Retina OCTOpen images. Deeper insights.

About

A searchable atlas of retinal optical coherence tomography B-scans, paired with expert biomarker annotations and clinical measurements.

78,283OCT B-scans
9,408Expert-annotated
16Biomarkers
87Patients

What this is

Most published OCT datasets are distributed as archives you download whole, unpack, and then write your own code to navigate. This atlas indexes one such dataset so it can be searched directly — filter by biomarker, eye, trial, visit, retinal thickness or visual acuity, and inspect any individual scan alongside its clinical context.

Every image is a single B-scan — one cross-sectional slice through the retina. Scans come in volumes of 49, so adjacent slices from the same eye and visit sit next to each other in the index.

Where biomarkers come from

Biomarker values are never anonymous. Each is attributed to whatever produced it, and the attribution is stored alongside the value:

KindMeaningConfidence
expert Labelled by clinicians and shipped with the source dataset—
model Predicted by a named, versioned machine-learning modelStored per value
algorithm Computed deterministically, e.g. layer thickness by image analysis—

This distinction matters. An expert annotation and a model prediction can disagree, and a reader must be able to tell which is which. Model output never overwrites human labels — both are kept, with their sources.

The 16 biomarkers

Annotated images carry binary labels for each of the following:

Prevalence is very uneven — intraretinal hyperreflective foci appear in 6,341 images, while vitreomacular traction appears in 10. Rare findings are shown in the filters only when the current selection contains them.

Clinical measurements

Beyond biomarkers, every image carries two measurements recorded at the visit:

How it is built

A static page served from Cloudflare Pages, with search handled by Pages Functions querying a Cloudflare D1 database. Images are stored in Backblaze B2 and served through a caching proxy. Source data is TIFF, which browsers cannot render, so each scan is converted to WebP at two sizes.

The full build — schema, extraction scripts, and a record of the decisions and mistakes along the way — is public on GitHub.

Research use only. This atlas is not a medical device, has no regulatory clearance, and must not inform diagnosis or treatment. Model-predicted biomarkers are unvalidated for clinical use.