AEYE — AI-Guided Retinal Screening (Mobile App)

AEYE is an Expo / React Native mobile application for diabetic retinopathy (DR) screening. Doctors capture or upload fundus photographs from the device, an EfficientNet-B4 model hosted on Hugging Face Spaces grades severity across five levels (No DR → Proliferative), and the results screen displays Grad-CAM heatmaps with a 5-colormap toggle, progression alerts vs. the patient's prior visit, and downloadable PDF reports.

The companion web platform lives in a separate repo: DRC_web. The ML inference service is hosted on Hugging Face Spaces — no local model setup required.

---

Features

• DR grading — 5-class severity (No DR / Mild / Moderate / Severe / Proliferative) via calibrated EfficientNet-B4
• Grad-CAM heatmaps — 5 colormaps (Turbo, Inferno, Jet, Viridis, Magma); Turbo + Inferno pre-rendered at upload, Magma / Viridis / Jet fetched on-demand via /recolor and cached client-side
• Progression alerts — rule-based red/green banner when DR worsens or improves vs. the patient's previous scan
• Multi-role auth — Patient, Doctor, Admin via Supabase magic-link email
• PDF reports — generated with expo-print, shared via expo-sharing
• Follow-up scheduling — doctors set follow-up dates; patients see reminders
• Dual-model support — AEYE v1 (calibrated) + optional partner model; model picker surfaces when both URLs are configured
• Camera + gallery uploads — expo-image-picker for capture or library selection

---

Tech Stack

Layer	Technology
Framework	Expo SDK 52, React Native, TypeScript
Styling	NativeWind (Tailwind CSS for React Native)
Navigation	React Navigation v7 (native stack + bottom tabs)
Auth + DB	Supabase JS client (@supabase/supabase-js)
Storage	Supabase Storage (fundus image bucket)
ML backend	FastAPI on Hugging Face Spaces (EfficientNet-B4 + Grad-CAM)
PDF	expo-print + expo-sharing
Image picker	expo-image-picker, expo-file-system

---

Project Structure

mobile/
├── App.tsx                  # Root component
├── app.json                 # Expo config (name, slug, permissions, icons)
├── eas.json                 # EAS Build profiles
├── index.ts                 # Entry point
├── assets/                  # Icons, splash, intro video
└── src/
    ├── screens/             # ResultsScreen, AnalysisScreen, HistoryScreen, …
    ├── components/          # AppHeader, WelcomeIntro, ui/* primitives
    ├── navigation/          # AppNavigator (role-based stack + tabs)
    ├── contexts/            # AuthContext
    ├── lib/
    │   ├── api.ts           # Supabase data layer (scans, profiles, follow-ups)
    │   ├── mlApi.ts         # ML backend client (/predict, /recolor)
    │   ├── progression.ts   # Rule-based DR progression utility
    │   └── supabase.ts      # Supabase client init
    └── hooks/

---

Prerequisites

• Node.js 20+
• Expo Go app on a physical device, or Android Studio / Xcode for emulators
• A Supabase project (same project as the web app, if both are deployed)

---

Setup

1. Clone and install

git clone https://github.com/faizanshoukat5/DRC-APP.git
cd DRC-APP/mobile
npm install

2. Environment variables

Create mobile/.env:

# Supabase
EXPO_PUBLIC_SUPABASE_URL=https://your-project.supabase.co
EXPO_PUBLIC_SUPABASE_ANON_KEY=your-anon-key

# AEYE ML backend (Hugging Face Space)
EXPO_PUBLIC_DR_API_URL=https://faizan055-dr-classifier.hf.space
EXPO_PUBLIC_DR_API_KEY=hf_xxxxxxxxxxxxxxxxxxxx

# Optional: partner model — leave unset to hide the model picker
EXPO_PUBLIC_DR_API_URL_PARTNER=https://hissanzahir-dr-detection-api.hf.space

All Expo-public envs must be prefixed with EXPO_PUBLIC_ to be inlined into the JS bundle.

3. Run on a device or simulator

npx expo start

Then either:

• Scan the QR code with Expo Go (iOS / Android)
• Press a to launch on Android emulator
• Press i to launch on iOS simulator
• Press w to open in web (limited compatibility)

For a fully native build on a connected Android device (no Expo Go needed):

npx expo run:android

---

ML Backend

The mobile app talks to a FastAPI service hosted on Hugging Face Spaces. No local setup is required — the Space is publicly accessible.

Endpoints used by mobile/src/lib/mlApi.ts:

Endpoint	Purpose
POST /predict	Upload multipart file; returns class_id, class_name, confidence, probabilities, heatmaps_b64 (Turbo + Inferno pre-rendered)
POST /recolor	Multipart file + colormap field; returns heatmap_b64 for Magma / Viridis / Jet on demand

Default Space: https://faizan055-dr-classifier.hf.space Source + weights: huggingface.co/spaces/faizan055/dr-classifier

To point the app at a different deployment, override EXPO_PUBLIC_DR_API_URL in mobile/.env.

---

Supabase Configuration

1. Create a project at supabase.com.
2. Copy the Project URL and anon/public key into mobile/.env.
3. Apply the schema (run from the companion DRC_web repo: npm run db:push).
4. In Authentication → URL Configuration, add aeye:// as a redirect URL for deep-link magic-link sign-in.

The mobile app shares the same Supabase project as the web app — scans uploaded on either platform appear on both.

---

User Roles

Role	Capabilities
patient	View own scans + follow-ups, download PDFs
doctor	Capture / upload fundus images, write clinical notes, schedule follow-ups, view assigned patients
admin	View all scans + profiles, approve doctor registrations

New accounts default to patient. Doctors must be approved by an admin (web Admin Dashboard or directly in the Supabase profiles table).

---

Building for Production

Android APK / AAB (EAS Build)

cd mobile
npx eas login
npx eas build --platform android --profile preview     # APK for sideload
npx eas build --platform android --profile production  # AAB for Play Store

iOS (EAS Build — requires Apple Developer account)

npx eas build --platform ios --profile production

Build profiles are configured in mobile/eas.json.

---

Companion Web Platform

The full clinician dashboard, admin tools, and PDF history live on the web. Repo: github.com/faizanshoukat5/DRC_web.

Both apps share the same Supabase project and ML backend — a scan uploaded on mobile shows up instantly on the web dashboard, and vice versa.

---

License

MIT