An integrative method for COVID-19 patients’ classification from chest X-ray using deep learning network with image visibility graph as feature extractor

We propose a method by integrating image visibility graph and deep neural network (DL) for classifying COVID-19 patients from their chest X-ray images. The computed assortative coefficient from each image horizonal visibility graph (IHVG) is utilized as a physical parameter feature extractor to improve the accuracy of our image classifier based on Resnet34 convolutional neural network (CNN). We choose the most optimized recently used CNN deep learning model, Resnet34 for training the pre-processed chest X-ray images of COVID-19 and healthy individuals. Independently, the preprocessed X-ray images are passed through a 2D Haar wavelet filter that decomposes the image up to 3 labels and returns the approximation coefficients of the image which is used to obtain the horizontal visibility graph for each X-ray image of both healthy and COVID-19 cases. The corresponding assortative coefficients are computed for each IHVG and was subsequently used in random forest classifier whose output is integrated with Resnet34 output in a multi-layer perceptron to obtain the final improved prediction accuracy. We employed a multilayer perceptron to integrate the feature predictor from image visibility graph with Resnet34 to obtain the final image classification result for our proposed method. Our analysis employed much larger chest X-ray image dataset compared to previous used work. It is demonstrated that compared to Resnet34 alone our integrative method shows negligible false negative conditions along with improved accuracy in the classification of COVID-19 patients. Use of visibility graph in this model enhances its ability to extract various qualitative and quantitative complex network features for each image. Enables the possibility of building disease network model from COVID-19 images which is mostly unexplored. Our proposed method is found to be very effective and accurate in disease classification from images and is computationally faster as compared to the use of multimode CNN deep learning models, reported in recent research works.