Pith. sign in

REVIEW 4 major objections 4 minor 40 references

UOPSL: Unpaired OCT Predilection Sites Learning for Fundus Image Diagnosis Augmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read UOPSL claims a learned predilection matrix can carry OCT spatial knowledge into fundus-only eye disease classification, outperforming benchmarks on nine datasets.

desk verdict Worth a look and worth a reject: the predilection-matrix idea is plausible and the experiments are extensive, but the proof that allows dropping OCT at inference is invalid, the key baseline is missing, and the domain-shift table is wrong. read the letter →

arxiv 2509.08624 v1 pith:RKUB6MPL submitted 2025-09-10 cs.CV cs.AI

classification cs.CVcs.AI
keywords UnpairedImagesPredilectionSitesRetinalDiagnosisFundusphotographyOCTContrastivelearningMultimodalmedicalimagingFoundationmodels
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 proposes that eye-disease classification from fundus photographs alone can inherit the spatial localization knowledge of OCT scans without ever needing a paired OCT image at test time. It pretrains on large unpaired collections of fundus photos and OCT B-scans that share only disease-category labels and expert-written text descriptions, and during this pretraining it learns a 'predilection sites matrix' inside the OCT latent space. At inference the OCT encoder is dropped; the matrix, fused with disease text through cross-attention, stands in for the missing OCT modality and guides where the model should look in fundus images. The authors report that this scheme beats existing retinal foundation models and single-modality baselines on transfer, domain-shift, and unseen-category evaluations across nine datasets and 28 disease categories.

What carries the argument

The predilection sites matrix P — a learnable weight array passed through a sigmoid so its entries lie in [0,1] — is the core object. It is multiplied element-wise (Hadamard product) with the OCT latent feature, and the result is fed with disease text into a cross-attention module; two CLIP-style contrastive losses align the fundus embedding with the reweighted OCT features and with the text. The matrix's role is to record, in latent space, the spatial locations where each disease tends to appear in OCT, so that at inference it can substitute for the OCT encoder entirely.

What would settle it

Regression test: on held-out OCT B-scans, fit F_O = aP + noise (or a linear map from P to F_O) and measure explained variance; if it is low or the dimensionalities mismatch (F_O is pooled to d per Eq. 1 while P is treated as an n×d spatial map), the substitution premise fails. Permutation test: shuffle the entries of P at inference; if downstream accuracy is unchanged, the matrix carries no spatial predilection information.

Watch

Extended reading notes

Core claim

UOPSL's central claim is that a single learnable matrix P, trained with CLIP-style contrastive losses on unpaired OCT and fundus images linked by disease text, captures where retinal diseases prefer to manifest. During pretraining, P is element-wise multiplied with OCT latent features; cross-attention ties the reweighted features to disease text, and two contrastive losses align fundus features with both the reweighted OCT features and the text. After pretraining, OCT input is removed: the pretrained P, fused through cross-attention with text prompts for the candidate diseases, acts as the spatial prior for fundus classification and zero-shot inference. The paper reports consistent improveme

Load-bearing premise

The load-bearing premise is that after sufficient pretraining the OCT image's latent feature is approximately a scaled copy of the learned predilection matrix plus small noise, so the matrix can replace OCT at inference; the paper asserts this relation rather than demonstrating it.

Editorial extensions

If this is right

  • A fundus-only pipeline can benefit from OCT-derived lesion-location cues without paying for paired OCT acquisition at test time.
  • The learned matrix transfers to held-out datasets and to disease categories never seen during training, suggesting the priors generalize beyond the training label set.
  • The ablation shows that the learned matrix outperforms both randomly selected OCT images and category-averaged OCT latents as stand-ins for the missing modality at fine-tuning.
  • Because text prompts are generated per disease category, the pretraining needs no image-level annotations beyond category labels.
  • The method is compatible with standard vision-language encoders and linear probing, so deployment adds a fixed matrix and a cross-attention layer rather than a full OCT model.

Reading between the lines

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

  • If P truly encodes predilection sites, it should be directly visualizable as a heatmap; comparing its high-weight regions to clinician-annotated lesion locations for diseases such as AMD or diabetic retinopathy would be a concrete external check.
  • The same unpaired-distillation recipe could transfer to other modality pairs with a cheap 2D image and an expensive tomographic volume, such as ultrasound to CT or dermoscopy to confocal microscopy.
  • A testable extension is to randomly permute the rows or entries of P at inference: if the gain is from spatial content, accuracy should degrade, whereas if P functions as a category-level bias it would not.
  • The proof's premise that OCT latent features equal a scaled P plus small noise could be checked directly on held-out OCT scans by linear regression; this would turn the theoretical justification into an empirical quantity.
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

4 major / 4 minor

Summary. The paper proposes UOPSL, an unpaired multimodal framework that learns a disease predilection matrix P from OCT latent features during pretraining, and then uses P (without OCT input) together with text descriptions to enhance fundus-image classification at fine-tuning and inference. The method is evaluated on nine fundus datasets across 28 categories and is reported to outperform several vision-language and retinal foundation models. The paper also contains a theoretical section (III-E–H) intended to justify replacing OCT features with P.

Significance. If the central claim holds, UOPSL would offer a practical way to transfer OCT-derived spatial knowledge to fundus-only diagnosis without requiring paired multimodal data at test time. The empirical scope is substantial: pretraining on ~80k OCT and ~84k fundus images, and evaluation across multiple public datasets. The ablation in Table V provides some evidence that the learned P is more useful than random or averaged OCT features. However, the theoretical justification is flawed in multiple ways, and some experimental claims are not supported by the presented results. The significance of the work therefore depends on whether the method can be re-validated after correcting these issues.

major comments (4)
  1. [§III-F–H, Eqs. (12)–(16)] The proof of rank preservation is invalid for three reasons: (i) the proof uses n×d matrices while the architecture uses d-dimensional vectors (Eq. 1 vs. §III-F); (ii) the hypothesis F_O = αP + B is assumed without any training mechanism enforcing it, making the argument circular; (iii) even under the hypothesis, monotonicity of squaring does not survive a learned linear projection and dot product, so the rank equivalence in Eq. (16) is false. The manuscript's claim of mathematical validation is therefore not supported.
  2. [§IV-B2, Table IV] Table IV, labeled as domain-shift generalization, contains the same dataset rows as Table III (JSIEC, Retina, RFMID, ODIR), whereas Section IV-A3 specifies that domain-shift testing should be on IDRID, APTOS2019, PAPILA, and Retina. This discrepancy means the claimed zero-shot domain-shift performance is not empirically demonstrated in the current manuscript.
  3. [§III-B, Eq. (2)] The paper claims a 'disease-specific' predilection matrix, but P in Eq. (2) is a single shared vector parameter. It is not conditioned on the disease label or text, so it cannot capture different predilection sites for different diseases. This is also inconsistent with the proof's assumption that all OCT features align with the same P. Please clarify the mechanism by which a single P encodes disease-specific spatial information, or redesign the model accordingly.
  4. [§IV-B4, Table V] The ablation in Table V shows that P outperforms random/averaged OCT features, but it does not isolate the contribution of P relative to the text-based contrastive learning. A control trained with only the fundus-text contrastive loss (Eq. 6) or with P removed at inference is needed to support the claim that OCT-derived spatial priors are the source of the gains. Please add such a comparison.
minor comments (4)
  1. [§IV-A4] The text says 'JEIEC114' but the dataset is JSIEC (also appearing in Table I). Please fix this typo.
  2. [References] Reference [6] is cited for CLIP, but the entry is for 'CLIP-art' (Conde and Turgutlu). The original CLIP model is referenced later as [39]. Please correct the citation in the introduction and related work.
  3. [Eq. (4)] The attention formulation yields a scalar attention weight because Q, K, V are d-dimensional vectors. The notation with softmax over a single value is unusual; clarify whether this is intended or whether multi-head attention over spatial locations is used.
  4. [Throughout] The notation alternates between P and P̂; please use a consistent symbol for the predilection matrix, especially in the proof section.

Circularity Check

1 steps flagged · score 7.0 of 10

The proof that P can replace OCT at inference assumes F_O = αP̂ and concludes P̂ alone suffices, so the central claim is the hypothesis; the rank-preservation argument also ignores the linear projection.

  1. self definitional [Section III-F/H, Eqs. (12)-(16)]
    "Our central hypothesis is that when the model is sufficiently well-trained, the information encoded in the OCT feature map FO is highly correlated with the spatial disease priors in P̂. That is, under ideal, well-trained conditions, FO = αP̂. ... Based on our analysis, substituting P̂ for the unavailable OCT features is a principled approach."

    The justification for removing the OCT encoder is the assumption itself. The section first posits FO = αP̂ + B, then substitutes this into the ideal attention (Eq. 12) and uses the substitution to conclude rank(a^(OP)) = rank(a^(P)) (Eq. 16). But this conclusion—that P̂ carries the same disease-location information as FO—is exactly the premise. Nothing in the contrastive losses (Eqs. 5-7) enforces FO ∝ P̂, and no experiment validates the proportionality. Thus the central claim that P can replace OCT at inference is not derived; it is the unproven hypothesis restated as a conclusion.

full rationale

The central claim is that a learned predilection matrix P suffices when OCT input is unavailable. The only mathematical support is the 'Proof of Consistency' in Section III-E-H, which explicitly assumes FO = αP̂ + B as its 'central hypothesis' and then concludes that substituting P̂ for FO preserves the attention ranking. That is a tautology: the proof's conclusion is its premise. It is not rescued by the rank-preservation argument because (i) the implemented encoder produces d-dimensional embeddings (Eq. 1: FO ∈ R^d, P ∈ R^d), whereas the proof postulates n×d spatial maps, so the objects in the proof are not the objects in the model; and (ii) even for n×d, elementwise monotonicity of squaring does not imply that the dot products (P̂_i² W_Q)·(F_T W_K) keep the same order after the linear map W_Q and the fixed key. Separately, Table IV duplicates Table III instead of reporting the promised domain-shift datasets, so the empirical demonstration of the substitution is also absent. There is no significant self-citation problem: the encoders (MetaCLIP, BioMed-BERT) and baselines are external, and the benchmark comparisons are independent. The circularity is localized to the theoretical derivation of the core OCT-free inference mechanism, so the score is 7 rather than 0.

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

The central claim rests on a fitted predilection matrix P, hand-set loss weights, and several unverified assumptions: that frozen CLIP features transfer to retinal modalities, that generated text captures predilection sites, and that after pretraining OCT features are approximately a scaled version of P. The latter is asserted without evidence and is the crux of the inference-time substitution.

free parameters (3)
  • predilection matrix P = learned vector of dimension d (CLIP embedding dim)
    Central learnable parameter trained on unpaired OCT data; used at inference as a proxy for OCT features. It is fitted to pretraining data, not derived from first principles.
  • loss weights λ1 and λ2 = 0.4 and 0.6
    Empirically set in Section III-C without justification; they control the balance between fundus-OCT and fundus-text contrastive losses.
  • temperature τ = not specified
    The contrastive temperature in Eqs (5) and (6) is a hyperparameter that affects feature alignment; its value is not reported.
assumptions (4)
  • domain assumption Frozen CLIP encoders (MetaCLIP ViT-B/16 and BioMed-BERT) transfer to retinal fundus and OCT domains without domain adaptation
    The method relies on a fixed CLIP vision encoder for both modalities and a fixed text encoder; no encoder fine-tuning is performed before contrastive learning on medical images.
  • ad hoc to paper ChatGPT-generated text descriptions accurately encode disease predilection sites
    Ten generated descriptions per disease are used as the text bridge; there is no validation that these texts capture the spatial prior information needed for the method to work.
  • ad hoc to paper After pretraining, OCT latent features are approximately a scaled version of the predilection matrix (FO ≈ αP)
    Stated in Section III-F as the central hypothesis, but asserted without evidence. It is the crux of the inference-time substitution and is never independently verified.
  • domain assumption The CLIP embedding used for OCT contains spatial structure that can be reweighted by a vector P
    The architecture uses pooled d-dimensional features (Eq 1), yet the proof in III-F assumes an n×d spatial feature map. This inconsistency suggests the spatial interpretation may not hold for the actual implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UOPSL: Unpaired OCT Predilection Sites Learning for Fundus Image Diagnosis Augmentation." pith.science (2026). https://pith.science/paper/RKUB6MPL

@misc{pith2026250908624,
  author       = {Pith},
  title        = {Pith review of: UOPSL: Unpaired OCT Predilection Sites Learning for Fundus Image Diagnosis Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RKUB6MPL}},
  note         = {Machine review of arXiv:2509.08624}
}
read the original abstract

Significant advancements in AI-driven multimodal medical image diagnosis have led to substantial improvements in ophthalmic disease identification in recent years. However, acquiring paired multimodal ophthalmic images remains prohibitively expensive. While fundus photography is simple and cost-effective, the limited availability of OCT data and inherent modality imbalance hinder further progress. Conventional approaches that rely solely on fundus or textual features often fail to capture fine-grained spatial information, as each imaging modality provides distinct cues about lesion predilection sites. In this study, we propose a novel unpaired multimodal framework \UOPSL that utilizes extensive OCT-derived spatial priors to dynamically identify predilection sites, enhancing fundus image-based disease recognition. Our approach bridges unpaired fundus and OCTs via extended disease text descriptions. Initially, we employ contrastive learning on a large corpus of unpaired OCT and fundus images while simultaneously learning the predilection sites matrix in the OCT latent space. Through extensive optimization, this matrix captures lesion localization patterns within the OCT feature space. During the fine-tuning or inference phase of the downstream classification task based solely on fundus images, where paired OCT data is unavailable, we eliminate OCT input and utilize the predilection sites matrix to assist in fundus image classification learning. Extensive experiments conducted on 9 diverse datasets across 28 critical categories demonstrate that our framework outperforms existing benchmarks.

Figures

Figures reproduced from arXiv: 2509.08624 by the authors.

Figure 1
Figure 1. Overview of our proposed UOPSL. During the model training phase, it concurrently accepts OCT, fundus, and disease text as inputs while dynamically learning the predilection sites matrix P. For fine-tuning in downstream classification tasks or the inference phase, the OCT module can optionally be excluded from the input. OCT images, and extended disease category texts with the same disease are jointly processed. Spec… view at source ↗
Figure 2
Figure 2. Overview of the fine-tuning and inference phases. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Generalization Capability on Unseen Classes [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 4 linked inside Pith

  1. [12]

    Multieye: Dataset and benchmark for oct-enhanced retinal disease recognition from fundus images,

    L. Wang, C. Qi, C. Ou, L. An, M. Jin, X. Kong, and X. Li, “Multieye: Dataset and benchmark for oct-enhanced retinal disease recognition from fundus images,”IEEE Transactions on Medical Imaging, 2024

  2. [1]

    The possibility of the combination of oct and fundus images for improving the diagnostic accuracy of deep learning for age- related macular degeneration: a preliminary experiment,

    T. K. Yoo, J. Y . Choi, J. G. Seo, B. Ramasubramanian, S. Selvaperumal, and D. W. Kim, “The possibility of the combination of oct and fundus images for improving the diagnostic accuracy of deep learning for age- related macular degeneration: a preliminary experiment,”Medical & biological engineering & computing, vol. 57, pp. 677–687, 2019

  3. [2]

    Application of machine learning in ophthalmic imaging modalities,

    Y . Tong, W. Lu, Y . Yu, and Y . Shen, “Application of machine learning in ophthalmic imaging modalities,”Eye and Vision, vol. 7, pp. 1–15, 2020

  4. [3]

    Ophthalmic fundus imaging: today and beyond,

    L. A. Yannuzzi, M. D. Ober, J. S. Slakter, R. F. Spaide, Y . L. Fisher, R. W. Flower, and R. Rosen, “Ophthalmic fundus imaging: today and beyond,”American journal of ophthalmology, vol. 137, no. 3, pp. 511– 524, 2004

  5. [4]

    Visionfm: a multi-modal multi-task vision foundation model for generalist ophthalmic artificial intelligence,

    J. Qiu, J. Wu, H. Wei, P. Shi, M. Zhang, Y . Sun, L. Li, H. Liu, H. Liu, S. Houet al., “Visionfm: a multi-modal multi-task vision foundation model for generalist ophthalmic artificial intelligence,”arXiv preprint arXiv:2310.04992, 2023

  6. [5]

    Multimodal machine learning: A survey and taxonomy,

    T. Baltru ˇsaitis, C. Ahuja, and L.-P. Morency, “Multimodal machine learning: A survey and taxonomy,”IEEE transactions on pattern anal- ysis and machine intelligence, vol. 41, no. 2, pp. 423–443, 2018

  7. [6]

    Clip-art: Contrastive pre-training for fine- grained art classification,

    M. V . Conde and K. Turgutlu, “Clip-art: Contrastive pre-training for fine- grained art classification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 3956–3960

  8. [7]

    Multi-modal retinal image clas- sification with modality-specific attention network,

    X. He, Y . Deng, L. Fang, and Q. Peng, “Multi-modal retinal image clas- sification with modality-specific attention network,”IEEE transactions on medical imaging, vol. 40, no. 6, pp. 1591–1602, 2021

Show all 40 references
  1. [8]

    Cross-modal attention network for retinal disease classification based on multi-modal images,

    Z. Liu, Y . Hu, Z. Qiu, Y . Niu, D. Zhou, X. Li, J. Shen, H. Jiang, H. Li, and J. Liu, “Cross-modal attention network for retinal disease classification based on multi-modal images,”Biomedical Optics Express, vol. 15, no. 6, pp. 3699–3714, 2024

  2. [9]

    Tagat: Topology-aware graph attention network for multi-modal retinal image fusion,

    X. Tian, N. Anantrasirichai, L. Nicholson, and A. Achim, “Tagat: Topology-aware graph attention network for multi-modal retinal image fusion,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 775–784

  3. [10]

    A multimodal imaging–based deep learning model for detecting treatment-requiring retinal vascular diseases: model development and validation study,

    E. Y .-C. Kang, L. Yeung, Y .-L. Lee, C.-H. Wu, S.-Y . Peng, Y .-P. Chen, Q.-Z. Gao, C. Lin, C.-F. Kuo, and C.-C. Lai, “A multimodal imaging–based deep learning model for detecting treatment-requiring retinal vascular diseases: model development and validation study,” JMIR Med...

  4. [11]

    Understanding multimodal contrastive learning and incorporating un- paired data,

    R. Nakada, H. I. Gulluk, Z. Deng, W. Ji, J. Zou, and L. Zhang, “Understanding multimodal contrastive learning and incorporating un- paired data,” inInternational Conference on Artificial Intelligence and Statistics. PMLR, 2023, pp. 4348–4380

  5. [13]

    A foundation model for generalizable disease detection from retinal images,

    Y . Zhou, M. A. Chia, S. K. Wagner, M. S. Ayhan, D. J. Williamson, R. R. Struyven, T. Liu, M. Xu, M. G. Lozano, P. Woodward-Courtet al., “A foundation model for generalizable disease detection from retinal images,”Nature, vol. 622, no. 7981, pp. 156–163, 2023

  6. [14]

    Urfound: Towards universal retinal foundation models via knowledge-guided masked modeling,

    K. Yu, Y . Zhou, Y . Bai, Z. D. Soh, X. Xu, R. S. M. Goh, C.-Y . Cheng, and Y . Liu, “Urfound: Towards universal retinal foundation models via knowledge-guided masked modeling,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer,...

  7. [15]

    A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision,

    J. Silva-Rodriguez, H. Chakor, R. Kobbi, J. Dolz, and I. B. Ayed, “A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision,”Medical Image Analysis, vol. 99, p. 103357, 2025

  8. [16]

    Ret-clip: A retinal image foundation model pre-trained with clinical diagnostic reports,

    J. Du, J. Guo, W. Zhang, S. Yang, H. Liu, H. Li, and N. Wang, “Ret-clip: A retinal image foundation model pre-trained with clinical diagnostic reports,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 709–719

  9. [17]

    Vilref: an expert knowledge enabled vision-language retinal foundation model,

    S. Yang, J. Du, J. Guo, W. Zhang, H. Liu, H. Li, and N. Wang, “Vilref: an expert knowledge enabled vision-language retinal foundation model,” arXiv preprint arXiv:2408.10894, 2024

  10. [18]

    Enhancing diagnostic accuracy in rare and common fundus diseases with a knowledge-rich vision-language model,

    M. Wang, T. Lin, A. Lin, K. Yu, Y . Peng, L. Wang, C. Chen, K. Zou, H. Liang, M. Chenet al., “Enhancing diagnostic accuracy in rare and common fundus diseases with a knowledge-rich vision-language model,” Nature Communications, vol. 16, no. 1, p. 5528, 2025

  11. [19]

    Visionunite: A vision-language foundation model for ophthalmology enhanced with clinical knowledge,

    Z. Li, D. Song, Z. Yang, D. Wang, F. Li, X. Zhang, P. E. Kina- han, and Y . Qiao, “Visionunite: A vision-language foundation model for ophthalmology enhanced with clinical knowledge,”arXiv preprint arXiv:2408.02865, 2024

  12. [20]

    Language enhanced model for eye (leme): An open-source ophthalmology-specific large language model,

    A. Gilson, X. Ai, Q. Xie, S. Srinivasan, K. Pushpanathan, M. B. Singer, J. Huang, H. Kim, E. Long, P. Wanet al., “Language enhanced model for eye (leme): An open-source ophthalmology-specific large language model,”arXiv preprint arXiv:2410.03740, 2024

  13. [21]

    Is there any predilection of left eyes for glaucoma?

    K. Ascher and W. Spurgeon, “Is there any predilection of left eyes for glaucoma?”American Journal of Ophthalmology, vol. 35, no. 10, pp. 1480–1483, 1952

  14. [22]

    Feature prediction across eye movements is location specific and based on retinotopic coordinates,

    A. Herwig, K. Weiß, and W. X. Schneider, “Feature prediction across eye movements is location specific and based on retinotopic coordinates,” Journal of vision, vol. 18, no. 8, pp. 13–13, 2018

  15. [23]

    Spatio-temporal signatures to predict retinal disease recurrence,

    W.-D. V ogl, S. M. Waldstein, B. S. Gerendas, C. Simader, A.-M. Glodan, D. Podkowinski, U. Schmidt-Erfurth, and G. Langs, “Spatio-temporal signatures to predict retinal disease recurrence,” inInternational Con- ference on Information Processing in Medical Imaging. Springer, 20...

  16. [24]

    Spatial analysis reveals vascular changes in retinal and choroidal vessel perfusion in intermediate amd with reticular pseudodrusen,

    J. Nam, L. Nivison-Smith, and M. Trinh, “Spatial analysis reveals vascular changes in retinal and choroidal vessel perfusion in intermediate amd with reticular pseudodrusen,”Investigative Ophthalmology & Visual Science, vol. 65, no. 2, pp. 33–33, 2024

  17. [25]

    Sd-layernet: Semi-supervised retinal layer segmentation in oct using disentangled representation with anatomical priors,

    B. Fazekas, G. Aresta, D. Lachinov, S. Riedl, J. Mai, U. Schmidt- Erfurth, and H. Bogunovi ´c, “Sd-layernet: Semi-supervised retinal layer segmentation in oct using disentangled representation with anatomical priors,” inInternational Conference on Medical Image Computing and C...

  18. [26]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

  19. [27]

    Demystifying clip data,

    H. Xu, S. Xie, X. E. Tan, P.-Y . Huang, R. Howes, V . Sharma, S.-W. Li, G. Ghosh, L. Zettlemoyer, and C. Feichtenhofer, “Demystifying clip data,”arXiv preprint arXiv:2309.16671, 2023

  20. [28]

    Biomedbert: A pre-trained biomedical language model for qa and ir,

    S. Chakraborty, E. Bisong, S. Bhatt, T. Wagner, R. Elliott, and F. Mosconi, “Biomedbert: A pre-trained biomedical language model for qa and ir,” inProceedings of the 28th international conference on computational linguistics, 2020, pp. 669–679

  21. [29]

    Evaluation of retinal image quality assessment networks in different color-spaces,

    H. Fu, B. Wang, J. Shen, S. Cui, Y . Xu, J. Liu, and L. Shao, “Evaluation of retinal image quality assessment networks in different color-spaces,” inMedical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October ...

  22. [30]

    Feedback on a publicly distributed image database: the messidor database,

    E. Decenci `ere, X. Zhang, G. Cazuguel, B. Lay, B. Cochener, C. Trone, P. Gain, J.-R. Ord ´o˜nez-Varela, P. Massin, A. Erginayet al., “Feedback on a publicly distributed image database: the messidor database,”Image Analysis & Stereology, pp. 231–234, 2014

  23. [31]

    Idrid: Diabetic retinopathy–segmentation and grading challenge,

    P. Porwal, S. Pachade, M. Kokare, G. Deshmukh, J. Son, W. Bae, L. Liu, J. Wang, X. Liu, L. Gaoet al., “Idrid: Diabetic retinopathy–segmentation and grading challenge,”Medical image analysis, vol. 59, p. 101561, 2020

  24. [32]

    APTOS 2019 Blindness Detection,

    A. E. Hospital, “APTOS 2019 Blindness Detection,” https://www. kaggle.com/competitions/aptos2019-blindness-detection, 2019, [Online; accessed 27-June-2019]

  25. [33]

    Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment,

    O. Kovalyk, J. Morales-S ´anchez, R. Verd´u-Monedero, I. Sell ´es-Navarro, A. Palaz ´on-Cabanes, and J.-L. Sancho-G ´omez, “Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment,”Scientific Data, vol. 9, no. 1, p. 291, 2022

  26. [34]

    A deep learning model for the detection of both advanced and early glaucoma using fundus photography,

    J. M. Ahn, S. Kim, K.-S. Ahn, S.-H. Cho, K. B. Lee, and U. S. Kim, “A deep learning model for the detection of both advanced and early glaucoma using fundus photography,”PloS one, vol. 13, no. 11, p. e0207982, 2018

  27. [35]

    Automatic detection of 39 fundus diseases and conditions in retinal photographs using deep neural networks,

    L.-P. Cen, J. Ji, J.-W. Lin, S.-T. Ju, H.-J. Lin, T.-P. Li, Y . Wang, J.-F. Yang, Y .-F. Liu, S. Tanet al., “Automatic detection of 39 fundus diseases and conditions in retinal photographs using deep neural networks,” Nature communications, vol. 12, no. 1, p. 4828, 2021

  28. [36]

    Retinal fundus multi-disease image dataset (rfmid): A dataset for multi-disease detection research,

    S. Pachade, P. Porwal, D. Thulkar, M. Kokare, G. Deshmukh, V . Sa- hasrabuddhe, L. Giancardo, G. Quellec, and F. M ´eriaudeau, “Retinal fundus multi-disease image dataset (rfmid): A dataset for multi-disease detection research,”Data, vol. 6, no. 2, p. 14, 2021

  29. [37]

    Peking university international competition on ocular disease intelligent recognition (odir-2019),

    G. Challenge, “Peking university international competition on ocular disease intelligent recognition (odir-2019),” 2019

  30. [38]

    Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities,

    M. Uzair Khattak, S. Kunhimon, M. Naseer, S. Khan, and F. Shah- baz Khan, “Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities,”arXiv e-prints, pp. arXiv–2412, 2024

  31. [39]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

  32. [40]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

Pith tools

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