Enhanced Detection of COVID-19 using Deep Learning and Multi-Agent Framework: The DLRPET Approach

Abstract

The ongoing global pandemic caused by novel coronavirus (COVID-19) has emphasized the urgent need for accurate and efficient methods of detection. Over the past few years, several methods were proposed by various researchers for detecting COVID-19, but there is still a scope of improvement. Considering this, an effective and highly accurate detection model is presented in this paper that is based on Deep learning and multi-Agent concepts. Our main objective is to develop a model that can not only detect COVID-19 with high accuracy but also reduces complexity and dimensionality issues. To accomplish this objective, we applied a Deep Layer Relevance Propagation and Extra Tree (DLRPET) technique for selecting only crucial and informative features from the processed dataset. Also, a lightweight ResNet based Deep Learning model is proposed for classifying the disease. The ResNet model is initialized three times creating agents which analyses the data individually. The novelty contribution of this work is that instead of passing the entire training set to the classifier, we have divided the training dataset into three subsets. Each subset is passed to a specific agent for training and making individual predictions. The final prediction in proposed network is made by implementing majority voting mechanism to determine whether an individual is COVID-19 positive or negative. The experimental outcomes indicated that our approach achieved an accuracy of 99.73% that is around 2% higher than standard best performing KISM model. Moreover, proposed model attained precision of 100%, recall of 99.73% and F1-score of 98.59 % respectively, showing an increase of 5% in precision, 4.73% in recall and 4.59% in F1score than best performing SVM model.