Pneumonia Detection with CNN & Grad-CAM

This project provides a complete deep learning pipeline for detecting pneumonia from chest X-ray images. It demonstrates how to transition an AI model from experimental notebooks into a deployable web application, featuring robust explainability using Grad-CAM.

Features

• Dual Model Architecture: Choose between training a baseline CNN from scratch or utilizing Transfer Learning with a pre-trained ResNet18 model.
• Resumable Training: The training loop automatically checkpoints the model state after every epoch (resnet_checkpoint.pth). If training is interrupted, simply run the cell again to seamlessly resume from exactly where it left off.
• Explainable AI (XAI): Integrated pytorch-grad-cam to generate heatmaps highlighting the exact regions of the lung the model focused on when making its prediction.
• Live Web Demo: A standalone app.py built with Gradio, ready to be deployed directly to Hugging Face Spaces for live inference.

Project Structure

├── pneumonia_detection.ipynb  # Core training workspace (Data Prep, Models, Checkpointing, Grad-CAM)
├── app.py                     # Standalone Gradio web interface for deployment
├── requirements.txt           # Python dependencies
└── .gitignore                 # Excludes dataset and large model checkpoints

(Note: The X-ray dataset and .pth checkpoints are ignored by git due to size constraints).

Quick Start

1. Installation

Clone the repository and install the dependencies:

git clone https://github.com/Himanshu-Vishwakarma-GH/Pneumonia-CNN.git
cd Pneumonia-CNN
pip install -r requirements.txt

2. Training the Model

1. Place the dataset inside a folder named xray_dataset_covid19/ (containing train and test subfolders).
2. Open pneumonia_detection.ipynb in Jupyter or VS Code.
3. Run the cells to initiate the resumable training loop. It will automatically save an epoch checkpoint (resnet_checkpoint.pth or cnn_checkpoint.pth).

3. Running the Live Demo

Once you have generated a checkpoint file from the notebook, you can launch the Gradio web interface:

python app.py

This will open a local web server (typically http://127.0.0.1:7860) where you can drag and drop X-rays and instantly view the AI's prediction alongside the Grad-CAM heatmap.

Technologies Used

• PyTorch & Torchvision
• Grad-CAM (Gradient-weighted Class Activation Mapping)
• Gradio (Web Deployment)
• OpenCV & PIL (Image processing)