Multiclass Classification of COVID-19, Pneumonia, or Nornal Lungs Based on Chest X-Ray Images with Ensemble Deep Learning

ABSTRACT

Coronavirus Disease of 2019 (COVID-19) has a high transmission and death rate. It is important to diagnose COVID-19 accurately and distinguish it clearly from other common lung diseases, e.g., pneumonia. Both diseases are detectable from chest X-Ray images. Therefore, an ensemble deep learning model is applied for multiclass classification of COVID-19, pneumonia, or normal lungs based on chest X-Ray images. ResNet50, VGG16, and InceptionV3 pretrained CNN models are employed to form an ensemble model. The chest X-Ray images are preprocessed in three steps, i.e., cropping, resizing, and normalization. Then, the pretrained models are trained with a new classifier at the top layer of the model. After the classifier is trained, then the pretrained ResNet50, VGG16, and InceptionV3 are fine-Tuned. Lastly, the decisions from each model are assembled using Soft Voting. The ensemble deep learning model which produces the best result, which is formed by combining pretrained and fine-Tuned ResNet50, VGG16, and InceptionV3 models, results weighted accuracy of 0.9752, weighted sensitivity of 0.9612, and weighted specificity of 0.9804. © 2022 IEEE.