Performance analysis of Pneumonia using Convolutional Neural Networks

Abstract

Pneumonia is a bacterial or viral infection disease occurs in human lungs. Pneumonia is a critical issue and affecting the human lungs, most of the time infections caused by bacteria called Streptococcus pneumonia. Early stage of pneumonia diagnosis is very much essential for effective and efficient treatment process to avoid future problems. According to World Health Organization (WHO), in India death rate is one among three due to pneumonia infection. Along with primary check-up and test, chest X-ray image is also needed by a doctor for the disease evaluation. Chest X-ray is one of the best tools in examining pneumonia, because examining is quite cheaper, and this service will be available in most of the hospitals and clinic. However, several traditional and manual diagnosis methods have been used for the pneumonia diagnosis. The traditional and manual method of diagnosis requires human time and intervention. This work proposes the development of the automated pneumonia diagnostic method using different Convolutional Neural Networks (CNN), which helps doctors for treatment and planning diagnosis. CNN architecture is developed for automatic pneumonia detection, avoid manual pre-processing, classification and segmentation, instead of it use chest radiography pixel information. CNN architectures are a multi-layer architecture with and without data augmentation was carried out with casual and pneumonia plain chest radiography for network training. CNN network plays an important role in diagnosing the disease. The work includes four CNN networks, namely DenseNet 121, VGG 16, Inception V3 and ResNet 50 with and without data argumentation was carried out for pneumonia diagnose. By the end of 10^th epoch the accuracy reached 90% in all the cases but in DenseNet it reached 98.5% (training set) and in Inception it reached 98.6% on validation set indicating that the Inception network has low variance compared to others.