Chest X-Ray Pneumonia Detector xAi + Web app

Project Overview

Developed a pneumonia detection system using MobileNet-V3 with transfer learning, achieving 91% accuracy in detecting pneumonia from chest X-rays. Optimized the model for edge devices through knowledge distillation to reduce model size and improve performance in resource-constrained environments. Deployed the model on AWS SageMaker using Flask, ensuring seamless integration and scalability for production environments. Implemented Grad-CAM for model interpretability, visualizing CNN focus areas to enhance prediction accuracy and provide insight into model decision-making.

Dataset Overview

• Normal images count in training set: 1349
• Pneumonia images count in training set: 3883
• Total Count of images: 5232

Project Structure

Chest_Xray_Pneumonia_Detector/
│── data/
│   ├── pneumonia/
│   ├── normal/
│── train/
│   ├── model_training_code.ipynb
│   ├── saving_model/
│       ├── pneumonia_model.keras
│── results/
│       ├── sample1.png
│       ├── sample3.png
│       ├──.....
│── deployment/
    ├── app.py
    ├── requirements.txt
    ├── index.html

Installation & Setup

pip install -r requirements.txt

Model Training (Google Colab)

1. Install dependencies

   !pip install -r requirements.txt

2. Prepare dataset

   Chest_Xray_Pneumonia_Detector/
       pneumonia /
           img1.jpg
           img2.jpg
       normal/
           img1.jpg
           img2.jpg
   

3. Train the model using model_training_code.ipynb

Deployment Guide

1. Local Deployment (Flask)

1. Navigate to the deployment/ folder and run:

   python app.py

2. The web app will be available at http://127.0.0.1:5000

2. AWS EC2 Deployment

1. Launch an EC2 instance with Ubuntu.
2. Install dependencies:

   sudo apt update && sudo apt install python3-pip
   pip install -r requirements.txt

3. Run the Flask app:

   python app.py

4. Configure security groups to allow inbound traffic on port 5000.

API Endpoints

• POST /predict - Upload a chest X-ray and get a pneumonia prediction.

Results & Interpretability

• Achieved 91% accuracy
• Used Grad-CAM to visualize model focus areas for interpretability(xAI).

---

Contact

👤 Author: Sai Krishna Chowdary Chundru
📩 Email: cchsaikrishnachowdary@gmail.com
🔗 LinkedIn: linkedin.com/in/sai-krishna-chowdary-chundru
💻 GitHub: github.com/sAI-2025

---

License

This project is licensed under the MIT License – feel free to use and modify!