Chest X-rays Pneumonia Detection using Convolutional Neural Network

A basic Tensorflow and Keras Implementation for beginners

Convolutional Neural Network (CNN) might seem intimidating for a beginner. However, this project will provide an overview of how to build a model from scratch to detect pneumonia using Tensorflow and Keras.

The content of the mini-project is listed as follows :

• Model Building
• Model Training
• Model Evaluation
• Final Model Prediction

What is Pneumonia?

Pneumonia is a lung inflammation caused by a viral or bacterial infection that can range from mild to severe cases. This inflammation makes the patient unable to breathe enough oxygen to reach the bloodstream. It happens when an infection makes the air sacs (alveoli) in the lungs fill with fluid or pus that might affect either one or both lungs. If your doctor thinks you might have pneumonia, a chest X-ray will be performed to find the infection in the patient's lungs and how far it’s spread.

Dataset

This experiment leveraging the data from Kaggle repository titled Chest X-Ray Images (Pneumonia). Check out the dataset here . The dataset composes of two classes which are normal lung and pneumonia lung as can be seen in the figure below.

Implementation

Before we start building the convolutional neural network, the following are the necessary libraries and dependencies.

Model Development

So let’s jump into the code. The model that we are going to develop composes several components which are:

• tf.keras.layers.Conv2D() : The convolution layer which improves image recognition by isolate images features
• tf.keras.layers.MaxPooling2D() : a layer to reduce the information in an image while maintaining features
• tf.keras.layers.Flatten() : flatten the result into 1-dimensional array
• tf.keras.layers.Dense() : add densely connected layer

A five-layer convolution network will be built where Conv2D() and MaxPooling2D() are stack together as one layer. Then, the output of the final convolutional layer will be flattened and fit to fully connected neurons.

Additionally, before the model is fitted for training, it is necessary to configure the specifications as follows:

• loss: pneumonia detection is using sigmoid activation in the final step, which resulted in either 0 or 1 (normal or pneumonia). Therefore, binary_crossentropy is the most suitable loss function
• optimizer: RMSprop (Root Mean Square Propagation) with a learning rate of 0.001 will be used
• metrics: accuracy is the measurement metric to obtain the prediction accuracy rate on every epoch

Following is the output of model.summary() .

The overview of the model that we have built can be illustrated as below: