-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Nicolás Baier Quezada edited this page Jun 5, 2026
·
5 revisions
DIRD+ is an open-source desktop application for diabetic retinopathy (DR) screening. It runs computer-vision models entirely on the user's device — no cloud, no backend, no per-screening cost. Packaged with Tauri v2 (Rust + WebView).
This wiki is the technical / developer reference for the current version of the app. For the project overview, epidemiology, and market context, see the README.
⚠️ DIRD+ is a research-and-development system, not an approved medical device. Every finding must be reviewed by a qualified ophthalmologist.
- Patients & sessions — register patients (clinical data) and create sessions (clinical visits).
- Upload fundus images per eye (OD/OI).
- On-device AI — ONNX detection + segmentation models run locally (WebAssembly) and find retinal lesions.
- Review — an interactive multi-layer canvas lets the clinician inspect, correct, and annotate the AI output.
- Classify — a pluggable clinical-guideline engine (ICDR 2024 / MINSAL 2017) turns detections into a severity grade with treatment and follow-up. This step is deterministic and local — no generative AI.
- Report — a configurable PDF. The clinical conclusion is written by the guideline; an optional local LLM (llama.cpp, in-process) only polishes the prose.
-
Export/Import — portable, encrypted
.dirdcontainers.
Everything is stored in a SQLite database encrypted at rest (SQLCipher, AES-256, Argon2id key derivation).
- 100% local / offline after the first model download. Patient data never leaves the device.
- Model-agnostic — plug in your own ONNX model via a model card. See 09. Model Interface.
- Guideline-agnostic — add clinical guidelines as JSON. See 10. Clinical Guidelines.
- Encrypted at rest with a dual-password model. See 11. Security & Privacy.
- No telemetry, no remote inference, no accounts.