Skip to content
RetinaGuard AIRetinal screening research prototype
UnivaBio · AI for Human Health

Earlier diabetic retinopathy screening, with AI that knows when not to answer.

RetinaGuard AI is a research prototype console for grading retinal fundus photographs on a five-level scale — built around image quality, uncertainty, and professional follow-up.

Analyze a fundus image Model & evaluation plan

Research prototype. Not for diagnosis or treatment decisions.

Illustrative demo — not a real patient or model prediction
Schematic · not a real image
Illustrative demo data
Demo case · DEMO-000

Example model output (illustrative)

Grade 2 · Moderate

  1. 0 · No apparent signs
  2. 1 · Mild
  3. 2 · Moderate
  4. 3 · Severe
  5. 4 · Proliferative

These values are hand-written to demonstrate the interface. No image was analyzed and no AI model produced them.

The problem

Damage can progress quietly.

Diabetic retinopathy is a complication of diabetes affecting the blood vessels of the retina. In its early stages it may cause no obvious symptoms, so people can feel their vision is fine while changes develop.

Regular retinal examination and review by eye-care professionals remain the foundation of detection. This project explores how AI research might help prioritize images for that professional review — never replace it.

Workflow

From photograph to follow-up

  1. 01

    Upload retinal image

    Fundus photograph, JPG/PNG

  2. 02

    Image-quality checks

    Type, size, gradability

  3. 03

    AI grading service

    Configurable external API

  4. 04

    Uncertainty-aware result

    Withheld when unreliable

  5. 05

    Clinician follow-up

    Professional review always

Sample case

How a result is presented

A demonstration of the results layout using the five-grade scale. Every value in this card is illustrative.

Illustrative demo — not a real patient or model prediction
Schematic · not a real image
Illustrative demo data
Demo case · DEMO-000

Example model output (illustrative)

Grade 2 · Moderate

  1. 0 · No apparent signs
  2. 1 · Mild
  3. 2 · Moderate
  4. 3 · Severe
  5. 4 · Proliferative

Example probability layout (illustrative values)

0 · No apparent signs
0.08
1 · Mild
0.17
2 · Moderate
0.52
3 · Severe
0.16
4 · Proliferative
0.07

These values are hand-written to demonstrate the interface. No image was analyzed and no AI model produced them.

Project pillars

Designed for responsible research

Transparent AI

Model plan, metrics and limitations are documented openly in a model card skeleton.

Quality awareness

Ungradable images produce no grade — only a recommendation for professional review.

Uncertainty

High model uncertainty suppresses reassuring language and surfaces a caution.

Privacy

No account, no storage. Images leave your browser only when you run a configured API.

What this prototype can and cannot do

This prototype can
  • Validate and preview a fundus photograph locally in your browser
  • Send it to a configured research inference API you choose to run
  • Display returned grade, probabilities, uncertainty and quality — only when returned
  • Withhold a grade and recommend professional review when output is unreliable
  • Export a printable research summary
It cannot
  • Diagnose, rule out, or treat any eye disease
  • Replace an examination by an eye-care professional
  • Claim clinical validation or regulatory clearance — it has neither
  • Produce a result when no trained model endpoint is connected
  • Interpret selfies, skin photos or ordinary images