Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

A retina-specialist vision model consistently outperforms large generalist models on ocular-disease and oculomics tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

On ten retinal imaging tasks, retina-trained RETFound-DINOv2 edged out larger generalist DINOv2/DINOv3 models, but the average advantage was small.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful, competently run benchmark of generalist vs specialist retinal FMs, but the 'consistently outperforms' headline outruns the data, and the AlzEye pretraining/evaluation overlap is a real confound that needs resolving. the 4 major comments →

arxiv 2509.03421 v1 pith:JD6G4AN2 submitted 2025-09-03 eess.IV cs.CV

Generalist versus Specialist Vision Foundation Models for Ocular Disease and Oculomics

Yukun Zhou , Paul Nderitu , Jocelyn Hui Lin Goh , Justin Engelmann , Siegfried K. Wagner , Anran Ran , Hongyang Jiang , Lie Ju
show 15 more authors
Ke Zou Sahana Srinivasan Hyunmin Kim Takahiro Ninomiya Zheyuan Wang Gabriel Dawei Yang Eden Ruffell Dominic Williamson Rui Santos Gabor Mark Somfai Carol Y. Cheung Tien Yin Wong Daniel C. Alexander Yih Chung Tham Pearse A. Keane
This is my paper
classification eess.IV cs.CV
keywords retinal foundation modelsRETFoundDINOv2DINOv3oculomicssystemic disease predictionfine-tuning vs linear probingdata efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether AI vision models pre-trained on billions of natural images can replace a model pre-trained specifically on retinal images for eye-disease detection and for predicting heart attack, stroke, heart failure, and Parkinson's disease from retinal photos. The paper benchmarks the generalist DINOv2 and DINOv3 model families against two specialist RETFound models across ten tasks, using both fine-tuning and linear probing. It finds that RETFound-DINOv2 — a generalist DINOv2 backbone further pre-trained on about 904,000 retinal images — gets the highest average AUROC (0.830 with fine-tuning), ahead of DINOv3-ViT-large (0.816), RETFound-MAE (0.809), and DINOv2-ViT-giant (0.800). The conclusion is that domain-specific pre-training still buys a real clinical advantage, but the gap to much larger generalist models is shrinking.

Core claim

On the paper's own terms, the core claim is that a retina-specialist foundation model, RETFound-DINOv2, consistently outperforms state-of-the-art generalist vision foundation models in ocular disease detection and systemic disease prediction from retinal images, and it does so with better data efficiency. The advantage holds under both fine-tuning and linear probing, and is most visible in external evaluation: when models fine-tuned on AlzEye are applied to UK Biobank, RETFound-DINOv2 reaches 0.599 average AUROC versus 0.571 for RETFound-MAE, 0.560 for DINOv2-ViT-giant, and 0.549 for DINOv3-ViT-large. The paper also reports that RETFound-DINOv2's image features are less self-similar (average

What carries the argument

The load-bearing object is RETFound-DINOv2, a Vision Transformer (ViT-large, about 300M parameters) initialised from DINOv2-ViT-large and further pre-trained on 904K retinal images from the AlzEye cohort. The paper compares it with RETFound-MAE (the same retinal pretraining recipe from a MAE-ViT-large start) and with generalists from the DINOv2 family (up to ViT-giant, about 1.1B parameters) and the DINOv3 family (up to ViT-huge+, with DINOv3-ViT-large the best performer). The evaluation protocol — ten tasks, fine-tuning and linear probing, bootstrapped confidence intervals, cross-dataset significance tests — carries the comparison, along with feature self-similarity analysis on 250 UK Bioba

Load-bearing premise

The comparison assumes the same RETFound training recipe — data augmentation, learning-rate schedule, batch size, and epoch count — is equally suitable for DINOv2 and DINOv3; no per-model hyperparameter tuning is reported.

What would settle it

A head-to-head repeat of the ten-task benchmark in which DINOv3-ViT-large and DINOv2-ViT-giant receive individually tuned hyperparameters while RETFound-DINOv2 is retrained under the same protocol: if their average AUROC then matches or exceeds RETFound-DINOv2, the specialist advantage claimed here would not hold.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Specialist retinal pretraining remains the better default for clinical eye-disease tasks, at least under a shared training recipe.
  • DINOv3-ViT-large is close enough to be a credible starting point for future medical foundation models, especially given its much smaller parameter count than DINOv2-ViT-giant.
  • Linear probing preserves most of the fine-tuned performance at a fraction of the memory and runtime, making frozen-feature pipelines practical in resource-constrained clinics.
  • Because different foundation models encode very different feature spaces, aggregating or aligning them — rather than picking a single one — is a promising next direction.
  • Scaling behaviour from natural images does not transfer directly to medical imaging: the largest DINOv3 variants tested did not always beat smaller ones.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If generalist models were given per-model hyperparameter tuning, the RETFound-DINOv2 advantage could shrink; a fair comparison with tuned schedules is the natural next experiment.
  • The low cross-model feature similarity suggests an ensemble or late fusion of RETFound-DINOv2 and DINOv3-ViT-large might outperform either alone, a test that has not been run here.
  • The relative success of DINOv3-ViT-large hints that continued scaling of generalist pretraining data may eventually erode the need for separately collected medical pretraining corpora.
  • Because linear probing performs well, frozen generalist embeddings could make external validation and privacy-preserving deployment easier, since patient images need not be stored for retraining.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper benchmarks two families of generalist vision foundation models (DINOv2 and DINOv3) against two specialist retinal foundation models (RETFound-MAE and RETFound-DINOv2) across ten ocular-disease and oculomics tasks, using fine-tuning and linear probing. It reports average AUROCs of 0.830 for RETFound-DINOv2, 0.816 for DINOv3-ViT-large, 0.809 for RETFound-MAE, and 0.800 for DINOv2-ViT-giant under fine-tuning, and uses a Wilcoxon signed-rank test to claim that RETFound-DINOv2 significantly outperforms DINOv3-ViT-large. The paper also reports data-efficiency and adaptation-efficiency comparisons and a feature-similarity analysis. The headline claim is that RETFound-DINOv2 consistently outperforms generalist models, implying that domain-specific pretraining remains valuable.

Significance. If the central claim is valid, the paper is a useful benchmark for the medical-imaging foundation-model community, with practical implications for whether to pretrain on domain data or rely on scaled generalist models. The manuscript's strengths include the use of several public ocular datasets, an external UK Biobank evaluation for oculomics, release of code and model weights, and a systematic comparison that includes data and adaptation efficiency. However, the main conclusion currently rests on two load-bearing assumptions that are not established in the text: (i) that the internal AlzEye oculomics evaluation is not contaminated by overlap with the AlzEye images used to pretrain RETFound, and (ii) that the shared RETFound-derived fine-tuning protocol is equally suitable for the generalist models. The paper's own statement that all four models 'performed comparably on most tasks' also stands in tension with the abstract's 'consistently outperforms.' The external evaluation partially mitigates these concerns but does not eliminate them. The underlying comparison is likely salvageable, but the evidence as presented does not fully support the strength of the claim.

major comments (4)
  1. [Datasets / Model description (Methods)] RETFound-DINOv2 was pretrained on 904K retinal images from the AlzEye cohort, and the same AlzEye cohort is used for model adaptation and internal evaluation of the four systemic-disease (oculomics) tasks. The manuscript never states that pretraining images or patients were excluded from the downstream 55:15:30 train/validation/test splits. If the same images, or different images from the same patients, appear in the internal test set, the reported oculomics AUROC advantage is partly a distribution-overlap effect rather than evidence of better transfer. Because four of the ten tasks are AlzEye-based and the headline averages over all ten, this directly affects the 'consistently outperforms' claim. Please report whether any pretraining images/patients appear in the downstream splits and, if they do, rerun the internal evaluation after exclusion. At minimum, report results separately for t
  2. [Model adaptation to downstream tasks (Methods)] All models are trained with the same augmentation schedule (RETFound's random crop and flip), batch size 24, 50 epochs, and learning-rate schedule 5e-4 to 1e-6, and checkpoints are selected by the same sum-of-AUROC-and-F1 criterion. No per-model hyperparameter tuning or sensitivity analysis is reported. The Discussion itself concedes that the RETFound models' reported results 'may not fully reflect optimised performance.' If this shared protocol is better suited to RETFound than to DINOv2/DINOv3, the small 0.014 AUROC advantage of RETFound-DINOv2 over DINOv3-ViT-large could shrink or disappear under a fairer protocol. This is a load-bearing issue for the central comparison. Please add a sensitivity analysis with alternative schedules/epoch sizes, or present the results as a comparison under the RETFound protocol rather than as general evidence about specialist-versus-generalist capabilit
  3. [Results, Figure 4 / Abstract] The text in the 'RETFound-DINOv2 outperformed...' section states that 'all four models performed comparably on most tasks,' yet the abstract claims RETFound-DINOv2 'consistently outperforms' the generalist models. The average AUROC difference between RETFound-DINOv2 and DINOv3-ViT-large is 0.014, which is small and, as the paper itself notes, often within per-task comparability. No clinically meaningful effect-size threshold is given, and the Wilcoxon test across only ten tasks with no multiple-comparison correction is a weak basis for 'consistent' superiority. Please temper the language, report per-task wins/losses, and provide a quantitative justification for what degree of AUROC difference is clinically meaningful.
  4. [Statistical analysis (Methods)] The bootstrap confidence-interval procedure appears incorrect. The text says: 'performed bootstrapping (n=200, repetitive sampling in full test size). The mean value and standard deviation across the 200 bootstraps were reported. The standard error was estimated as (standard deviation / √200), and 95% CI were obtained as 1.96 × standard error.' This computes the standard error of the bootstrap mean, not the bootstrap standard error of the AUROC statistic; the latter is simply the standard deviation of the bootstrap replicates. Using sd/√200 shrinks the CIs by a factor of about 14.1, making reported uncertainty intervals misleading. Additionally, if test data contain multiple images from the same patient, the bootstrap should be clustered by patient. Please correct the bootstrap procedure and recalculate the CIs, as these CIs are used in Figures 2–4 and Supplementary Table 3.
minor comments (5)
  1. [Supplementary Figure 11 caption] The caption says models are 'externally evaluating models via cross-validation on three diabetic retinopathy datasets.' This is cross-dataset transfer, not external validation; the terminology should be clarified.
  2. [Statistical analysis (Methods)] The Wilcoxon signed-rank p-values are computed across ten tasks. Because multiple pairwise comparisons are made (DINOv2 series, DINOv3 series, and the four-model comparison), exact p-values and a note on multiple comparisons would be more informative than asterisks alone.
  3. [Feature analysis (Results)] The claim that RETFound-DINOv2's lower self-similarity 'likely contributes' to downstream performance is speculative. The paper does not establish a causal or correlational link between feature self-similarity and task AUROC; please soften or provide supporting analysis.
  4. [Methods, image preprocessing] The text says all images are resized to 256×256 with cubic interpolation, but the augmentation description then states random crops are resized to 224×224. Clarify the relationship between the 256×256 resizing and the 224×224 training resolution.
  5. [Discussion, adaptation efficiency] The statement that fine-tuning DINOv2-ViT-giant requires '~5 times more memory and ~3 times longer runtime' is based on a single dataset (APTOS-2019); this is a useful estimate but should be labelled as dataset-specific rather than general.

Circularity Check

0 steps flagged

No circularity: this is an empirical benchmark comparison with no fitted-parameter or definitional reduction; self-citations are protocol references and are not load-bearing evidence.

full rationale

The paper is an empirical comparison of pretrained vision models after fine-tuning or linear probing on ten downstream tasks. There is no derivation chain in which a predicted quantity is constructed from its own target: AUROC values are measured on held-out test splits (AlzEye 55:15:30, with external UK Biobank evaluation and public ocular datasets for DR/glaucoma). RETFound-DINOv2's advantage is an observed outcome of this protocol, not an algebraic consequence of pretraining on AlzEye. The self-citations to RETFound [7] and RETFound-DINOv2 [24] supply the model checkpoints and the training recipe; they do not assert a uniqueness theorem or forbid alternative comparisons, and DINOv2/DINOv3 checkpoints are external (Meta/FAIR). The shared RETFound recipe (augmentation, LR schedule) for all models is a protocol-choice concern that could bias the comparison, but it is not circular: 'same recipe' does not make the outcome equal to the input. The AlzEye pretraining/evaluation overlap is a valid external-validity caveat about distribution overlap, but it is not a reduction of the claim to its inputs by construction; the paper even includes UK Biobank external evaluation and public cross-dataset DR checks. No equation is defined in terms of another by construction, no fitted parameter is renamed as a prediction, and no load-bearing self-citation replaces evidence. Therefore the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The comparison relies on fixed upstream weights, label quality, and a single shared adaptation protocol; no new entities or fitted constants are introduced. The two listed selection choices are hand-chosen and affect the reported comparisons.

free parameters (2)
  • Shared fine-tuning schedule = LR 5e-4 to 1e-6, 50 epochs, batch size 24
    Taken from RETFound [7] and applied uniformly; no evidence it is optimal for DINOv2 and DINOv3, so the comparison depends on this hand-chosen protocol.
  • Checkpoint selection criterion = max AUROC + F1 on validation set
    Hand-chosen for all models; affects reported test AUROC and could favor models with smoother validation curves.
axioms (4)
  • domain assumption Upstream pretrained weights are accepted as given and are comparable starting points.
    Model description relies on published weights from DINOv2, DINOv3, and RETFound; they cannot be independently re-trained in this study.
  • domain assumption Labels in all datasets are sufficiently accurate to rank models.
    No label-noise or adjudication analysis is provided; all AUROC and F1 calculations assume the ground truth is correct.
  • standard math Bootstrapping and Wilcoxon signed-rank tests are valid for correlated task-level scores.
    The statistical analysis section reports n=200 bootstraps and two-sided Wilcoxon tests, but no multiple-testing correction is applied across the many pairwise model comparisons.
  • domain assumption Single left-eye image per visit controls patient-level dependence.
    Datasets section states 'For each patient, we include a single left-eye image per visit to avoid visit-related bias.'

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalist versus Specialist Vision Foundation Models for Ocular Disease and Oculomics." pith.science (2026). https://pith.science/paper/JD6G4AN2

@misc{pith2026250903421,
  author       = {Pith},
  title        = {Pith review of: Generalist versus Specialist Vision Foundation Models for Ocular Disease and Oculomics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JD6G4AN2}},
  note         = {Machine review of arXiv:2509.03421}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Medical foundation models, pre-trained with large-scale clinical data, demonstrate strong performance in diverse clinically relevant applications. RETFound, trained on nearly one million retinal images, exemplifies this approach in applications with retinal images. However, the emergence of increasingly powerful and multifold larger generalist foundation models such as DINOv2 and DINOv3 raises the question of whether domain-specific pre-training remains essential, and if so, what gap persists. To investigate this, we systematically evaluated the adaptability of DINOv2 and DINOv3 in retinal image applications, compared to two specialist RETFound models, RETFound-MAE and RETFound-DINOv2. We assessed performance on ocular disease detection and systemic disease prediction using two adaptation strategies: fine-tuning and linear probing. Data efficiency and adaptation efficiency were further analysed to characterise trade-offs between predictive performance and computational cost. Our results show that although scaling generalist models yields strong adaptability across diverse tasks, RETFound-DINOv2 consistently outperforms these generalist foundation models in ocular-disease detection and oculomics tasks, demonstrating stronger generalisability and data efficiency. These findings suggest that specialist retinal foundation models remain the most effective choice for clinical applications, while the narrowing gap with generalist foundation models suggests that continued data and model scaling can deliver domain-relevant gains and position them as strong foundations for future medical foundation models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Super-Generalist: Towards Comprehensive and Accurate Medical Image Understanding via Generalist-Specialist Synergy

    cs.CV 2026-07 conditional novelty 6.0

    Injecting multi-expert anatomy/lesion segmentation priors into vision–language alignment and calibrating text attention with lesion masks yields broad CT diagnosis plus specialist-level tumor performance and lesion grounding.

  2. Representation Transfer of Foundation Models for Ultra-Widefield Retinal Imaging

    cs.CV 2026-08 conditional novelty 5.0

    Pretraining objective strongly affects frozen feature transfer in MIL for ultra-widefield retinal imaging; DINO-based encoders outperform MAE on five-class DR grading, a gap mostly closed by partial fine-tuning.

Reference graph

Works this paper leans on

39 extracted references · 28 canonical work pages · cited by 2 Pith papers · 1 internal anchor

  1. [1]

    On the Opportunities and Risks of Foundation Models

    Bommasani R, Hudson DA, Adeli E, Altman R, Arora S, von Arx S, et al. On the Opportunities and Risks of Foundation Models. arXiv [cs.LG]. 2021. Available: http://arxiv.org/abs/2108.07258

    Pith/arXiv arXiv 2021

  2. [2]

    On the Opportunities and Risks of Foundation Models for Natural Language Processing in Radiology

    Wiggins WF, Tejani AS. On the Opportunities and Risks of Foundation Models for Natural Language Processing in Radiology. Radiol Artif Intell. 2022;4: e220119

    work page 2022

  3. [3]

    Foundation models for generalist medical artificial intelligence

    Moor M, Banerjee O, Abad ZSH, Krumholz HM, Leskovec J, Topol EJ, et al. Foundation models for generalist medical artificial intelligence. Nature. 2023;616: 259–265

    work page 2023

  4. [4]

    Foundation Model for Advancing Healthcare: Challenges, Opportunities and Future Directions

    He Y, Huang F, Jiang X, Nie Y, Wang M, Wang J, et al. Foundation Model for Advancing Healthcare: Challenges, Opportunities and Future Directions. IEEE Rev Biomed Eng. 2025;18: 172–191

    work page 2025

  5. [5]

    On the challenges and perspectives of foundation models for medical image analysis

    Zhang S, Metaxas D. On the challenges and perspectives of foundation models for medical image analysis. Med Image Anal. 2024;91: 102996

    work page 2024

  6. [6]

    A new foundation model for multimodal ophthalmic images: Advancing disease detection and prediction

    Chia MA, Zhou Y, Keane PA. A new foundation model for multimodal ophthalmic images: Advancing disease detection and prediction. NEJM AI. 2024;1. doi:10.1056/aie2401024

    work page doi:10.1056/aie2401024 2024

  7. [7]

    A foundation model for generalizable disease detection from retinal images

    Zhou Y, Chia MA, Wagner SK, Ayhan MS, Williamson DJ, Struyven RR, et al. A foundation model for generalizable disease detection from retinal images. Nature. 2023;622: 156–163

    work page 2023

  8. [8]

    Advanced Intelligent Computing Technology and Applications: 21st International Conference, ICIC 2025, Ningbo, China, July 26–29, 2025, Proceedings, Part VIII

    Huang D-S, Chen H, Li B, Zhang Q. Advanced Intelligent Computing Technology and Applications: 21st International Conference, ICIC 2025, Ningbo, China, July 26–29, 2025, Proceedings, Part VIII. Springer Nature; 2025

    work page 2025

  9. [9]

    Integrated image-based deep learning and language models for primary diabetes care

    Li J, Guan Z, Wang J, Cheung CY, Zheng Y, Lim L-L, et al. Integrated image-based deep learning and language models for primary diabetes care. Nat Med. 2024;30: 2886–2896

    work page 2024

  10. [10]

    Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning

    Tiu E, Talius E, Patel P, Langlotz CP, Ng AY, Rajpurkar P. Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning. Nat Biomed Eng. 2022. doi:10.1038/s41551-022-00936-9

    doi:10.1038/s41551-022-00936-9 2022

  11. [11]

    Foundation model for cancer imaging biomarkers

    Pai S, Bontempi D, Hadzic I, Prudente V, Sokač M, Chaunzwa TL, et al. Foundation model for cancer imaging biomarkers. Nat Mach Intell. 2024;6: 354–367

    work page 2024

  12. [12]

    Collaboration between clinicians and vision-language models in radiology report generation

    Tanno R, Barrett DGT, Sellergren A, Ghaisas S, Dathathri S, See A, et al. Collaboration between clinicians and vision-language models in radiology report generation. Nat Med. 2025;31: 599–608

    work page 2025

  13. [13]

    A visual-language foundation model for pathology image analysis using medical Twitter

    Huang Z, Bianchi F, Yuksekgonul M, Montine TJ, Zou J. A visual-language foundation model for pathology image analysis using medical Twitter. Nat Med. 2023;29: 2307–2316

    work page 2023

  14. [14]

    A visual-language foundation model for computational pathology

    Lu MY, Chen B, Williamson DFK, Chen RJ, Liang I, Ding T, et al. A visual-language foundation model for computational pathology. Nat Med. 2024;30: 863–874

    work page 2024

  15. [15]

    A whole-slide foundation model for digital pathology from real-world data

    Xu H, Usuyama N, Bagga J, Zhang S, Rao R, Naumann T, et al. A whole-slide foundation model for digital pathology from real-world data. Nature. 2024;630: 181–188

    work page 2024

  16. [16]

    Momentum contrast for unsupervised visual representation learning

    He K, Fan H, Wu Y, Xie S, Girshick R. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2020. pp. 9729–9738

    work page 2020

  17. [17]

    Emerging properties in self-supervised vision transformers

    Caron M, Touvron H, Misra I, Jégou H, Mairal J, Bojanowski P, et al. Emerging properties in self-supervised vision transformers. arXiv [cs.CV]. 2021. pp. 9650–9660. Available: http://openaccess.thecvf.com/content/ICCV2021/html/Caron_Emerging_Properties_in_Self-Supervised_Vision_Transformers_ICCV_2021_paper.html

    work page 2021

  18. [18]

    DINOv2: Learning robust visual features without supervision

    Oquab M, Darcet T, Moutakanni T, Vo H, Szafraniec M, Khalidov V, et al. DINOv2: Learning robust visual features without supervision. arXiv [cs.CV]. 2023. Available: http://arxiv.org/abs/2304.07193

    Pith/arXiv arXiv 2023

  19. [19]

    Masked Autoencoders Are Scalable Vision Learners

    He K, Chen X, Xie S, Li Y, Doll’ar P, Girshick RB. Masked Autoencoders Are Scalable Vision Learners. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2021; 15979–15988

    work page 2021

  20. [20]

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

    Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, et al. An image is worth 16x16 words: Transformers for image recognition at scale. Int Conf Learn Represent. 2020;abs/2010.11929. doi:10.11929/1000

    Pith/arXiv arXiv 2020

  21. [21]

    Toward expert-level medical question answering with large language models

    Singhal K, Tu T, Gottweis J, Sayres R, Wulczyn E, Amin M, et al. Toward expert-level medical question answering with large language models. News@nat,Com. 2025. doi:10.1038/s41591-024-03423-7

    doi:10.1038/s41591-024-03423-7 2025

  22. [22]

    Block Expanded DINORET: Adapting Natural Domain Foundation Models for Retinal Imaging Without Catastrophic Forgetting

    Zoellin J, Merk C, Buob M, Saad A, Giesser S, Spitznagel T, et al. Block expanded DINORET: Adapting natural domain foundation models for retinal imaging without catastrophic forgetting. arXiv [cs.CV]. 2024. Available: http://arxiv.org/abs/2409.17332

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    FusionFM: Fusing eye-specific foundational models for optimized ophthalmic diagnosis

    Zou K, Goh JHL, Zhou Y, Lin T, Yew SME, Srinivasan S, et al. FusionFM: Fusing eye-specific foundational models for optimized ophthalmic diagnosis. arXiv [cs.CV]. 2025. Available: http://arxiv.org/abs/2508.11721

    Pith/arXiv arXiv 2025

  24. [24]

    Revealing the Impact of Pre-training Data on Medical Foundation Models

    Zhou Y, Wang Z, Wu Y, Ong AY, Wagner S, Ruffell E. Revealing the Impact of Pre-training Data on Medical Foundation Models. 2025. Available: https://www.researchsquare.com/article/rs-6080254/latest

    work page 2025

  25. [25]

    Siméoni O, Vo HV, Seitzer M, Baldassarre F, Oquab M, Jose C, et al. DINOv3. arXiv [cs.CV]. 2025. Available: http://arxiv.org/abs/2508.10104

    Pith/arXiv arXiv 2025

  26. [26]

    Can a natural image-based foundation model outperform a retina-specific model in detecting ocular and systemic diseases? Ophthalmol Sci

    Hou Q, Zhou Y, Goh JHL, Zou K, Yew SME, Srinivasan S, et al. Can a natural image-based foundation model outperform a retina-specific model in detecting ocular and systemic diseases? Ophthalmol Sci. 2025; 100923

    work page 2025

  27. [27]

    Insights into Systemic Disease through Retinal Imaging-Based Oculomics

    Wagner SK, Fu DJ, Faes L, Liu X, Huemer J, Khalid H, et al. Insights into Systemic Disease through Retinal Imaging-Based Oculomics. Transl Vis Sci Technol. 2020;9: 6

    work page 2020

  28. [28]

    Oculomics: Current concepts and evidence

    Zhu Z, Wang Y, Qi Z, Hu W, Zhang X, Wagner SK, et al. Oculomics: Current concepts and evidence. Prog Retin Eye Res. 2025;106: 101350

    work page 2025

  29. [29]

    AlzEye: longitudinal record-level linkage of ophthalmic imaging and hospital admissions of 353 157 patients in London, UK

    Wagner S, Hughes F, Cortina-Borja M, Pontikos N, Struyven R, Liu X, et al. AlzEye: longitudinal record-level linkage of ophthalmic imaging and hospital admissions of 353 157 patients in London, UK. BMJ Open. 2022;12. doi:10.1136/bmjopen-2021-058552

    doi:10.1136/bmjopen-2021-058552 2022

  30. [30]

    Idrid: Diabetic retinopathy--segmentation and grading challenge

    Porwal P, Pachade S, Kokare M, Deshmukh G, Son J, Bae W, et al. Idrid: Diabetic retinopathy--segmentation and grading challenge. Med Image Anal. 2020;59: 101561

    work page 2020

  31. [31]

    Automated Analysis of Retinal Images for Detection of Referable Diabetic Retinopathy

    Abràmoff, ; Folk MD, ; Han JC, ; Walker DP, ; Williams JD, ; Russell DF, et al. Automated Analysis of Retinal Images for Detection of Referable Diabetic Retinopathy. JAMA Ophthalmol. 2013;131: 351–357

    work page 2013

  32. [32]

    Feedback on a publicly distributed image database: The Messidor database

    Decencière E, Zhang X, Cazuguel G, Lay B, Cochener B, Trone C, et al. Feedback on a publicly distributed image database: The Messidor database. Image Anal Stereol. 2014;33: 231

    work page 2014

  33. [33]

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

    Kovalyk O, Morales-Sánchez J, Verdú-Monedero R, Sellés-Navarro I, Palazón-Cabanes A, Sancho-Gómez J-L. PAPILA: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Sci Data. 2022;9: 291

    work page 2022

  34. [34]

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

    Ahn JM, Kim S, Ahn K-S, Cho S-H, Lee KB, Kim US. Correction: A deep learning model for the detection of both advanced and early glaucoma using fundus photography. PLoS One. 2019;14: e0211579

    work page 2019

  35. [35]

    The UK Biobank resource with deep phenotyping and genomic data

    Bycroft C, Freeman C, Petkova D, Band G, Elliott LT, Sharp K, et al. The UK Biobank resource with deep phenotyping and genomic data. Nature. 2018;562: 203–209

    work page 2018

  36. [36]

    AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline

    Zhou Y, Wagner SK, Chia MA, Zhao A, Woodward-Court P, Xu M, et al. AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline. Transl Vis Sci Technol. 2022;11: 12

    work page 2022

  37. [37]

    PaliGemma: A versatile 3B VLM for transfer

    Beyer L, Steiner A, Pinto AS, Kolesnikov A, Wang X, Salz D, et al. PaliGemma: A versatile 3B VLM for transfer. arXiv [cs.CV]. 2024. Available: http://arxiv.org/abs/2407.07726

    Pith/arXiv arXiv 2024

  38. [38]

    Qwen Technical Report

    Bai J, Bai S, Chu Y, Cui Z, Dang K, Deng X, et al. Qwen Technical Report. arXiv [cs.CL]. 2023. Available: http://arxiv.org/abs/2309.16609

    Pith/arXiv arXiv 2023

  39. [39]

    Dino U-Net: Exploiting high-fidelity dense features from foundation models for medical image segmentation

    Gao Y, Li H, Yuan F, Wang X, Gao X. Dino U-Net: Exploiting high-fidelity dense features from foundation models for medical image segmentation. arXiv [cs.CV]. 2025. Available: http://arxiv.org/abs/2508.20909 Supplementary materials Supplementary Tables: https://docs.google.com/spreadsheets/d/1eRD44imIhicHehRrP7w7LBSHj8ztLEFv/edit?usp=sharing&ouid=115097033...

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.