Photographs of the back of the eye show small blood vessels that share their developmental origin and much of their physiology with the vessels of the brain. Deep networks trained on these photographs can flag people at risk of stroke, and of the "silent" brain infarcts that only MRI normally reveals. The data that would make such models trustworthy across populations sit in hospitals and biobanks in different countries, and privacy law prevents pooling them. Decentralized training, where hospitals average their models peer-to-peer with no trusted central server, is the natural fit. It has a known weakness, though. With no central referee, every peer has to decide for itself which incoming model updates to accept. The standard test asks whether an update points in roughly the same direction as the consensus, and that test cannot tell a malicious hospital from an honest one whose camera or patient population differs. For retinal images the confusion is severe, because models trained on raw fundus images readily learn camera and site quirks.
We propose to move the acceptance test out of weight space and into vessel space. Every hospital trains a small model whose input is the extracted graph of retinal vessels (junctions as nodes, vessel segments as edges) rather than the raw photograph. Before an update is admitted, each peer runs it on a shared public panel of vessel graphs and checks whether its stroke-risk predictions agree with those of the other peers. Converting to vessel graphs strips out most camera-specific signal, so honest hospitals should agree closely on the panel. Disagreement then becomes evidence of poisoning rather than of optics. The claim is testable: under a fixed poisoning budget, screening on vessel-graph predictions should wrongly exclude far fewer honest minority sites than weight-cosine screening, and it should block attacks at least as well. The whole check stays cheap and deterministic enough to run on hospital hardware and record on a shared ledger.
Core Claim
Two findings in the file motivate the idea. Feature Isolation for Hypothesis Testing in Retinal Imaging trained stroke classifiers on full fundus images and on U-Net vessel maps alone. It found that the full-image AUC of 0.987 was driven by camera and site artifacts, while the vessel-only model's AUC of 0.754 reflected generalizable vascular signal. On Biases in a UK Biobank-based Retinal Image Classification Model showed a 10% AUC gap between assessment centers and a 15% gap between age groups, even with identical cameras and protocols. Together these say that honest hospital updates on raw images genuinely diverge. A defense that throws away divergent updates, such as the cosine-similarity filter of PPBFL (which flags updates with negative cosine to the global update, evaluated on OCTMNIST and PathMNIST), or the sign-consistency mask of Fair-FedMOE, will therefore be most likely to discard exactly the elderly-heavy or unusual cohorts that carry the most stroke signal.
A third paper supplies the model. Predicting Stroke through Retinal Graphs and Multimodal Self-supervised Learning showed that a graph attention network on vessel topology, aligned with clinical tabular data, predicts stroke on 82,569 UK Biobank participants at 71.92% AUROC. That is slightly above a raw-image ResNet50 at 71.54%, with 3.4M instead of 28.2M parameters. A model that small makes peer-to-peer weight exchange practical, and its input space is the one in which honest sites should agree.
The screening rule is simple. Let $R$ be a public reference panel of vessel graphs and $f_i$ the model proposed by site $i$ in a round. Site $i$'s disagreement score is
$$d_i = \operatorname{median}_{j \neq i} \frac{1}{|R|} \sum_{r \in R} \left| f_i(r) - f_j(r) \right|,$$
the typical gap between site $i$'s predicted risk and each peer's predicted risk, averaged over the panel. An update is admitted if $d_i$ lies within a robust multiple of the median absolute deviation of all $d_j$. We use the median over peers so that a minority of colluding attackers cannot drag the reference point. The panel is built from the public artery/vein-annotated datasets named in the file (AV-DRIVE, LES-AV, HRF, MAGIC, plus DRIVE and CHASE_DB1 via the SGL vessel segmentation code). It is expanded to several thousand graphs by subgraph sampling and caliber jitter, and it is stratified so that no single vessel morphology dominates. The panel carries no stroke labels, so the screen tests whether peers agree, not whether they are accurate. Accuracy is left to local validation.
This is different from what already exists at the intersection. Federated Learning in Ocular Imaging reviews server-based federation for eye diseases and does not treat decentralization or cerebral outcomes. PPBFL is decentralized and runs on OCT data, but it screens raw weights under costly multi-party computation and targets four intra-ocular disease classes. Fair-FedMOE and FairFedMed fix demographic gaps for glaucoma and diabetic retinopathy through a central server, and both need demographic labels. The swarm learning studies in histopathology and breast MRI show that decentralized training can match pooled training. Neither addresses poisoning, and neither involves the retina. No prior work defines the robustness test on biologically constrained inputs.
Research Program
The pipeline has three parts. The first is local feature extraction. Each site segments arteries and veins with a shared, frozen segmenter and converts the result to a vessel graph, following the construction in the retinal-graph stroke paper. Freezing the segmenter matters: if it were trained jointly, it could reintroduce site-specific signal into the graphs.
The second is local training of the small graph model with the clinical-tabular branch, in rounds.
The third is the peer protocol. Each round, a site publishes three things: its model weights, a hash of those weights, and its predictions on the reference panel. Every peer computes all the $d_i$ scores identically. A smart contract on a permissioned ledger records the scores, the admitted set, and the resulting average, so any hospital can recompute and audit the aggregate. Participation receipts follow the idea in I Trained That!, so a hospital can prove its update was included.
Only the hashes, scores and panel predictions go on the ledger. Panel predictions are a few thousand floats per site. The weights themselves, about 13 MB for 3.4M parameters, move peer-to-peer off-ledger. This sidesteps the ledger-bloat problem that BCFL addresses through sparsification.
The analysis side asks three questions. First, how much do honest sites disagree on the panel under raw-image versus vessel-graph models, as a function of how far apart the sites are in age mix and assessment center? Second, what is the separation between honest and poisoned updates under each screen? Third, does the vessel-only input cost real predictive signal?
The third question is the main risk. The feature-isolation paper's 0.754 was lower than an artifact-inflated baseline, and we do not know how much genuine non-vascular signal, such as textures of the retina's background tissue, raw models also use. We will measure that cost directly rather than assume it away.
Baselines
The naive reference points are local-only training at each simulated site, which shows how much collaboration buys a small site, and centralized pooled training of the same graph model on all data. Pooled training has no privacy protection and no attack surface, and it serves as the ceiling. The published figure it should reproduce is the 71.92% AUROC reported for Multimodal-CL-graph on UK Biobank stroke prediction in Predicting Stroke through Retinal Graphs and Multimodal Self-supervised Learning. That is the strongest reported number in the file on the benchmark we use. DeepRETStroke's 0.901 internal AUC comes from a different, non-public cohort and is not directly comparable.
The strongest decentralized competitor is plain swarm-style weight averaging of the same graph model with no screening, as in the histopathology swarm study. We add PPBFL-style cosine screening on the same weights, and Krum, which PPBFL itself compares against. The file reports no numeric OCTMNIST or PathMNIST accuracy for PPBFL, so we will rerun it ourselves rather than quote it. We also include DFPL, which exchanges class prototypes instead of weights and reported 82.96% versus BLADE-FL's 67.85% on non-IID CIFAR-10. It is the strongest communication-light decentralized baseline.
On the representation side, we include frozen RETFound (A foundation model for generalizable disease detection from retinal images) with only a small shared head trained by swarm averaging. We also run Fair-FedMOE from its released code, with its sign-consistency mask executed by every peer rather than a server.
Two ablations isolate the key ingredient. One applies our panel-agreement screen to a raw-image model; for that model the reference panel is the underlying public fundus images. The other applies cosine screening to the vessel-graph model. If vessel inputs alone explain the gain, the second ablation will match our method. If functional screening alone explains it, the first will.
For a fair comparison, all methods see identical site partitions, attack schedules and round counts. Each gets the same hyperparameter search budget, tuned on each site's local validation split only, and all run on single-GPU nodes. Our method beats the field if three things hold together. Its clean-data AUROC must be within one point of the pooled graph model. Its honest-site false-exclusion rate must be at least halved relative to cosine screening, with attack success no higher than Krum's. And the minority age group must keep its AUROC rather than having parity bought by lowering overall performance, the trade-off the UK Biobank bias paper observed for GroupDRO.
Why not just…?
Why not just use a central server with FedAvg and robust aggregation? The screening idea transfers to a server setting, and we will report that variant. The failure of weight-space screening on heterogeneous retinal sites happens with or without a server. The decentralized setting is the one where the problem cannot be patched by a trusted operator inspecting updates by hand, and where the test must be deterministic and auditable, which ours is. If server-side Krum on the graph model already matched our false-exclusion rate, the contribution would narrow to "use vessel inputs," and the cosine-on-graph ablation will show whether that is the case.
Why not just freeze RETFound and share a small head? This is the most serious threat. It is cheap, communication-light, and starts from strong features. Its weakness is that frozen raw-image features still encode camera and center identity, which is the concern raised by the feature-isolation result. So the heads trained at different sites may still diverge in ways a screen misreads as attacks. We run it head-to-head under the same attacks. If it matches our false-exclusion rate at higher AUROC, we will conclude that the gain comes from low-dimensional sharing rather than vessel space, and we will redirect the method to screening RETFound heads on a panel.
Why not just pool de-identified images? The histopathology swarm study already shows decentralized training can match or beat merged data (0.7736 versus 0.7567 AUROC for BRAF). Cross-border pooling of retina and stroke-registry records is legally blocked in the settings that motivate DeepRETStroke's call for international validation. Pooling therefore appears here only as the ceiling.
Why not just use clinical risk scores? DeepRETStroke reported 0.797 AUC for silent infarct detection versus 0.633 from clinical traits. Our models keep the clinical-tabular branch, so the comparison is additive rather than either-or.
Proposed Experiments
The main testbed is UK Biobank stroke prediction, partitioned into simulated hospitals by assessment center, the axis along which the bias study found a 10% gap. Some sites are deliberately skewed toward older participants to create honest minority cohorts. We vary the number of sites, and we run a harder cross-cohort split that treats the MCRS stroke cases and the SCES, SiMES and SiNDI controls from the feature-isolation study as separate peers. Attacks include label flipping, scaled-update model replacement, and a backdoor keyed to a vessel-graph pattern, at 10–30% malicious peers. OCTMNIST gives a direct replication point against PPBFL's own setting.
We measure stroke AUROC overall and per age group and center, the honest-site false-exclusion rate, attack success rate, bytes per round, ledger growth, and wall-clock time per round on single-GPU nodes.
Two checks decide the outcome. The first is whether the distribution of honest-site panel disagreement collapses when moving from raw-image to vessel-graph models, while poisoned updates stay separated. If it does not collapse, the premise fails. The second is whether the vessel-only model under our screen keeps within one AUROC point of the pooled 71.92% while cosine screening, at matched attack blocking, drops the elderly-skewed sites and loses minority-group AUROC. A final stress test gives an attacker full knowledge of the reference panel. This measures how much an adaptive adversary can game panel agreement, and whether a held-out, per-round random subsample of the panel restores the defense.
References
- Feature Isolation for Hypothesis Testing in Retinal Imaging: An Ischemic Stroke Prediction Case Study
- On Biases in a UK Biobank-Based Retinal Image Classification Model
- Privacy-Preserving Poisoning-Resistant Blockchain-Based Federated Learning for Data Sharing in the Internet of Medical Things
- Fair-FedMOE: Group-Fair One-Shot Federated Learning via Prototype-Guided Experts for Medical Imaging Analysis
- Predicting Stroke Through Retinal Graphs and Multimodal Self-supervised Learning
- Federated Learning in Ocular Imaging: Current Progress and Future Direction
- FairFedMed: Benchmarking Group Fairness in Federated Medical Imaging With FairLoRA
- Swarm learning for decentralized artificial intelligence in cancer histopathology
- Swarm learning with weak supervision enables automatic breast cancer detection in magnetic resonance imaging
- I Trained That! Client-Side Proof of Participation in Federated Learning
- A Fast Blockchain-Based Federated Learning Framework With Compressed Communications
- DFPL: Decentralized Federated Prototype Learning Across Heterogeneous Data Distributions
- A foundation model for generalizable disease detection from retinal images