Basira — retinal screening decision support
Three AI models second-read an eye scan and report how far they agree; a doctor always makes the final call.
My partBuilt solo · a prototype, not a medical device
In short
A prototype web platform for eye clinics in Lebanon in which three independent retinal-image models each analyze a fundus or OCT image, a consensus engine reports agreement, urgency, and a referral suggestion, and the reviewing doctor confirms or overrides before a report is issued.
- What it is
- A decision-support prototype that second-reads retinal images for clinics and always leaves the decision to a doctor.
- Why it matters
- Retinal screening in a clinic depends on scarce specialist time. A second reader that prioritizes and flags, running on an ordinary CPU on site, could help, but only if it never pretends to diagnose.
- My part
- I built the platform, the model integration, and the deployment on my own.
Step by step
- 1
Gate
An image-quality gate runs before any model; ungradable images are reported as such.
- 2
Read three times
Each council member serves predictions only when its encoder and a trained head both load; an untrained head never produces a diagnosis, and the API reports stubs explicitly.
- 3
Merge
The consensus engine reports unanimous, majority, or split agreement, an urgency level, and a referral recommendation, with explainability heatmaps.
- 4
Review
The doctor’s confirmation or override is the record; the PDF report is generated from that decision.
Results
- ~2 s per eye
- Council latency
- all three models on a plain CPU, no GPU
For engineersHow it is built, its limits, and its keywords.
How it works
A FastAPI backend, a separate ML service, and a Next.js frontend share one data directory. Three independent retinal-image models, each with a classification head trained on openly licensed public datasets (DDR, IDRiD, RFMiD, PAPILA), read the image; a consensus engine merges the three; the doctor confirms or overrides; a trilingual PDF report is generated.
Limits
- Basira is in development. Benchmarks on public datasets guide the engineering but say nothing about clinical performance; the clinical validation study has not been run, and Basira is not a medical device.
- The live deployment uses a demo clinic and demo data only.
Keywords
- retinal screening
- decision support
- FastAPI
- Python
- ophthalmology