Federated validation for
infectious keratitis AI.
K-ERA is a research platform for evaluating infectious keratitis AI across hospitals while raw patient data remain on local systems.
Sites apply on the web, continue in the desktop app on a hospital PC, and contribute local cases under institutional approval.
Dept. of Ophthalmology · Jeju National University Hospital
Founding benchmark results
from the initial single-site cohort.
The public benchmark is intentionally strict: white-light images only, patient-disjoint evaluation, visit-level prediction, and leakage-aware controls.
The app can still register multimodal visits, including fluorescein, but the current default AI path remains anchored to that white-light benchmark. A single-center feasibility result tells you what is possible. Multi-site validation is what makes that evidence credible across institutions.
Early treatment decisions are difficult,
and external validation remains limited.
Infectious keratitis remains one of the leading causes of preventable corneal blindness worldwide. Bacterial and fungal keratitis can present with overlapping features, making early differential diagnosis difficult even for experienced clinicians.
What remains limited is not only model performance, but the ability to build broader validation cohorts under real privacy and governance constraints.
Approval on the web.
Case work on the desktop.
Review stays in the loop.
Case authoring, image upload, and AI assessment all run from the K-ERA desktop app installed on a hospital PC. The web portal handles account approval only — patient images never reach a web server. This is a deliberate security boundary, not a technical limitation.
Raw data stays on-site.
Validation reaches across hospitals.
Multi-site validation is difficult when institutions cannot transfer raw images and patient identifiers. A federated workflow allows each hospital to keep patient data on-site while still contributing to a shared research network.
In K-ERA, participating sites contribute reviewed weight updates, de-identified metadata, and limited review assets rather than raw clinical data. This keeps the privacy boundary intact while making broader validation possible.
Hospital B → local training → Δ weights → Central review
Hospital C → local training → Δ weights
Reviewed aggregation → redistributed to approved nodes
All patient-facing case work and local training remain on the K-ERA desktop app installed on a hospital PC. The browser is used only for approval and access management.
After institutional approval, sites can install the desktop app, register cases locally, and join the review and aggregation workflow.
- →Multi-site publications follow ICMJE authorship criteria and written contribution agreements established before submission.
- →Access to the K-ERA AI inference model and confidence outputs for enrolled cases.
- →Similar-case retrieval across the shared research corpus via the DINO embedding index.
- →Federated weight updates distributed to approved nodes after central review.
Multi-site validation is the next step.
dr.jinho.jeong@gmail.com