🏥 OrOeye: Deep Learning Oral Lesion Analysis

This isn't just another classification script.

OrOeye is a production-grade, full-stack medical AI platform designed to detect potential oral anomalies with explainable precision. We moved beyond simple "black box" predictions to build a system that medical professionals can actually verify.

It combines Computer Vision (CNNs) for detection, Heuristic Forensics for data validation, and Generative AI for system analytics into a single, cohesive architecture.

---

⚡ The "Heavy Lifting" (Technical Deep Dive)

We didn't just pip install tensorflow and call it a day. Here is what's actually happening under the hood:

1. Explainable AI (XAI) with Grad-CAM

Most AI models just say "Cancer: 90%". That's useless for a doctor.

• What we built: We implemented Gradient-weighted Class Activation Mapping (Grad-CAM).
• How it works: We hook into the final convolutional layer of our InceptionResNetV2 model to extract the gradients. We then overlay a heatmap on the original image, showing you exactly which pixels triggered the prediction.
• The Result: You don't just get a diagnosis; you get visual proof of the lesion the AI detected.

2. Heuristic Image Forensics (The "Gatekeeper")

Garbage In, Garbage Out. We built a custom validation engine using OpenCV that runs before the AI even touches the image. It uses physics-based heuristics to reject invalid data:

• Spectral Analysis: Rejects images with dominant blue/sky tones (outdoors/non-medical).
• Flesh-Tone Ratios: Calculates the saturation of specific HSV ranges to ensure the image contains mucosal tissue.
• Face-Rejection Logic: Uses Haar Cascades to detect if a user uploaded a "selfie" instead of a zoomed-in oral cavity shot. If it sees a face, it blocks the upload.

3. Hybrid AI Architecture

We aren't relying on one model.

• The Eye: A fine-tuned InceptionResNetV2 (CNN) handles the visual diagnostics.
• The Brain: We integrated LLMs (GPT-4) into the Admin Dashboard (/admin). It analyzes system logs, user patterns, and scan statistics to provide natural language insights to administrators.

---

🛠️ System Architecture

• Backend: Python (Flask) serving as the REST API and inference engine.
• Database: MongoDB Atlas (NoSQL) for flexible user data and scan history storage.
• Auth: Secure session management with bcrypt encryption.
• Reporting: Automated PDF generation engine (ReportLab) that compiles the original image, the AI heatmap, and clinical notes into a downloadable medical report.

---

🚀 How to Run It

1. Clone & Prep

git clone https://github.com/thisistanishq/oroeye.git
cd oroeye

2. Virtual Env (Recommended)

python -m venv venv
source venv/bin/activate  # Mac/Linux
# venv\Scripts\activate   # Windows

3. Install Dependencies

pip install -r backend/requirements.txt

4. Run the Server

# We handle the frontend serving directly from Flask for speed
python wsgi.py
# OR
cd backend && python app.py

Visit http://127.0.0.1:5001.

---

🔐 Admin Dashboard

The system includes a fully equipped Admin Dashboard for clinic managers.

• Real-Time Analytics: Live tracking of daily/weekly scans, cancer detection rates, and user growth.
• System Logs: Real-time event logging engine that aggregates user actions, scans, and system warnings.
• Message Center: Full management interface for contact form submissions.
• User Management: Search, filter, and manage user accounts and access roles.

---

⚠️ Simulation Mode

Note: The trained model files (.h5 / .keras) are massive (~200MB+). If you clone this repo without pulling the LFS files, the system detects the missing model and automatically switches to Simulation Mode.

• It will function normally but return simulated "Non-Cancerous" results to prevent crashing.
• To get real predictions, ensure you have the model files in backend/model/.

---

Built by Tanishq. Code that actually works.