An Improved Classification of Chest X-ray Images Using Adaptive Activation Function

Abstract

Machine vision techniques particularly convolutional neural networks (CNNs) have attained major breakthrough in medical image analysis and classification because of their ability to learn representative features from the input in a hierarchical manner. A couple of years back performing an effective and accurate CNN based classification was a tremendous challenge due to non-availability of large and good quality chest X-ray image (CXR) database. In this paper, we have presented the experiment based on state of the art deep CNN architectures like AlexNet, Res Net and VGG16. These experiments were conducted based on two types of study, one containing dataset with chest Xray images of subjects who contracted Covid-19, viral pneumonia and no respiratory disorder(normal) mentioned as study II and the other dataset containing only Covid-19 and healthy subjects mentioned as study I. A comparison has been drawn with the proposed architecture and classification results based on standard metrics have been carried out on test dataset. The raw chest Xray (CXR) images were passed to the CNN during the training phase without any prior image processing techniques applied on them. Also, we have proposed a new CNN architecture which incorporates the use of an adaptive activation function and it classified the above mentioned studies(I and II) with an accuracy of 96.89 % and 96.75 % and proved to be better than some of the very deep and much more advanced CNN architectures in terms of number of parameters, training time and the amount of space it occupied.