Performance Evaluation of Convolutional Neural Network Architectures for Diagnosis of Childhood Pneumonia

Abstract

Pneumonia, a bacterial or viral infection of the lungs that causes the inflammation of the air sacs, is one of the leading causes of mortality of children in the world. Chest x-rays, one of the golden standard tools in determining pneumonia, is mainly used to detect malignancy in the lungs. However, the process of analyzing may be time-consuming for the radiologist, and costly to hospitals. Inter-observer variability with the diagnosis is very high since childhood pneumonia can be difficult to diagnose amongst radiologists. Considering that the design of convolutional neural networks makes it suited to process spatially distributed input such as images, the application of convolutional neural networks trained with chest X-rays to automate the diagnosis of pneumonia is viable. This study evaluates the performance of four well known architectures in literature using a childhood pneumonia dataset: (1) VGGNet, (2) ResNet, (3) DenseNet, and (4) AlexNet. Based on our simulations, VGGNet obtained the highest accuracy and sensitivity, followed by ResNet, which obtained the highest specificity, DenseNet, and AlexNet. Using gradient-weighted class activation to validate the learnt features, we observed that sufficiently deep architectures can effectively learn the features of pneumonia. In addition, the increase in depth improves the information flow at the cost of computational time, which is evident in DenseNet.