An end-to-end deep learning solution for classifying chest X-ray images into three categories: COVID-19, Normal, and Viral Pneumonia. This project features a fine-tuned ResNet50 model, optimized with TensorFlow Lite, and deployed as an interactive web application using Streamlit.
The timely and accurate diagnosis of respiratory diseases like COVID-19 and Viral Pneumonia is crucial for patient treatment and controlling public health crises. Manual analysis of chest X-rays can be time-consuming and requires expert radiologists. This project aims to address this challenge by developing an automated tool that uses deep learning to provide a fast, accurate, and accessible preliminary diagnosis from chest X-ray images, assisting healthcare professionals in the screening process.
- Multi-Class Classification: Accurately distinguishes between COVID-19, Normal, and Viral Pneumonia.
- High-Performance Model: Built using transfer learning with a fine-tuned ResNet50 architecture, achieving over 96% accuracy on the test set.
- Model Optimization: The final model was quantized using TensorFlow Lite, reducing its size by ~75% (from 240 MB to ~60 MB) for efficient deployment.
- Interactive Web Application: A user-friendly and responsive web interface built with Streamlit for easy image uploads and instant prediction results.
- Deployment Ready: The application is structured and prepared for one-click deployment on Streamlit Community Cloud.
The project was executed in several key phases:
- Data Preprocessing: The image dataset was loaded and preprocessed on the fly using Keras's
ImageDataGenerator. This included resizing all images to a consistent 224x224, applying aggressive data augmentation (rotation, zoom, flips) to prevent overfitting, and normalizing pixel values. - Model Building: A ResNet50 model, pre-trained on the ImageNet dataset, was used as the base. The top classification layer was removed and replaced with a new custom head suitable for our three classes.
- Training & Fine-Tuning: The model was trained in two phases:
- First, only the custom head was trained with the base model's layers frozen.
- Then, the final few blocks of the base model were unfrozen and the entire model was fine-tuned with a very low learning rate to adapt its features to the X-ray data. Class weights were used to handle the imbalanced dataset.
- Model Quantization: After training, the final
.kerasmodel was converted to a.tflite(TensorFlow Lite) model. Post-training quantization was applied to reduce the file size and improve inference speed. - Application Development: An interactive web application was developed using Streamlit to provide an easy-to-use interface for the model.
| Category | Technologies & Tools |
|---|---|
| Machine Learning | TensorFlow, Keras, Scikit-learn, NumPy, Pillow, OpenCV |
| Web Application | Streamlit |
| Deployment & Version Control | Git, Git LFS, GitHub, Anaconda |
The model's performance was evaluated on a held-out test set, demonstrating high accuracy and balanced precision/recall across all classes.
| Class | Precision | Recall | F1-Score |
|---|---|---|---|
| COVID | 1.00 | 0.97 | 0.98 |
| Normal | 0.98 | 0.95 | 0.97 |
| Viral Pneumonia | 0.99 | 0.98 | 0.98 |
| Overall Accuracy | 96.7% |
Follow these steps to set up and run the project on your local machine.
# Install Git LFS (only needs to be done once per machine)
git lfs install
# Clone the repository
git clone [https://github.com/itz-Mayank/Chest-X-Ray-Diagnostics.git](https://github.com/itz-Mayank/Chest-X-Ray-Diagnostics.git)
cd Chest-X-Ray-Diagnostics.git
conda create -n cxr_diagnostics python=3.10 -y
conda activate cxr_diagnostics
pip install -r requirements.txt
streamlit run app.py
The application will open in your web browser at http://localhost:8501.
Created and developed by Mayank Meghwal.