Chest X-ray Classification of Pneumonia and COVID19 Using Modified Capsule Networks

Abstract

Many studies are already done on Deep Learning-based diagnosis, specially using Convolutional Neural Network (CNN), to assist identifying lung disease cases based on radiology imaging. In this study three types of chest X-ray images are taken to be classified by convolutional neural network (CNN), e.g. 1583 normal or healthy chest X-rays, 4273 pneumonia diagnosed chest X-rays and 262 COVID19 diagnosed chest X-ray images. Five various proved architectures (VGG16, VGG19, Xception, InceptionV3, Inception-ResNetV2) are tested on diagnosis of the above classes of X-rays images. Then this above five convolutional architectures are used as feature extractors for a capsule layer of 16 capsule dimension and 4 routings. Total ten CNN architectures are tested to perform the task. The main advantages of capsule networks is that the part-whole relation can be captured through the capsules of consecutive layers. Among the tested main five CNNs VGG16 performs the best with 96.65% accuracy over this task. Among the other five capsulated CNNs VGG16 based capsule network outperforms any other architecture tested with an accuracy of 96.81%. Hopefully the proposed CNN architecture may be an alternative method to diagnose any X-ray classification by providing fast and accurate screening.