Detection and localization of COVID 19 and Pneumonia on chest radiographs

Overview

This repository contains the implementation of deep learning techniques for detecting and localizing COVID-19 lung lesions on chest radiographs. We utilized a multi-class classification approach with three deep learning architectures: DenseNet169, VGG16, and a custom sequential model.

Key Points

• Objective: Early diagnosis of COVID-19 using deep learning algorithms on chest X-rays.
• Models Used: DenseNet169, VGG16, and a non-pretrained sequential architecture.
• Techniques: Transfer learning and ensemble learning were employed to classify radiographs into three categories: "COVID", "Pneumonia", and "Normal".
• Database: The dataset consists of 3225 chest radiographs selected by a radiologist from the COVIDx-CXR version 8 database.
• Results: Achieved state-of-the-art results with accuracy values above 83% for individual models and up to 96% for ensemble models.
• Visualization: Class activation mapping (CAM) techniques were used to localize and visualize COVID lesions on chest radiographs.

Getting Started

To get started with the code, please follow the instructions below.

Prerequisites

• Python 3.x
• TensorFlow
• Keras
• OpenCV
• NumPy
• Matplotlib

Dataset

Available on : https://doi.org/10.6084/m9.figshare.25917340.v1

Models

Available on : https://drive.google.com/drive/folders/1RI_BU9Ew6b_HtK1wljF4Vt_Ug6dW2Jem?usp=drive_link

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

For any questions, please contact [ahmed.balaazi@etudiant-fmt.utm.tn]