● Edge AI · Open Source · Digital Public Good nominee

Diabetic retinopathy screening,
100% on the device

DIRD+ is an open-source desktop application that detects signs of diabetic retinopathy with local AI. No cloud, no per-screening fee — patient data never leaves the device.

⬇ Download View source Documentation
License AGPLv3 DOI Platforms
⚠️
Research tool. DIRD+ is not an approved medical device and must not be used as the sole diagnostic criterion. Every finding must be reviewed by a qualified ophthalmologist.
Why DIRD+

Screening that respects privacy, cost and connectivity

Diabetic retinopathy is the leading cause of blindness in working-age adults — and over 80% of it is preventable with timely screening. DIRD+ brings AI screening to clinics that lack specialists, budget, or stable internet.

🔒

Full privacy

All AI inference runs on the device. Zero transmission to external servers. The database is encrypted at rest (AES-256).

📡

Works offline

After a one-time model download, everything runs locally. Built for areas with no reliable connectivity.

💸

Zero cost

No licenses, no per-screening fee, no proprietary hardware. Works with any existing fundus camera.

🧩

Model-agnostic

Plug in your own ONNX model via a simple model card — calibrated to your own population, no code changes.

📋

Guideline-agnostic

Clinical guidelines are JSON files. Ships with ICDR 2024 and MINSAL Chile 2017; add new ones without coding.

📖

Open source

Fully auditable under GNU AGPLv3. Algorithms, thresholds and criteria are transparent and verifiable.

How it works

From fundus image to clinical report — on-device

A deterministic, auditable pipeline. The clinical guideline makes the decision; the optional local LLM only writes the prose.

Capture & upload

Load fundus images (OD/OI) from any camera into the desktop app.

AI detection

Local ONNX models find lesions (hemorrhages, exudates, edema, microaneurysms…).

Review

Inspect, correct and annotate on a multi-layer canvas with measurement tools.

Classify

The pluggable guideline grades severity with treatment and follow-up — deterministic, no black box.

Report

A configurable PDF; an optional in-process LLM polishes the prose, never the decision.

Export

Portable, encrypted .dird containers move data between installations.

Open by design

A platform, not a black box

🧩 Bring your own model

Train on your own cohort and deploy inside DIRD+ — no vendor lock-in.

  • Standard ONNX, input [1,3,640,640]
  • Declare classes & metadata in a JSON model card
  • Schema + tensor-shape + inference validation on load
  • CLI validator and a documented contract
Model interface →

📋 Pluggable clinical guidelines

Adapt classification to any national protocol — guidelines are data, not code.

  • 5+ severity levels with treatment & urgency
  • Per-quadrant spatial analysis (ICDR 4-2-1 rule)
  • Add a country's guideline as a JSON file
  • Human corrections preserved and traceable
Clinical guidelines →
Tauri v2RustReact + TypeScript ONNX Runtimellama.cpp (local LLM)SQLite + SQLCipher Argon2idAES-256
Digital Public Good

Built for equitable, sovereign health screening

DIRD+ aligns with the Digital Public Goods Standard: open licensing, clear ownership, platform independence, privacy by design, and open standards (ONNX, PDF, SQLite, JSON, AES-256, Argon2id).

Explore the project AI models (AGPLv3)