The Differentiation of Viral and Bacterial Pneumonia using Deep Learning

This project aims to find out whether a Convolutional Neural Network can be used to classify x-ray scans as having either bacterial or viral Pneumonia.

This project aims to find out whether a Convolutional Neural Network can be used to classify x-ray scans as having either bacterial or viral Pneumonia. First, its use is discussed, as it is meant to be used to assist with pulmonary healthcare in underprivileged areas. Then, its structure is outlined as a CNN set up using Keras and Tensorflow that is trained on a dataset consisting of 5856 x-ray scans of patients one to five years old from Guangzhou Women and Children’s Medical Center. After testing multiple setups, the final model, based on Inception ResNet v2, differentiates between the two types of Pneumonia with a 75% test accuracy, using minimal epochs in order to increase efficiency. The model was further improved by using dropout, the implementation of early stopping, L2 regularization and augmentation of the training data in order to combat overfitting. Lastly, the model’s limitations and future are discussed. This model proves one can predict and identify types of Pneumonia using a CNN with relatively high accuracy and shows the potential of deep learning in the context of medical imaging to assist medical staff.