JoWUA

Abstract

Evolution of artificial intelligence has significant impact on medical field, especially in disease classification. Lung disease often referred as respiratory infections, sound severe threat to human life. Early interpretation is very much essential for patient survival. Deep network models, especially the technique of ensemble aid the practitioners in automating disease prediction process. In this work, various ensemble strategies such as average, weighted average and stacked ensemble model for four class lung disease classification are proposed. Deep convolutional neural network (CNN) models namely Resnet101, Inception V3 and Mobilenet V2 are used as first level or base models for the ensemble. To study the potency of different meta learners on ensemble stacking, five machine learning models namely Logistic Regression, Support Vector Machine (SVM), Linear Discriminant Analysis, Decision Tree and Naive Bayes are opted as the meta learner models. Here, QaTa-COV 19 dataset is utilized which comprises of 21165 chest x-ray images under four classes - covid19, lung opacity, healthy and viral pneumonia. The experimental results demonstrated that proposed average, weighted average and deep stacked ensemble models outperform, the base predictors and among best, stacked ensemble with support vector machine as meta learner with an overall model accuracy of 98.6%. The proposed model could reduce the variations in heterogeneous base model predictions and resulted in a highly efficient stacking ensemble model with better generalization capability for lung disease prediction.