Improving COVID 19 Detection based on a hybrid data mining approach

Abstract

The worldwide spread of coronavirus disease (COVID-19) has become a threatening risk for global public health. Currently, doctors resort to PCR analysis, however, it suffers from low accuracy problems. On the other hand, Convolutional neural network (CNN) and despite its high accuracy incorrect classification, it takes a long time to train data, in addition it requires large training dataset. In this paper, we propose a hybrid approach for COVID-19 detection and diagnosis. Our contribution consists of two phases to provide high detection accuracy. In the first phase, we propose a hybrid features-fusion phase that works by fusing four common features extracted from medical image, Row pixel intensity, Colour histogram, Harlick texture and Threshold. Each single classifier is fed with these four features and yielded a 4 different predictions for each feature. A well-known voting technique is then applied to provide final predication result for each classifier. Secondly, the ensemble stacking technique is employed to fuse predication of each classifier, which significantly improves final detection accuracy. The proposed approach has been quantitatively evaluated on a public dataset of 5000 CT-images. The proposed approach yields accuracy of 99.3% and overcome traditional approaches such as KNN (K-nearest neighbours) that yields 92%, and SVM (Support vector machines) that yields 92% comparable computational time that is approximately 4.9 minutes.