FusionFM: Fusing Eye-specific Foundational Models for Optimized Ophthalmic Diagnosis

Ching-Yu Cheng; Gabor M. Somfai; Haoyu Chen; Huazhu Fu; Jocelyn Hui Lin Goh; Ke Zou; Meng Wang; Pearse A. Keane; Rui Santos; Sahana Srinivasan

REVIEW 5 major objections 4 minor 2 cited by

FusionFM: Fusing Eye-specific Foundational Models for Optimized Ophthalmic Diagnosis

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read FusionFM, a systematic benchmark of four ophthalmic foundation models, reports that DINORET and RetiZero lead on eye and systemic disease tasks, while gated fusion yields modest, task-specific gains.

desk verdict A useful benchmark question buried in an unreadable file: the abstract promises the first systematic comparison of four ophthalmic foundation models, but with the full text corrupted, the ranking is currently unverifiable. read the letter →

arxiv 2508.11721 v1 pith:GDCFCD6Q submitted 2025-08-15 cs.CV cs.AI

Ke Zou , Jocelyn Hui Lin Goh , Yukun Zhou , Tian Lin , Samantha Min Er Yew , Sahana Srinivasan , Meng Wang , Rui Santos

show 6 more authors

Gabor M. Somfai Huazhu Fu Haoyu Chen Pearse A. Keane Ching-Yu Cheng Yih Chung Tham

This is my paper · ORCID

classification cs.CVcs.AI

keywords ophthalmicfoundationmodelsretinalimagingmodelfusiongating-basedexternalvalidationsystemicdiseasepredictionglaucomadiabeticretinopathy

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 tries to answer a practical question the field has not settled: among current ophthalmic foundation models, which should a researcher or clinician actually use, and does combining them help? FusionFM benchmarks four pretrained retinal-image models on the same multi-country datasets for glaucoma, diabetic retinopathy, age-related macular degeneration, diabetes, and hypertension. It finds that DINORET and RetiZero outperform RETFound and VisionFM on most tasks, with RetiZero generalizing best to external cohorts. A gating-based fusion head adds modest improvements for glaucoma, AMD, and hypertension, but systemic disease prediction remains difficult, especially hypertension on external data. If correct, the study gives an evidence-based ranking and a reusable evaluation protocol for future ophthalmic foundation models.

What carries the argument

FusionFM is the central object: an evaluation suite that standardizes datasets and splits across four foundation models, plus two fusion approaches. The load-bearing fusion mechanism is the gating-based head, which learns per-sample weights over the feature vectors of RETFound, VisionFM, RetiZero, and DINORET; this is the component responsible for the reported gains on glaucoma, AMD, and hypertension. The standardized multi-country evaluation protocol is what makes the cross-model ranking meaningful.

What would settle it

Train or fine-tune each of the four foundation models on one shared public retinal dataset with a single fixed preprocessing pipeline and several random splits; if RETFound or VisionFM then match or beat DINORET and RetiZero, the reported ranking is an artifact of dataset handling rather than a property of the models. For the external-generalization claim, the test is whether RetiZero's advantage over DINORET reproduces on an unseen dataset from a different camera and ethnicity mix.

Watch

Extended reading notes

Core claim

The paper's central claim is that not all ophthalmic foundation models are interchangeable. When evaluated on standardized multi-country datasets with identical protocols, DINORET and RetiZero achieve superior AUC and F1 across both ophthalmic diseases and systemic diseases predicted from retinal images, while RETFound and VisionFM trail. RetiZero shows the strongest performance on external datasets. The paper also claims that a gating-based fusion of the four models' features yields small but consistent gains for glaucoma, age-related macular degeneration, and hypertension, indicating that the models carry complementary information even when their single-model rankings differ. The authors p

Load-bearing premise

The standardized multi-country datasets are labeled, preprocessed, and split equivalently for all four foundation models, so that AUC differences reflect model quality rather than data artifacts, and the external cohorts reflect the training-domain distribution closely enough for generalization claims to hold.

Editorial extensions

If this is right

A clinic or research group choosing a single ophthalmic foundation model would be better served by DINORET or RetiZero than by RETFound or VisionFM on the evaluated tasks.
RetiZero's external-cohort advantage makes it the safer default for deployment on images from unseen sites, devices, or populations.
Fusing models can recover a few additional points of AUC for glaucoma, AMD, and hypertension, but not enough to make systemic disease prediction clinically reliable on its own.
Systemic disease prediction from retinal images, especially hypertension on external data, remains an open problem even with fusion.
Future ophthalmic foundation model papers should report both within-dataset and external-cohort performance and compare against the models ranked here.

Reading between the lines

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

A reader should not read the modest fusion gains as proof that the individual models are redundant; a natural follow-up would inspect the learned gating weights per disease to see whether gains come from one dominant model or from genuine complementarity.
The ranking likely depends on how labels, preprocessing, and data splits are standardized across models; a testable extension would rerun the four models on one public dataset with a fixed preprocessing pipeline and multiple splits to see whether the DINORET/RetiZero advantage persists.
Because hypertension remains hardest on external data, the retinal signal for hypertension appears weak or site-specific; a follow-up could test whether domain adaptation or federated training on external cohorts closes the gap.
The paper's contribution may end up mattering more as a benchmarking protocol than as a deployable fused model, since the fusion improvements are small and task-specific.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: the paper asks the right questions—which ophthalmic FM is best, are they equally good across tasks, and does fusion help—and it claims an evidence-based answer. The abstract is clear and the framing is honest: DINORET and RetiZero come out ahead, fusion gives only modest gains, and systemic disease prediction stays hard. That is genuinely useful for model selection, and it is not over-hyped.

What is new: a simultaneous head-to-head of four FMs with standardized multi-country datasets and two fusion baselines. The fusion methods themselves are standard, but the systematic comparison is a real service to the field. The authors also deserve credit for reporting that fusion helps only in some tasks and that hypertension prediction remains poor; that kind of negative result is easy to bury.

The soft spots are serious but mostly unanswerable from what I can read. The provided full text is corrupted beyond the abstract, so I cannot check the tables, the protocol, or the confidence intervals. That is not the authors' fault in my copy, but it means the central ranking is currently an assertion, not a demonstration. The stress-test concern is on point: each FM has its own native resolution, normalization, and optimal fine-tuning recipe. If the authors applied one pipeline to all four, the ranking could reflect preprocessing matching rather than model quality. I cannot tell from the abstract whether they tuned each model fairly. Author overlap with some of the benchmarked FMs is another neutrality question, but that alone is not disqualifying—most benchmark papers come from groups with vested interests. The cure is transparency: protocol details, per-model hyperparameters, and ideally released evaluation code.

Bottom line: if the full paper shows comparable training setups and reports uncertainty, this is a solid empirical contribution that belongs in the literature. As it stands, I cannot verify the headline ranking, and I would not cite it yet. But it deserves a serious referee—someone should read the actual methods and tables, not desk-reject on the abstract. Send it to review; ask for the protocol and the code.

Referee Report

5 major / 4 minor

Summary. The paper proposes FusionFM, an evaluation suite and two fusion strategies (including a gating-based approach) for four ophthalmic foundation models (RETFound, VisionFM, RetiZero, DINORET). The authors state that they benchmark these models on standardized datasets from multiple countries for glaucoma, diabetic retinopathy, age-related macular degeneration, diabetes, and hypertension, reporting AUC and F1. The main claims are that DINORET and RetiZero achieve superior performance in both ophthalmic and systemic disease tasks, that RetiZero generalizes better on external datasets, and that gating fusion yields modest improvements in glaucoma, AMD, and hypertension, while systemic disease prediction—especially hypertension—remains challenging. The submitted full text is corrupted beyond section headers, so no methods, results tables, equations, or protocol details are readable. As a result, the evidence supporting the central claims cannot be verified from the manuscript as received.

Significance. If the claims hold, this would be the first systematic comparison and fusion study of ophthalmic foundation models and would provide useful, clinically relevant evidence for model selection. The authors are honest about the modest size of fusion gains and about persistent difficulty in systemic disease prediction, which is a strength. The study also covers multiple countries and external datasets, which is valuable if the datasets and splits are properly described. However, the current submission does not provide the evidence needed to assess the ranking: there are no readable tables, confidence intervals, per-model protocol details, or external cohort definitions. The potential contribution is real, but the manuscript in its present form is not evaluable.

major comments (5)

[Full text] The supplied manuscript text is corrupted beyond section headers; all tables, figures, equations, and result details are unreadable. The central claims—DINORET/RetiZero superiority, RetiZero's external generalization, and gating fusion gains—are empirical and require the full protocol and results. Please resubmit a readable manuscript with all tables, figure captions, equations, and appendices. Without this, no claim can be independently checked.
[Abstract / Methods] The abstract reports that datasets were 'standardized,' but the full text does not describe the preprocessing, input resolution, normalization, augmentation, fine-tuning schedule, learning rates, number of epochs, frozen layers, or classifier heads used for each foundation model. These four models have different native input specifications and training pipelines; if a single preprocessing pipeline was applied to all, models whose pretraining matches that pipeline will be favored, and models expecting different preprocessing will be penalized. The paper must report per-model protocol details or justify that the chosen pipeline is equally fair to all models. Otherwise the observed AUC gaps may be artifacts of protocol mismatch.
[Abstract / Results] The abstract reports AUC and F1 values without confidence intervals or significance tests. The claims of 'superior performance' and 'stronger generalization' are comparative; the paper should provide bootstrap confidence intervals or paired significance tests (e.g., DeLong tests) for each task and dataset, with appropriate multiple-comparison control. Since the fusion gains are described as modest, statistical reliability is essential to distinguish genuine improvements from noise.
[External generalization claim] The claim that RetiZero exhibits stronger generalization on external datasets requires a clear definition of the external cohorts: acquisition device, image resolution, population ethnicity, disease severity, label noise, and potential overlap with pretraining data. The reader cannot assess whether the external cohorts match RetiZero's training distribution more closely than the other models' distributions. Please report per-cohort AUC/F1 for every model and describe the external datasets in sufficient detail.
[Competing interests / Evaluation neutrality] Several co-authors are also developers of some of the benchmarked foundation models. This does not by itself invalidate the study, but it is a load-bearing neutrality concern. The manuscript should disclose these relationships and describe safeguards: a fixed evaluation protocol, blinded or third-party execution, publicly released code, or a pre-registered analysis. As written, the comparison is not demonstrably independent, and the reader cannot exclude protocol-selection bias.

minor comments (4)

[Abstract] The phrase 'first study to systematically evaluate...' should be supported by a related-work section that explicitly discusses prior comparative evaluations of ophthalmic FMs; if none exist, say so with citations to the closest works.
[Abstract / Code availability] FusionFM is described as a 'comprehensive evaluation suite,' but no code or data availability link is given. If the suite is intended for community use, include a repository URL.
[Corrupted text] Some equations and fragments appear in the corrupted text but cannot be read. Ensure all equations are rendered correctly in the resubmission and that notation for the gating fusion is defined precisely.
[Results / Fusion gains] Where fusion improvements are reported, specify the baseline: improvement over the best single model or over the average of all models, and report the absolute AUC/F1 differences rather than only a qualitative 'modest' characterization.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the ranking is an empirical benchmark result, not a derivation from fitted inputs or self-cited theorems.

full rationale

The manuscript is a comparative empirical study: it benchmarks four released ophthalmic foundation models on standardized datasets and evaluates two fusion approaches. The central claims (DINORET/RetiZero achieve superior AUC/F1, RetiZero generalizes better externally, gating gives modest gains) are empirical outcomes contingent on datasets, preprocessing, and fine-tuning protocol. No derivational chain is present that could reduce to a fitted input or to a definitionally equivalent quantity: the paper does not define any model quality measure in terms of the outputs it later reports, nor does it fit a parameter to a subset and call the related quantity a prediction. The abstract and the recoverable text do not exhibit any equation-level identity of the form "prediction = input by construction." The only concern evident from the author list is non-independence, since some co-authors also developed benchmarked models; however, that is a neutrality/fairness issue, not circularity of reasoning, and the corrupted full text prevents any specific verification of protocol fairness. That unverifiability is a correctness/risk concern, not a circularity finding. No load-bearing self-citation chain or imported uniqueness theorem is recoverable from the provided text. Under the rules requiring quote and specific reduction, no circular step can be identified.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities; it relies on existing FMs and learned fusion weights. The main ledger items are the fitted gating parameters and task heads, plus the implicit assumptions that the datasets are comparable and that frozen embeddings are fair basis for comparison.

free parameters (2)

Gating fusion weights = learned per task on validation folds
The gating-based fusion approach trains per-task weights that determine how much each FM contributes; the reported fusion improvements are conditional on these fitted weights.
Linear probe / task classifier heads = trained per dataset and task
Each FM embedding is decoded by a separately trained classifier; per-task AUC/F1 depends on these heads, and no cross-validation details are given in the abstract.

assumptions (3)

domain assumption Frozen FM embeddings from different models are directly comparable when fed into the same task heads.
The entire benchmark assumes each FM can be treated as a fixed feature extractor and that embedding scale or distribution differences do not distort the comparison.
domain assumption The 'standardized datasets from multiple countries' have consistent labels and representative disease distributions.
External generalization claims in the abstract depend on label harmonization and similar prevalence/severity across datasets.
domain assumption No significant train-test overlap exists between the pretraining data of the four FMs and the evaluation datasets.
If any FM was pretrained on data that includes the test sets, its AUC would be inflated; the abstract does not report any overlap analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FusionFM: Fusing Eye-specific Foundational Models for Optimized Ophthalmic Diagnosis." pith.science (2026). https://pith.science/paper/GDCFCD6Q

@misc{pith2026250811721,
  author       = {Pith},
  title        = {Pith review of: FusionFM: Fusing Eye-specific Foundational Models for Optimized Ophthalmic Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GDCFCD6Q}},
  note         = {Machine review of arXiv:2508.11721}
}

read the original abstract

Foundation models (FMs) have shown great promise in medical image analysis by improving generalization across diverse downstream tasks. In ophthalmology, several FMs have recently emerged, but there is still no clear answer to fundamental questions: Which FM performs the best? Are they equally good across different tasks? What if we combine all FMs together? To our knowledge, this is the first study to systematically evaluate both single and fused ophthalmic FMs. To address these questions, we propose FusionFM, a comprehensive evaluation suite, along with two fusion approaches to integrate different ophthalmic FMs. Our framework covers both ophthalmic disease detection (glaucoma, diabetic retinopathy, and age-related macular degeneration) and systemic disease prediction (diabetes and hypertension) based on retinal imaging. We benchmarked four state-of-the-art FMs (RETFound, VisionFM, RetiZero, and DINORET) using standardized datasets from multiple countries and evaluated their performance using AUC and F1 metrics. Our results show that DINORET and RetiZero achieve superior performance in both ophthalmic and systemic disease tasks, with RetiZero exhibiting stronger generalization on external datasets. Regarding fusion strategies, the Gating-based approach provides modest improvements in predicting glaucoma, AMD, and hypertension. Despite these advances, predicting systemic diseases, especially hypertension in external cohort remains challenging. These findings provide an evidence-based evaluation of ophthalmic FMs, highlight the benefits of model fusion, and point to strategies for enhancing their clinical applicability.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

Generalist versus Specialist Vision Foundation Models for Ocular Disease and Oculomics
eess.IV 2025-09 conditional novelty 5.0 of 10

On ten retinal imaging tasks, retina-trained RETFound-DINOv2 edged out larger generalist DINOv2/DINOv3 models, but the average advantage was small.
Uncertainty-Aware Last-Layer Adaptation of RETFound for Referable Diabetic Retinopathy Screening Under Dataset Shift
cs.CV 2026-06 conditional novelty 3.5 of 10

Last-layer Bayesian and SNGP heads on frozen RETFound improve APTOS selective-referral safety, but false-negative gains are not unique to Bayesian modeling and SNGP fails under APTOS-to-DDR transfer.

Reference graph

Works this paper leans on

1 extracted references · 1 linked inside Pith · cited by 2 Pith papers

[1]

�� ...

arXiv 2025

Pith tools

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

[1] [1]

�� ...

arXiv 2025