Pith. sign in

REVIEW 2 major objections 5 minor 37 references

This paper argues that in frozen patch-based MIL for UWF retinal images, the pretraining objective matters: supervised and self-distillation ViTs transfer much better than MAE, with DINOv3 and DINOv1 reaching QWK ≈ 0.86 versus MAE's 0.53, a

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 ↗

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.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful benchmark for frozen encoder choice in UWF MIL, with a large DINO/supervised vs MAE gap on five-class DR, but the pooling confound and an overreaching abstract keep it from being a clean accept. the 2 major comments →

arxiv 2608.00586 v1 pith:AYAS4ZVG submitted 2026-08-01 cs.CV

Representation Transfer of Foundation Models for Ultra-Widefield Retinal Imaging

Mingya Alexa Gong , Da Ma , Lovre Antonio Budimir , Ivana Matovinovic , Sven Loncaric , Myeong Jin Ju , Yukun Zhou , Siegfried K. Wagner
show 2 more authors
Pearse A. Keane Marinko V. Sarunic
This is my paper
classification cs.CV
keywords ultra-widefield retinal imagingdiabetic retinopathy gradingmultiple instance learningvision transformerself-supervised pretrainingmasked autoencoderDINO self-distillationrepresentation transfer
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

The paper asks why one foundation model might work better than another when used as a frozen feature extractor for medical images. In a controlled patch-based multiple instance learning (MIL) setup on 10,404 ultra-widefield retinal images, it compares ViT-B encoders pretrained on ImageNet with supervised, masked autoencoder (MAE), and self-distillation objectives. The result: supervised and self-distillation (DINOv1) encoders vastly outperform MAE for five-class diabetic retinopathy grading (QWK 0.84–0.86 vs 0.53), and the same ranking appears on two additional UWF tasks. A larger-scale DINOv3 model performs best overall (QWK 0.863) but statistically ties DINOv1. Partial fine-tuning of the last encoder block nearly erases MAE's disadvantage, showing the gap is specific to frozen transfer.

Core claim

Within one fixed MIL pipeline—25 overlapping 224×224 patches per image, attention-based aggregation, frozen encoder—the choice of pretraining objective changes both the quality and the behavior of transferred representations. On MMRDR, ViT-MAE(ImageNet) achieves QWK 0.5325 while ViT-Supervised reaches 0.8396, ViT-DINOv1 0.8602, and ViT-DINOv3 0.8629, a margin the paper attributes to the objective rather than architecture, capacity, or downstream module. The MIL attention also diverges: MAE features yield near-uniform attention, effectively reducing attention pooling to mean pooling, while supervised and DINO features concentrate weight on central retinal patches. Unfreezing only the final en

What carries the argument

The load-bearing setup is patch-based MIL with frozen encoders: each UWF image is tiled into a 5×5 grid of 224×224 patches, each patch is embedded by a pretrained ViT, a fixed 2D positional encoding is added, and an attention-based MIL pooling layer (Ilse et al.) aggregates patch embeddings into a bag representation for classification. The encoders are the only thing varied; the comparison isolates pretraining objective (supervised, MAE reconstruction, self-distillation) by using ViT-B ImageNet checkpoints, plus DINOv3 and a domain-specific MAE. The feature pooling per encoder (CLS vs average) is chosen according to the pretraining objective, which is the mechanism's main uncontrolled knob.

Load-bearing premise

The main comparison assumes that using each model's recommended patch-pooling method (CLS token for supervised and DINOv1, average pooling for MAE and DINOv3) does not itself create the performance gap; since only MAE and DINOv3 were tested with both, the objective effect could be partly a pooling effect.

What would settle it

On the same MMRDR MIL setup, repeat the comparison with identical pooling (e.g., average pooling over all patch tokens) and identical preprocessing for every encoder. If MAE's QWK rises from 0.53 toward the supervised/DINO range while the others stay near 0.86, the claimed pretraining-objective effect is largely a pooling or preprocessing artifact; if MAE stays near 0.53, the effect is real.

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

If this is right

  • For frozen-feature medical imaging pipelines, selecting a supervised or self-distillation pretrained ViT over a MAE encoder can be worth about 0.3 QWK in DR grading before any fine-tuning.
  • Reconstruction-based encoders should not be dismissed: with a single unfrozen block, MAE(ImageNet) reaches frozen-supervised-level performance, suggesting a fine-tuning budget can substitute for pretraining choice.
  • MIL attention behavior is representation-dependent: MAE collapses to uniform weights, so attention-based MIL confers little benefit with such features.
  • Patch-based attention aggregation beats global and max pooling for the same encoder, supporting the use of high-resolution patch evidence in UWF imaging.
  • DINOv1 and DINOv3 perform comparably despite DINOv3's much larger pretraining data, so scale alone did not translate into a large frozen-transfer gain.

Where Pith is reading between the lines

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

  • Because pooling strategy is not held constant across objectives, a fair re-test would use identical token pooling for all encoders; the paper's own sensitivity data for MAE and DINOv3 show pooling can move results by roughly 0.1 QWK, which is smaller than the objective gap but not negligible.
  • The common grayscale preprocessing and 1024×1024 downsampling may favor some pretraining objectives over others, so the ranking should be rechecked at native Optomap resolution and with each checkpoint's native preprocessing.
  • The DINOv1-versus-DINOv3 tie hints that for medical patch features, the marginal value of larger self-supervised pretraining data is small relative to the objective family, perhaps because the downstream task rewards fine-grained local discriminability more than global semantic richness.
  • The attention statistics themselves could serve as a cheap diagnostic for whether a frozen representation is usable in MIL: near-uniform attention weights may indicate the encoder's patch features are not locally discriminative enough.
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

2 major / 5 minor

Summary. The paper evaluates frozen Vision Transformer (ViT) encoders pretrained with supervised, masked autoencoding (MAE), and self-distillation (DINOv1/DINOv3) objectives inside a patch-based, attention-based multiple instance learning (MIL) framework for ultra-widefield retinal imaging. On the MMRDR five-class diabetic retinopathy grading task, supervised and DINO-based ViTs reach QWK around 0.84–0.86, while ImageNet-MAE reaches QWK 0.53; DINOv3 is best overall (QWK 0.8629). The authors also report binary DR and intraocular-tumour transfer results, attention-pooling statistics, UMAP visualisations, and partial fine-tuning experiments. The central claim is that the choice of pretraining objective substantially influences frozen representation transfer and patch-aggregation behaviour within MIL, with reconstruction-based MAE representations transferring worst.

Significance. The question addressed is practically important: many medical-imaging pipelines use frozen ImageNet or large-scale self-supervised ViTs as feature extractors, and the interaction between pretraining strategy and weakly supervised MIL aggregation is underexplored. The study's strengths are its use of public datasets and checkpoints, three-seed repeats, a second DR dataset, a cross-disease dataset, a pooling sensitivity analysis, and a candid limitations section. If the attribution to pretraining objective is supported, the result gives valuable guidance for encoder selection in high-resolution UWF imaging. The main weakness is that the controlled ImageNet-ViT-B comparison confounds pretraining objective with the per-encoder token pooling convention, so the causal claim is not fully identified as stated.

major comments (2)
  1. [III.C.2, Appendix C, Table II, Eq. (4)] The controlled ImageNet ViT-B comparison confounds pretraining objective with token pooling. ViT-Supervised and ViT-DINOv1 use CLS-token features, while ViT-MAE uses average pooling over spatial tokens (Appendix C). Thus the large MMRDR gap (DINOv1 QWK 0.8602 vs MAE 0.5325, Table II) is not attributable to pretraining objective alone. Appendix Table I provides a sensitivity analysis only for ViT-MAE and ViT-DINOv3; it does not test ViT-Supervised or ViT-DINOv1 under average pooling, which are two of the three encoders in the central controlled comparison. Because the MIL attention mechanism in Eq. (4) receives different input distributions for CLS vs average-pooled features, the Fig. 5 attention-concentration claims are entangled with the same confound. The MAE disadvantage is plausibly robust (DINOv3 with CLS is also strong), but the causal attribution to pretraining objective requires
  2. [Abstract and Section IV.A.2, Table II] The abstract and conclusion state that supervised and self-distillation-based models outperform MAE, but Table II shows that on the DeepDRiD binary DR task ViT-MAE(ImageNet) achieves AUROC 0.8131, essentially tied with ViT-DINOv1 (0.8136). The ordering is task-dependent, as the Discussion itself acknowledges. The summary-level generalisation overstates the MMRDR-specific ranking and should be qualified, e.g., 'on five-class DR grading' or 'in most settings evaluated here.'
minor comments (5)
  1. [Section IV.A.1 / Table I] The representation-strategy comparison uses a reduced protocol (instance dropout disabled, patience 5, single run) as stated in Section III.C.1. The conclusion that attention pooling is the 'most informative' strategy would be stronger with multiple seeds or a caveat that the comparison is preliminary.
  2. [Section VI / Table II] The intraocular tumour dataset has no patient-level identifiers, so the same eye may appear across cross-validation folds. The authors disclose this in Limitations, but the potential for inflated AUROC should also be flagged where the tumour results are first reported in Table II, since the paper uses them as cross-disease validation.
  3. [Appendix B] Consider adding a reproducibility statement with a code repository link or explicit instructions for reproducing the checkpoints, preprocessing, and MIL training loop. The current text names timm and torchvision but no project URL.
  4. [Section III.B.1] The two positional encodings (the encoder's internal ViT patch-position encoding and the added 5x5 bag-level sine-cosine grid encoding) are described only briefly. A sentence clarifying that the bag-level grid is added after encoder feature extraction would reduce potential confusion.
  5. [Figure 5] The quantitative attention statistics are computed from unnormalised softmax weights before min-max visualisation. Please specify how many test images are aggregated and whether the reported values are per-image means or medians, to make the figure reproducible.

Circularity Check

0 steps flagged

No significant circularity: the central claims rest on external test-set measurements with public checkpoints and datasets, not on self-citation or fitted inputs.

full rationale

This is an empirical evaluation paper, not a derivation. The main claim—that pretraining objective affects frozen-representation transfer in a patch-based MIL framework—is supported by measuring QWK/AUROC on held-out test splits of public datasets (MMRDR, DeepDRiD, and the Sun et al. tumour dataset) using externally released checkpoints (timm/torchvision for MAE, DINOv1, DINOv3, supervised ViT, ResNet). No parameter is fitted to a subset of the evaluation data and then reported as a prediction of that same data; the MIL head is trained on the training split and evaluated on unseen test images, with fixed splits and seeds. The domain-specific AlzEye MAE is trained by the authors, but the paper explicitly and repeatedly warns that it is not part of the controlled pretraining-objective comparison (see Section III.C.2 and the Limitations section), so it is not used to force the central conclusion. The strongest potential concern is the co-variation of token pooling strategy with pretraining objective (CLS for supervised/DINOv1, average pooling for MAE/DINOv3), which is a genuine identifiability limitation. However, this is a validity/confounding issue, not a circular one: the comparison does not reduce by definition to its inputs, no equation equates the output with a fitted quantity, and the authors include a sensitivity analysis for MAE and DINOv3 in Appendix Table I. The attention-concentration observations are empirical summaries of the learned MIL weights, not predictions derived from the pretraining objectives by construction. Consequently, there is no circular step under the stated criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

No theoretical constants are fitted. The paper is an empirical benchmark; all reported numbers are test-set measurements. The key modeling choices (patch grid, overlap, instance dropout, learning rates, pooling per encoder) are experimental settings, not free parameters fit to the test set. The main assumptions are domain assumptions about preprocessing and the validity of public checkpoints.

axioms (3)
  • domain assumption Pretrained checkpoints faithfully represent their pretraining objectives (supervised, MAE, DINO) and pooling conventions
    The comparison assumes differences between checkpoints are due to objective, not checkpoint quality or implementation quirks.
  • domain assumption Resizing UWF images to 1024x1024 and converting to grayscale preserves the task-relevant features for DR grading and tumour detection
    Downsampling and grayscale conversion are applied uniformly; if these destroy peripheral lesion information, the measured effects could shift.
  • domain assumption Attention-based MIL pooling (Ilse et al.) is an appropriate aggregation mechanism and its learned attention weights reflect meaningful patch importance
    The framework treats attention weights as evidence of aggregation behavior; this is standard in MIL but not externally validated here.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Representation Transfer of Foundation Models for Ultra-Widefield Retinal Imaging." pith.science (2026). https://pith.science/paper/AYAS4ZVG

@misc{pith2026260800586,
  author       = {Pith},
  title        = {Pith review of: Representation Transfer of Foundation Models for Ultra-Widefield Retinal Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AYAS4ZVG}},
  note         = {Machine review of arXiv:2608.00586}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Despite the widespread adoption of foundation models as feature extractors for medical imaging, relatively little is understood about how different pretraining strategies influence the transferability of learned representations to weakly supervised ophthalmic imaging tasks. We investigate this question in ultra-widefield (UWF) retinal imaging by evaluating foundation model representations within a patch-based multiple instance learning (MIL) framework for disease classification on UWF images. We compare Vision Transformer encoders pretrained with supervised, Masked Autoencoder (MAE), and self-distillation objectives, while keeping the downstream aggregation architecture unchanged. Within a controlled comparison of ViT-B encoders pretrained on ImageNet-1k, the choice of pretraining objective substantially influenced frozen representation transfer, with supervised and self-distillation-based models outperforming MAE. A contemporary DINOv3 model pretrained at a larger scale achieved the strongest overall performance, with a quadratic weighted kappa of 0.863 for five-class diabetic retinopathy grading, comparable with DINOv1. Attention analysis further revealed distinct patch-aggregation behaviours associated with the different pretrained representations, while partial fine-tuning substantially reduced the performance gap for MAE. These findings suggest that pretraining strategy influences both representation transferability and the subsequent aggregation of patch-level evidence within MIL, resulting in differences in downstream classification performance.

Figures

Figures reproduced from arXiv: 2608.00586 by Da Ma, Ivana Matovinovic, Lovre Antonio Budimir, Marinko V. Sarunic, Mingya Alexa Gong, Myeong Jin Ju, Pearse A. Keane, Siegfried K. Wagner, Sven Loncaric, Yukun Zhou.

Figure 1
Figure 1. Figure 1: FIGURE 1 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIGURE 3 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIGURE 4 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIGURE 5 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIGURE 6 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIGURE 7 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    W. Khan, S. Leem, K. B. See, J. K. Wong, S. Zhang, and R. Fang, ‘‘A comprehensive survey of foundation models in medicine,’’IEEE Rev. Biomed. Eng., 2025

    work page 2025

  2. [2]

    Jiaoet al., ‘‘Foundation model for medical imaging: A comprehen- sive review,’’IEEE Trans

    L. Jiaoet al., ‘‘Foundation model for medical imaging: A comprehen- sive review,’’IEEE Trans. Artif. Intell., 2025

    work page 2025

  3. [3]

    M. Huh, P . Agrawal, and A. A. Efros, ‘‘What makes ImageNet good for transfer learning?,’’arXiv preprint arXiv:1608.08614, 2016

    Pith/arXiv arXiv 2016

  4. [4]

    Dosovitskiyet al., ‘‘An image is worth 16×16 words: Transformers for image recognition at scale,’’ inProc

    A. Dosovitskiyet al., ‘‘An image is worth 16×16 words: Transformers for image recognition at scale,’’ inProc. Int. Conf. Learn. Represent. (ICLR), 2021

    work page 2021

  5. [5]

    K. He, X. Chen, S. Xie, Y . Li, P . Dollár, and R. Girshick, ‘‘Masked autoencoders are scalable vision learners,’’ inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022

    work page 2022

  6. [6]

    Caronet al., ‘‘Emerging properties in self-supervised vi- sion transformers,’’ inProc

    M. Caronet al., ‘‘Emerging properties in self-supervised vi- sion transformers,’’ inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Montreal, QC, Canada, Oct. 2021, pp. 9630–9640, doi: 10.1109/ICCV48922.2021.00951

    arXiv 2021

  7. [7]

    Siméoniet al., ‘‘DINOv3,’’arXiv preprint arXiv:2508.10104, 2025

    O. Siméoniet al., ‘‘DINOv3,’’arXiv preprint arXiv:2508.10104, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Zhouet al., ‘‘A foundation model for generalizable disease detection from retinal images,’’Nature, vol

    Y . Zhouet al., ‘‘A foundation model for generalizable disease detection from retinal images,’’Nature, vol. 622, no. 7981, pp. 156–163, Oct. 2023, doi: 10.1038/s41586-023-06555-x

    doi:10.1038/s41586-023-06555-x 2023

  9. [9]

    Generalist versus Specialist Vision Foundation Models for Ocular Disease and Oculomics

    Y . Zhouet al., ‘‘Generalist versus specialist vision foundation models for ocular disease and oculomics,’’arXiv preprint arXiv:2509.03421, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    A. S. Neubauer, M. Kernt, C. Haritoglou, S. G. Priglinger, A. Kampik, and M. W. Ulbig, ‘‘Nonmydriatic screening for diabetic retinopathy by ultra-widefield scanning laser ophthalmoscopy (Optomap),’’Graefes Arch. Clin. Exp. Ophthalmol., vol. 246, no. 2, pp. 229–235, Feb. 2008, doi: 10.1007/s00417-007-0631-4

    work page doi:10.1007/s00417-007-0631-4 2008

  11. [11]

    Talebi and P

    H. Talebi and P . Milanfar, ‘‘Learning to resize images for computer vision tasks,’’ inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 497–506

    work page 2021

  12. [12]

    Maron and T

    O. Maron and T. Lozano-Pérez, ‘‘A framework for multiple-instance learning,’’ inAdv. Neural Inf. Process. Syst., vol. 10, 1997

    work page 1997

  13. [13]

    Carbonneau, V

    M.-A. Carbonneau, V . Cheplygina, E. Granger, and G. Gagnon, ‘‘Mul- tiple instance learning: A survey of problem characteristics and ap- 14 Gonget al.: Preprint TABLE II.Full performance metrics for five-class DR classification on MMRDR. Values are reported as mean±standard deviation across three runs. Best results are shown in bold. Model QWK AUROC F1 AC...

    work page arXiv 1981

  14. [14]

    Gadermayr and M

    M. Gadermayr and M. Tschuchnig, ‘‘Multiple instance learning for digital pathology: A review of the state-of-the-art, limitations, and future potential,’’Comput. Med. Imaging Graph., vol. 112, Art. no. 102337, Mar. 2024, doi: 10.1016/j.compmedimag.2024.102337

    arXiv 2024

  15. [15]

    Liet al., ‘‘A multi-resolution model for histopathology im- age classification and localization with multiple instance learning,’’ Comput

    J. Liet al., ‘‘A multi-resolution model for histopathology im- age classification and localization with multiple instance learning,’’ Comput. Biol. Med., vol. 131, Art. no. 104253, Apr. 2021, doi: 10.1016/j.compbiomed.2021.104253

    work page arXiv 2021

  16. [16]

    Z. Su, M. Rezapour, U. Sajjad, M. N. Gurcan, and M. K. K. Ni- azi, ‘‘Attention2Minority: A salient instance inference-based multi- ple instance learning for classifying small lesions in whole slide im- ages,’’Comput. Biol. Med., vol. 167, Art. no. 107607, Dec. 2023, doi: 10.1016/j.compbiomed.2023.107607

    work page arXiv 2023

  17. [17]

    Muelleret al., ‘‘Multiple instance learning detects peripheral arterial disease from high-resolution color fundus photography,’’Sci

    S. Muelleret al., ‘‘Multiple instance learning detects peripheral arterial disease from high-resolution color fundus photography,’’Sci. Rep., vol. 12, no. 1, Art. no. 1389, Jan. 2022, doi: 10.1038/s41598-022-05169-z

    work page doi:10.1038/s41598-022-05169-z 2022

  18. [18]

    Tanget al., ‘‘A multimodal retinal image dataset for diabetic retinopathy detection using foundation models,’’Sci

    Z. Tanget al., ‘‘A multimodal retinal image dataset for diabetic retinopathy detection using foundation models,’’Sci. Data, Mar. 2026, doi: 10.1038/s41597-026-07005-9

    work page doi:10.1038/s41597-026-07005-9 2026

  19. [19]

    S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, ‘‘Transformers in vision: A survey,’’ACM Comput. Surv., vol. 54, no. 10s, pp. 1–41, 2022

    work page 2022

  20. [20]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, ‘‘ImageNet: A large-scale hierarchical image database,’’ inProc. IEEE Conf. Com- put. Vis. Pattern Recognit. (CVPR), Miami, FL, USA, Jun. 2009, pp. 248–255, doi: 10.1109/CVPR.2009.5206848

    arXiv 2009

  21. [21]

    Guiet al., ‘‘A survey on self-supervised learning: Algorithms, ap- plications, and future trends,’’IEEE Trans

    J. Guiet al., ‘‘A survey on self-supervised learning: Algorithms, ap- plications, and future trends,’’IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 12, pp. 9052–9071, 2024

    work page 2024

  22. [22]

    Qiuet al., ‘‘Development and validation of a multimodal multitask vi- sion foundation model for generalist ophthalmic artificial intelligence,’’ NEJM AI, vol

    J. Qiuet al., ‘‘Development and validation of a multimodal multitask vi- sion foundation model for generalist ophthalmic artificial intelligence,’’ NEJM AI, vol. 1, no. 12, Art. no. AIoa2300221, 2024

    work page 2024

  23. [23]

    Moranoet al., ‘‘Multimodal foundation model and benchmark for comprehensive retinal OCT image analysis,’’npj Digit

    J. Moranoet al., ‘‘Multimodal foundation model and benchmark for comprehensive retinal OCT image analysis,’’npj Digit. Med., vol. 8, no. 1, Art. no. 576, 2025

    work page 2025

  24. [24]

    S. Y uet al., ‘‘MIL-VT: Multiple instance learning enhanced vision transformer for fundus image classification,’’ inMedical Image Com- puting and Computer Assisted Intervention–MICCAI 2021, Lecture Notes in Computer Science, vol. 12908. Cham, Switzerland: Springer, Sep. 2021, pp. 45–54, doi: 10.1007/978-3-030-87237-3_5

    work page doi:10.1007/978-3-030-87237-3_5 2021

  25. [25]

    Biet al., ‘‘MIL-ViT: A multiple instance vision transformer for fundus image classification,’’J

    Q. Biet al., ‘‘MIL-ViT: A multiple instance vision transformer for fundus image classification,’’J. Vis. Commun. Image Represent., vol. 97, Art. no. 103956, Dec. 2023, doi: 10.1016/j.jvcir.2023.103956

    work page arXiv 2023

  26. [26]

    Y ang, Z

    Y . Y ang, Z. Cai, S. Qiu, and P . Xu, ‘‘A novel transformer model with multiple instance learning for diabetic retinopathy classifica- tion,’’IEEE Access, vol. 12, pp. 6768–6776, 2024, doi: 10.1109/AC- CESS.2024.3351473

    arXiv 2024

  27. [27]

    Liuet al., ‘‘DeepDRiD: Diabetic retinopathy—grading and image quality estimation challenge,’’Patterns, vol

    R. Liuet al., ‘‘DeepDRiD: Diabetic retinopathy—grading and image quality estimation challenge,’’Patterns, vol. 3, no. 6, Art. no. 100512, Jun. 2022, doi: 10.1016/j.patter.2022.100512

    arXiv 2022

  28. [28]

    Sunet al., ‘‘An ultra-wide-field fundus image dataset for intelligent diagnosis of intraocular tumors,’’Sci

    J. Sunet al., ‘‘An ultra-wide-field fundus image dataset for intelligent diagnosis of intraocular tumors,’’Sci. Data, vol. 12, no. 1, Art. no. 1521, Aug. 2025, doi: 10.1038/s41597-025-05864-2

    work page doi:10.1038/s41597-025-05864-2 2025

  29. [29]

    Paszkeet al., ‘‘PyTorch: An imperative style, high-performance deep learning library,’’ inAdv

    A. Paszkeet al., ‘‘PyTorch: An imperative style, high-performance deep learning library,’’ inAdv. Neural Inf. Process. Syst., vol. 32, 2019

    work page 2019

  30. [30]

    M. Ilse, J. Tomczak, and M. Welling, ‘‘Attention-based deep multiple instance learning,’’ inProc. 35th Int. Conf. Mach. Learn. (ICML), Stockholm, Sweden, 2018, pp. 2127–2136

    work page 2018

  31. [31]

    K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for image recognition,’’ inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Las V egas, NV , USA, Jun. 2016, pp. 770–778

    work page 2016

  32. [32]

    S. V . Chilukoti, L. Shan, V . S. Tida, A. S. Maida, and X. Hei, ‘‘A reliable diabetic retinopathy grading via transfer learning and ensemble learning with quadratic weighted kappa metric,’’BMC Med. Inform. Decis. Mak., vol. 24, no. 1, Art. no. 37, Feb. 2024, doi: 10.1186/s12911- 024-02446-x

    doi:10.1186/s12911- 2024

  33. [33]

    S. K. Wagneret al., ‘‘AlzEye: Longitudinal record-level linkage of ophthalmic imaging and hospital admissions of 353 157 patients in London, UK,’’BMJ Open, vol. 12, no. 3, Art. no. e058552, Mar. 2022, doi: 10.1136/bmjopen-2021-058552

    work page doi:10.1136/bmjopen-2021-058552 2022

  34. [34]

    Engelmann and M

    J. Engelmann and M. O. Bernabeu, ‘‘Training a high-performance retinal foundation model with half-the-data and 400 times less com- pute,’’Nat. Commun., vol. 16, no. 1, Art. no. 6862, Jul. 2025, doi: 10.1038/s41467-025-62123-z

    doi:10.1038/s41467-025-62123-z 2025

  35. [35]

    V aswaniet al., ‘‘Attention is all you need,’’ inAdv

    A. V aswaniet al., ‘‘Attention is all you need,’’ inAdv. Neural Inf. Process. Syst., vol. 30, 2017

    work page 2017

  36. [36]

    Wightman, ‘‘PyTorch Image Models,’’ GitHub repository, 2019

    R. Wightman, ‘‘PyTorch Image Models,’’ GitHub repository, 2019. 15 Gonget al.: Preprint [Online]. Available: https://github.com/huggingface/pytorch-image- models, doi: 10.5281/zenodo.4414861

    doi:10.5281/zenodo.4414861 2019

  37. [37]

    [Online]

    TorchVision maintainers and contributors, ‘‘TorchVision: PyTorch’s computer vision library,’’ GitHub repository, 2016. [Online]. Available: https://github.com/pytorch/vision. 16

    work page 2016

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