot actually enforce decoupling and the reported gains would come from another mechanism.

Editorial extensions

If this is right

  • If DRiFt's claim holds, then few-shot adaptation of medical vision-language models can be done without full fine-tuning, since only low-rank adapters and prompt tokens are updated.
  • Explicit spurious-feature neutralization can be added to any prompt-based method as a loss term, potentially improving out-of-distribution behavior.
  • The decoupling works on both modalities, so degradation from text-side spurious cues (e.g., incidental findings in reports) is also addressed.
  • Cross-dataset transfer suggests that models trained on one imaging protocol can preserve accuracy on another, which matters for multi-site deployment.
  • Using two prompt tokens rather than one nearly doubled average accuracy in the experiments, so prompt capacity itself appears to be part of the mechanism.

Reading between the lines

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

  • Because the projection functions that split embeddings are not trained and their architecture is not specified, the decoupling effect may be sensitive to how those projections are chosen; a direct test is to make them learnable or compare several fixed choices.
  • The few-shot numbers come from a single 16-sample-per-class split with no seeds or error bars, so the +11.4 point gap might not survive resampling; reporting multiple splits would separate mechanism from sampling luck.
  • The caption-generation step relies on an instruction-tuned captioning model; if caption quality degrades for rare diseases, the invariant alignment could reinforce wrong text-image associations, making caption validation a hidden dependency.
  • The same loss design could transfer to other high-stakes domains with known spurious cues, such as pathology or retinal OCT, where domain shift follows similar patterns, but this extrapolation is not tested in the paper.
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.

Referee Report

5 major / 5 minor

Summary. DRiFt is a parameter-efficient few-shot adaptation framework for medical vision-language models under distribution shift. It decomposes image and text embeddings into invariant (clinically relevant) and spurious components via unspecified projection functions, fine-tunes only LoRA adapters and prompt tokens, and trains with a combination of invariant contrastive alignment, KL-based spurious neutralization, and conditional-independence (HSIC) losses. The authors add InstructBLIP-generated captions to MedIMeta and evaluate on eight in-distribution tasks and cross-dataset transfer against MaPLe-IC and CoOp-OOD-IC. The abstract reports +11.4% Top-1 and +3.3% Macro-F1 over prior prompt-based methods while maintaining robustness across unseen datasets.

Significance. The intended contribution—explicit cross-modal decoupling of clinical and class-agnostic features with parameter-efficient tuning—is timely and, if valid, would be useful for few-shot medical VLM adaptation. The paper has concrete strengths: code is released, two prompt-based baselines are adapted to the image-caption setting, and ablations cover shots, prompt depth, prompt tokens, and backbone. However, the central decoupling mechanism is not actually specified, the HSIC estimator and loss weights are absent, and the reported in-distribution AUC is at chance for most tasks. As a result, the paper's core claims—both the mechanism and the robustness of the gains—are not currently established.

major comments (5)
  1. [Section 2.1] The decomposition is undefined. The paper states only z_{M i,u}=phi_M(zMi) and calls these 'two projection functions'; it neither defines the spurious component z_{M i,s}, nor specifies the architecture of phi_M (linear/MLP/random), nor includes phi_M in the trainable set Theta=theta∪gamma. If phi_M is frozen, the invariant/spurious split is fixed before training and Eqs. (1)-(3) cannot adaptively enforce decoupling; if phi_M is learned, it is omitted from the implementation. Either way, the advertised cause of the +11.4% gain is unverifiable.
  2. [Section 2.2] The conditional-independence losses L_conv and L_cont in Eq. (2) are written as HSIC terms, but no estimator, kernel, or conditioning procedure is given. The paper does not report how I(z_v;z_s|Y) is computed in practice, so Eq. (2) and the overall loss L in Eq. (3) are not reproducible. This is a load-bearing component of the claimed decoupling.
  3. [Section 2.2 and 3.1] The loss weights alpha and beta in Eq. (3) are never reported, and the implementation details do not describe a validation split for choosing them, the LoRA rank r, or the caption-generation thresholds. Ablations vary prompt depth and token count without indicating where selection occurs, raising a risk of test-set-based selection. Reproducibility and the interpretation of the reported gains are therefore limited.
  4. [Table 1] AUC is at or below chance for most tasks (bus 34.9, skinl_derm 49.5, derm 46.5, glaucoma 51.2, pneumonia 52.4, mammo_mass 45.9; average 49.1 with chance 50), directly contradicting the abstract's 'strong robustness' and 'reliable' claims. In addition, all results come from a single 16-sample-per-class draw with no error bars or significance tests. With few-shot sample sizes, the Top-1 improvements may be within run-to-run noise.
  5. [Table 2 and Section 3.2] The cross-dataset generalization claim is not supported. DRiFt is best only for skinl_derm (10.4 vs 6.4-6.5) and is worse on derm (17.8 vs 18.8-18.9) while roughly tying on glaucoma and fundus. No average or statistical comparison is reported, so 'maintaining strong robustness across unseen datasets' is unsubstantiated. Moreover, the method extends prior work [9], but [9] is not among the baselines, leaving the incremental contribution unclear.
minor comments (5)
  1. [Section 3.1] Typo: 'howvere' should be 'however'. Also 'spuriousrelationshipswhichcanbeverysensitive' is missing spaces.
  2. [Section 2.2] Notation is inconsistent: Eq. (2) defines L_conv and L_cont, but Eq. (3) introduces L_con = avg(L_conv, L_cont); the names should be unified.
  3. [Section 2.2] The reference to 'Figure 4(left)' appears in Section 2.2, before the experiments are described; this should be moved to the results discussion.
  4. [Table 1 / Section 3.2] The pneumonia row for DRiFt (37.5% Acc, 18.2% Macro-F1) is identical to CoOp-OOD-IC, yet the text says 'pneumonia accuracy improves'; the comparison baseline should be stated precisely.
  5. [Section 4a] The shot-ablation text says bus 'increases' from 50% to 42.3% (decrease); the sentence is internally contradictory and should be corrected.

Circularity Check

1 steps flagged · score 4.0 of 10

Load-bearing decoupling premise is self-cited to the authors' Dimple preprint; no equation-level circularity in the reported accuracy gains.

  1. self citation load bearing [Section 2.1, 'Decoupled Feature Representation']
    "To enhance generalization, we decompose each image and text embedding into invariant and spurious components [9]. Given an input embeddingzMi, two projection functions are defined byz Mi,u =ϕ M (zMi), whereM=v,t, withvandtrepresenting visual and textual embeddings, respectively. This decomposition ensures clinically meaningful features are preserved while mitigating domain-specific artifacts."

    The paper's central mechanism—explicit decoupling into invariant and spurious components—is introduced by citing [9], the authors' own prior arXiv preprint (Dimple), rather than by deriving or validating the projection functions. The projection functions phi_M are not defined, parameterized, or included in the trainable set Θ=θ∪γ, so the claimed guarantee that the decomposition preserves clinical features and mitigates artifacts rests entirely on the self-cited prior work. This is load-bearing: if [9] is not accepted, the losses L_ce^u, L_sp^r, and L_con in Section 2.2 have no specified spurious component to act on, so the mechanism behind the +11.4% improvement is not independently established in this paper.

full rationale

The central accuracy claims in Table 1 are arithmetic differences against external baselines (MaPLe-IC and CoOp-OOD-IC), so the headline +11.4%/+3.3% is not circular. No equation-level reduction of a prediction to a fitted parameter was found; the losses in Eq. (3) are not equivalent to the reported metrics by construction. However, the mechanism said to produce the gain—the invariant/spurious split—is imported from the authors' prior work [9] (Dimple) and the projection functions phi_M are never specified or included in the trainable set Θ=θ∪γ. The paper's only justification for 'This decomposition ensures clinically meaningful features are preserved' is the self-citation to [9], which is an unreviewed preprint by the same authors. This makes the decoupling premise load-bearing self-citation rather than independent derivation. The underspecification of phi_M is also a reproducibility gap, but it is not itself circular; it only reinforces that the claimed mechanism rests on [9]. Hyperparameter choices (n=2, J=3) are reported and ablated on test data, a mild selection feedback, but they are not presented as predictions derived from the model. Overall: some self-citation load-bearing; the main empirical comparison is independent.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several unstated assumptions: the base VLM is not named; the projection functions are undefined and untrained; the caption filtering is manual; and the regularization weights are not reported. These are load-bearing for the method to work, and none are independently validated.

free parameters (6)
  • alpha (spurious neutralization weight)
    In Eq. (3), the weight for L^sp_r is never reported; it is a hyperparameter likely tuned per dataset.
  • beta (independence loss weight)
    In Eq. (3), the weight for L_con is never reported; value unknown.
  • Number of prompt tokens = 2
    Chosen in ablation because n=2 improves average accuracy from 27.9 to 45.8 on the test set (Sec. 4b).
  • Prompt depth (J) = 3
    Chosen because deeper prompts degrade performance; J=3 used in main results, ablation shows steep declines at J=9/12 (Sec. 4c).
  • LoRA rank r
    Not reported; affects capacity of adapters.
  • Caption generation minimum image threshold = 100+
    Selected by hand to ensure diversity; not justified quantitatively.
assumptions (5)
  • domain assumption The pre-trained ViT-B/16-based vision-language model provides sufficient transferable visual and textual representations.
    Used as backbone; exact model not specified, but all evaluations rely on it.
  • domain assumption InstructBLIP-generated captions, after manual filtering, are clinically relevant and provide the supervision signal for decoupling.
    Section 2.1 caption generation; filtering criteria not quantified or validated by external clinical experts.
  • ad hoc to paper A fixed projection phi_M can separate invariant from spurious features without being trained or verified.
    Section 2.1: phi_M is defined but excluded from trainable parameter set, so the separation is assumed valid.
  • domain assumption HSIC-based conditional independence measures enforce the intended invariance in both modalities.
    Section 2.2 Eq. (2); no estimator details or justification given.
  • domain assumption The MedIMeta labels and task splits are reliable and representative.
    All in-distribution results depend on these datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupling Clinical and Class-Agnostic Features for Reliable Few-Shot Adaptation under Shift." pith.science (2026). https://pith.science/paper/2ME5JMIU

@misc{pith2026250909397,
  author       = {Pith},
  title        = {Pith review of: Decoupling Clinical and Class-Agnostic Features for Reliable Few-Shot Adaptation under Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ME5JMIU}},
  note         = {Machine review of arXiv:2509.09397}
}
read the original abstract

Medical vision-language models (VLMs) offer promise for clinical decision support, yet their reliability under distribution shifts remains a major concern for safe deployment. These models often learn task-agnostic correlations due to variability in imaging protocols and free-text reports, limiting their generalizability and increasing the risk of failure in real-world settings. We propose DRiFt, a structured feature decoupling framework that explicitly separates clinically relevant signals from task-agnostic noise using parameter-efficient tuning (LoRA) and learnable prompt tokens. To enhance cross-modal alignment and reduce uncertainty, we curate high-quality, clinically grounded image-text pairs by generating captions for a diverse medical dataset. Our approach improves in-distribution performance by +11.4% Top-1 accuracy and +3.3% Macro-F1 over prior prompt-based methods, while maintaining strong robustness across unseen datasets. Ablation studies reveal that disentangling task-relevant features and careful alignment significantly enhance model generalization and reduce unpredictable behavior under domain shift. These insights contribute toward building safer, more trustworthy VLMs for clinical use. The code is available at https://github.com/rumaima/DRiFt.

Figures

Figures reproduced from arXiv: 2509.09397 by the authors.

Figure 1
Figure 1. Impact of spurious correlations in medical imaging. (a) Conceptual illustration of multi-modal decoupling. (b) Shows how spurious features can dominate embeddings, leading to misclassification, critical in healthcare contexts. scheduling, can mislead a model if these spurious elements become incorrectly associated with diagnostic findings (e.g., lung infiltrates) [5]. Conventional fine-tuning strategies such as [6,7… view at source ↗
Figure 2
Figure 2. DRiFt: Our proposed feature disentanglement framework. (a) Captions are generated for images from the MedIMeta dataset. (b) Multi-modal feature decoupling is applied using LoRA fine-tuning and learnable prompts to focus on invariant features. Transformer layers, LoRA injects low-rank matrices into the original weight ma￾trix W0 ∈ R dout×din , decomposing it as, Wθ = W0 + αAB, where A ∈ R dout×r , B ∈ R r×din , and θ… view at source ↗
Figure 3
Figure 3. (left) In-distribution performance comparison across MaPLe-IC, CoOp-OOD￾IC, and DRiFt, with ours achieving the highest average Top-1 accuracy across 8 MedIMeta tasks. (middle) Impact of loss components on in-distribution performance, showing Top-1 accuracy variations for the mammo_mass and glaucoma tasks. (right) Cross-dataset evaluation of loss components, highlighting Top-1 accuracy changes. on textual prompts, wh… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: (left) Comparison of cross-dataset generalization for MaPLe-IC, CoOp-OOD￾IC, and DRiFt (Ours). (middle) Comparison of ViT-B/16 and ViT-B/32 backbones on MedIMeta datasets, showing that ViT-B/16 achieves higher average Top-1 accuracy. (right) Increasing the number of pr…
Figure 5
Figure 5. Figure 5: (left) Influence of the number of training shots (1, 2, 4, 8, 16), showing that additional shots improve performance for some datasets while others remain stable. (right) Impact of varying prompt depth (J) on Top-1 accuracy across datasets, high￾lighting performance de…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 5 linked inside Pith

  1. [9]

    Dimple– disentangled multi-modal prompt learning: Enhancing out-of-distribution align- ment with invariant and spurious feature separation.arXiv preprint arXiv:2506.21237, 2025

    Umaima Rahman, Mohammad Yaqub, and Dwarikanath Mahapatra. Dimple– disentangled multi-modal prompt learning: Enhancing out-of-distribution align- ment with invariant and spurious feature separation.arXiv preprint arXiv:2506.21237, 2025

    arXiv 2025

  2. [1]

    Which clinical decisions benefit from automation? a task complexity approach.International journal of medical infor- matics, 70(2-3):309–316, 2003

    Vitali Sintchenko and Enrico W Coiera. Which clinical decisions benefit from automation? a task complexity approach.International journal of medical infor- matics, 70(2-3):309–316, 2003

    2003

  3. [2]

    Mrishift: Disentangled repre- sentation learning for 3d mri lesion segmentation under distributional shifts

    Umaima Rahman, Guangyi Chen, and Kun Zhang. Mrishift: Disentangled repre- sentation learning for 3d mri lesion segmentation under distributional shifts. In 2024 12th European Workshop on Visual Information Processing (EUVIP), pages 1–6. IEEE, 2024

    2024

  4. [3]

    Heterogeneity in systematic reviews of medical imaging diagnos- tic test accuracy studies: a systematic review.JAMA Network Open, 7(2):e240649– e240649, 2024

    Samuel J White, Qi Sheng Phua, Lucy Lu, Kaspar L Yaxley, Matthew DF Mcinnes, and Minh-Son To. Heterogeneity in systematic reviews of medical imaging diagnos- tic test accuracy studies: a systematic review.JAMA Network Open, 7(2):e240649– e240649, 2024

    2024

  5. [4]

    Reducing reliance on spurious features in medical image clas- sification with spatial specificity

    Khaled Saab, Sarah Hooper, Mayee Chen, Michael Zhang, Daniel Rubin, and Christopher Ré. Reducing reliance on spurious features in medical image clas- sification with spatial specificity. InMachine Learning for Healthcare Conference, pages 760–784. PMLR, 2022

    2022

  6. [5]

    When more is less: Incorporating additional datasets can hurt performance by introducing spuri- ous correlations

    Rhys Compton, Lily Zhang, Aahlad Puli, and Rajesh Ranganath. When more is less: Incorporating additional datasets can hurt performance by introducing spuri- ous correlations. InMachine Learning for Healthcare Conference, pages 110–127. PMLR, 2023

    2023

  7. [6]

    Maple: Multi-modal prompt learning

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113–19122, 2023

    2023

  8. [7]

    Conditional prompt learning for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816–16825, 2022

    2022

Show all 28 references
  1. [8]

    Umaima Rahman, Raza Imam, Mohammad Yaqub, Boulbaba Ben Amor, and Dwarikanath Mahapatra. Can language-guided unsupervised adaptation improve medical image classification using unpaired images and texts? In2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), page...

    2025

  2. [10]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    2022

  3. [11]

    Test- time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models.arXiv preprint arXiv:2407.15913, 2024

    Raza Imam, Hanan Gani, Muhammad Huzaifa, and Karthik Nandakumar. Test- time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models.arXiv preprint arXiv:2407.15913, 2024

    2024 arXiv

  4. [12]

    A compre- hensive and easy-to-use multi-domain multi-task medical imaging meta-dataset (medimeta).arXiv preprint arXiv:2404.16000, 2024

    Stefano Woerner, Arthur Jaques, and Christian F Baumgartner. A compre- hensive and easy-to-use multi-domain multi-task medical imaging meta-dataset (medimeta).arXiv preprint arXiv:2404.16000, 2024. 10 Rahman et al

    2024

  5. [13]

    Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023

    2023

  6. [14]

    Domain generalization on medical imaging classification us- ing episodic training with task augmentation.Computers in biology and medicine, 141:105144, 2022

    Chenxin Li, Xin Lin, Yijin Mao, Wei Lin, Qi Qi, Xinghao Ding, Yue Huang, Dong Liang, and Yizhou Yu. Domain generalization on medical imaging classification us- ing episodic training with task augmentation.Computers in biology and medicine, 141:105144, 2022

    2022

  7. [15]

    Domain generalization for medical image analysis: A review.Proceedings of the IEEE, 2024

    Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A Mazurowski, and Heung-Il Suk. Domain generalization for medical image analysis: A review.Proceedings of the IEEE, 2024

    2024

  8. [16]

    Evaluation of do- main generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine.Scientific reports, 12(1):2726, 2022

    Lin Lawrence Guo, Stephen R Pfohl, Jason Fries, Alistair EW Johnson, Jose Posada, Catherine Aftandilian, Nigam Shah, and Lillian Sung. Evaluation of do- main generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine.Scientific r...

    2022

  9. [17]

    Few-shot adaptation of medical vision-language models

    Fereshteh Shakeri, Yunshi Huang, Julio Silva-Rodríguez, Houda Bahig, An Tang, Jose Dolz, and Ismail Ben Ayed. Few-shot adaptation of medical vision-language models. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, pages 553–563. Springer, 2024

    2024

  10. [18]

    Clipath: Fine-tune clip with visual feature fusion for pathology image analysis towards minimizing data collection efforts

    Zhengfeng Lai, Zhuoheng Li, Luca Cerny Oliveira, Joohi Chauhan, Brittany N Dugger, and Chen-Nee Chuah. Clipath: Fine-tune clip with visual feature fusion for pathology image analysis towards minimizing data collection efforts. InProceedings of the IEEE/CVF International Confer...

    2023

  11. [19]

    Noise is an efficient learner for zero-shot vision- language models.arXiv preprint arXiv:2502.06019, 2025

    Raza Imam, Asif Hanif, Jian Zhang, Khaled Waleed Dawoud, Yova Kementched- jhieva, and Mohammad Yaqub. Noise is an efficient learner for zero-shot vision- language models.arXiv preprint arXiv:2502.06019, 2025

    2025 arXiv

  12. [20]

    Meduna: Language guided unsupervised adaptation of vision-language models for medical image classification.arXiv preprint arXiv:2409.02729, 2024

    Umaima Rahman, Raza Imam, Dwarikanath Mahapatra, and Boulbaba Ben Amor. Meduna: Language guided unsupervised adaptation of vision-language models for medical image classification.arXiv preprint arXiv:2409.02729, 2024

    2024 arXiv

  13. [21]

    Learning disentangled representations in the imaging domain.Medical Image Analysis, 80:102516, 2022

    Xiao Liu, Pedro Sanchez, Spyridon Thermos, Alison Q O’Neil, and Sotirios A Tsaftaris. Learning disentangled representations in the imaging domain.Medical Image Analysis, 80:102516, 2022

    2022

  14. [22]

    Domain adaptation and generalization of func- tional medical data: A systematic survey of brain data.ACM Computing Surveys, 56(10):1–39, 2024

    Gita Sarafraz, Armin Behnamnia, Mehran Hosseinzadeh, Ali Balapour, Amin Meghrazi, and Hamid R Rabiee. Domain adaptation and generalization of func- tional medical data: A systematic survey of brain data.ACM Computing Surveys, 56(10):1–39, 2024

    2024

  15. [23]

    Shenzhen hospital chest x-ray (cxr) set

    Shenzhen Hospital. Shenzhen hospital chest x-ray (cxr) set. https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray- Datasets/Shenzhen-Hospital-CXR-Set/index.html

  16. [24]

    Montgomery county chest x-ray (cxr) set

    Montgomery County. Montgomery county chest x-ray (cxr) set. https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray- Datasets/Montgomery-County-CXR-Set/MontgomerySet/index.html

  17. [25]

    Indian diabetic retinopathy image dataset (idrid): a database for diabetic retinopathy screening research.Data, 3(3):25, 2018

    Prasanna Porwal, Samiksha Pachade, Ravi Kamble, Manesh Kokare, Girish Desh- mukh, Vivek Sahasrabuddhe, and Fabrice Meriaudeau. Indian diabetic retinopathy image dataset (idrid): a database for diabetic retinopathy screening research.Data, 3(3):25, 2018

    2018

  18. [26]

    Isic archive: International skin imaging collaboration dataset

    International Skin Imaging Collaboration. Isic archive: International skin imaging collaboration dataset. https://www.isic-archive.com/

  19. [27]

    On the robustness of medical vision-language models: Are they truly generalizable?arXiv preprint arXiv:2505.15425, 2025

    Raza Imam, Rufael Marew, and Mohammad Yaqub. On the robustness of medical vision-language models: Are they truly generalizable?arXiv preprint arXiv:2505.15425, 2025. Decoupling Clinical and Class-Agnostic Features under Shift 11

    2025 arXiv

  20. [28]

    Amend to alignment: Decoupled prompt tuning for mitigating spuri- ous correlation in vision-language models

    Jie Zhang, Xiaosong Ma, Song Guo, Peng Li, Wenchao Xu, Xueyang Tang, and Zicong Hong. Amend to alignment: Decoupled prompt tuning for mitigating spuri- ous correlation in vision-language models. InForty-first International Conference on Machine Learning

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.