README.md

metadata

title: OJOS AI
emoji: 👁️
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false

👁️ OJOS AI — Advanced Eye Disease Detection Web App

A state-of-the-art deep learning web application that detects and classifies multiple eye diseases with 96.69% validation accuracy using fine-tuned Convolutional Neural Networks (CNNs). Equipped with a premium glassmorphic dark-mode UI.

---

🎯 Key Features

• Multi-Disease Classification: Detects 5 distinct conditions:
  • 👁️ Cataract
  • ⚠️ Diabetic Retinopathy
  • 🔴 Glaucoma
  • 👓 Myopia
  • ✅ Healthy Eye
• High Performance: Achieves 96.69% accuracy on clinical validation datasets.
• Advanced Preprocessing: Uses LAB color space CLAHE (Contrast Limited Adaptive Histogram Equalization) to enhance contrast in retinal/eye images, matching clinical standards.
• Premium User Interface: Dark glassmorphic theme with sliding login/signup panel, drag-and-drop upload, scanning progress animation, and multi-probability confidence charts.
• Robust Authentication: Fully integrated SQLite database system for secure user registration and login session management.
• Keras 3 Runtime Compatibility: Includes a custom monkeypatch to bypass quantization config parameters when loading legacy .h5 model files.

---

🛠️ Tech Stack

• Backend: Python, Flask, SQLite3
• Deep Learning: TensorFlow / Keras (EfficientNetV2 backbone)
• Computer Vision: OpenCV, Pillow (PIL)
• Frontend: HTML5, Vanilla CSS3 (Glassmorphism, Neon Accents, CSS Keyframes)
• Deployment: Docker, Hugging Face Spaces (Git LFS)

---

🚀 How to Run Locally

1. Prerequisites

Ensure you have Python 3.9+ installed. Install the required libraries:

pip install tensorflow opencv-python numpy pillow flask

2. Clone the Repository

git clone https://github.com/Sharad9084/OJOS-AI.git
cd OJOS-AI

3. Run the Application

python app.py

Open http://localhost:5000 in your web browser.

---

📁 Project Structure

OJOS-AI/
├── best_eye_disease_model.h5   # Fine-tuned EfficientNetV2 multiclass model (Git LFS)
├── class_indices.json          # Mapping of class names to model indices
├── app.py                      # Core Flask server and prediction pipeline
├── database.py                 # SQLite database utility functions
├── requirements.txt            # Python dependencies
├── Dockerfile                  # Containerization config for Hugging Face Spaces
├── Procfile                    # Web process config
├── static/
│   ├── css/
│   │   └── style.css           # Premium glassmorphic styling sheet
│   ├── logo.png                # Custom circular eye logo
│   └── uploads/                # Directory for temporary preview images
└── templates/
    ├── home.html               # Main landing page
    ├── login.html              # Login & signup portal
    ├── prediction.html         # Real-time scan and multi-probability results
    ├── about.html              # Team / info page
    └── diseases_info.html      # Visual medical encyclopedia

---

💡 How It Works

1. Image Preprocessing & Contrast Enhancement

Retinal images can have inconsistent illumination. To normalize this:

• The input image is converted to the LAB color space.
• CLAHE is applied on the L-channel (lightness) to enhance local contrast without over-amplifying noise.
• The image is merged back, converted to RGB, and resized to 300x300.
• EfficientNetV2 normalization (preprocess_input) is applied.

2. Inference & Classification

• The preprocessed image is passed into best_eye_disease_model.h5.
• The model outputs logits/probabilities for all 5 target categories.
• The front-end renders a interactive bar chart displaying the confidence percentage for every condition.

---

⚕️ Medical Compliance & Disclaimer

• Screening Tool Only: This application is designed as an assistive screening tool. It does not replace professional medical diagnosis, advice, or treatment by a qualified ophthalmologist.
• Data Privacy: Images uploaded are saved temporarily on the server for visualization and are not shared.

---

📞 Contact & Support

Developed by Sharad Sharma (sharmasharad9794@gmail.com).

⭐ If you found this project helpful, please star the repository!