Pneumonia detection using capsule networks

Abstract

Pneumonia is an infection that inflames the air sacs known as alveoli in the lungs. Symptoms include fever, chest pain, cough, and shortness of breath. It affects approximately 450 million people globally (7percent of the population) and results in about four million deaths per year [1]. Currently the best available method to diagnose pneumonia and other pathol- ogy types are chest X-rays. In this thesis, we present a machine learning model using capsule networks (CapsNet) to be able to detect pathology like pneumonia and other common disease types using the chest X-ray images. CapsNet were recently proposed to address the shortcomings of the traditional convolutional neural networks (CNNs) in computer vision tasks. One of the main issues of CNNs is that they do not recognize orientation or relative spatial relationships between the different elements of the image, which is an important feature when detecting pneumonia from X-ray images. They also require a large amount of training data, which is not the case for our problem. To train and eval- uate our CapsNet approach, we use a dataset of 5,855 X-ray images [2] of infected and healthy individuals (we refer to this dataset as the pneumonia dataset), and then use a larger dataset [3] of 112,120 X-ray images to detect more common disease types (we refer to this dataset as the ChestX-ray14 dataset). Using CapsNet we were able to achieve 0.94 accuracy on the pneumonia dataset. And on the large dataset, ChestX-ray14, we were able to outperform the ResNet-50 model of Wang et. al (2017) [3] (the creators of the dataset) on average AUROC and on most disease types including pneumonia.