Pneumonia Detection

Overview

This project classifies images of chest X-rays into two categories: Normal and Pneumonia. It uses a Convolutional Neural Network (CNN) for image classification using Keras.

Dataset

The dataset used for this project is sourced from Kaggle. It contains chest X-ray images categorized into two folders: NORMAL and PNEUMONIA. The NORMAL folder contains X-ray images of healthy patients, while the PNEUMONIA folder contains X-ray images of patients diagnosed with pneumonia.

The above images shows a batch of 5 images that are Normal

The above images shows a batch of 5 images that are Pneumonia

Convolutional Models

Custom CNN Model

Architecture

The custom CNN model consists of the following layers:

1. Conv2D: 32 filters, kernel size (3, 3), activation 'relu', input shape (224, 224, 3)
2. MaxPooling2D: pool size (2, 2)
3. Conv2D: 64 filters, kernel size (3, 3), activation 'relu'
4. MaxPooling2D: pool size (2, 2)
5. Conv2D: 128 filters, kernel size (3, 3), activation 'relu'
6. MaxPooling2D: pool size (2, 2)
7. Flatten
8. Dense: 512 units, activation 'relu'
9. Dropout: rate 0.5
10. Dense: 1 unit, activation 'sigmoid'

Training Process

1. Data Preprocessing: Images are resized to 224x224 pixels and normalized.
2. Compilation: The model is compiled with binary cross-entropy loss and the Adam optimizer.
3. Training: The model is trained on the training dataset with a validation split, using data augmentation techniques to improve generalization.

Pretrained VGG16 Model

Architecture

The VGG16 model is a deep convolutional network pre-trained on the ImageNet dataset. For this project, the following modifications are made:

1. VGG16 Base: Pre-trained VGG16 model without the top layers.
2. Flatten
3. Dense: 256 units, activation 'relu'
4. Dropout: rate 0.5
5. Dense: 1 unit, activation 'sigmoid'

Training Process

1. Data Preprocessing: Images are resized to 224x224 pixels and normalized.
2. Feature Extraction: The pre-trained VGG16 model is used to extract features from the images.
3. Fine-Tuning: The custom top layers are trained on the extracted features, while the lower layers of VGG16 are optionally fine-tuned.
4. Compilation: The model is compiled with binary cross-entropy loss and the Adam optimizer.
5. Training: The model is trained on the training dataset with a validation split, using data augmentation techniques to improve generalization.

UI

The user interface for this project is built using Streamlit. It allows users to upload chest X-ray images and get predictions from the trained models.

Running the Streamlit App

To run the Streamlit app, use the following command:

streamlit run app.py

Screenshots