X-COVNet: Externally Validated Model for Computer-Aided Diagnosis of Pneumonia-Like Lung Diseases in Chest X-Rays Based on Deep Transfer Learning

Abstract

Since the appearance of COVID-19, the accurate diagnosis of pneumonia-type lung diseases by chest radiographs has been a challenging task for experts, mainly due to the similarity of patterns between COVID-19 and viral or bacterial pneumonia. To address this challenge, a model for the computer-aided diagnosis of chest X-Rays has been developed in this research. This model might contribute to substantially increasing the accuracy of the diagnosis. This approach is based on supervised learning using neural networks, where the quality of the result depends on the quality of the dataset used during training. Image data augmentation techniques, hyperparameter adjustments and dropout layer contributed to achieve high performance values on test data in multi-class classification. The experiments conducted to evaluate the model yielded that it detects and classifies domain classes with an accuracy of 99.45% on training data, 99.27% on validation data and 99.06% on selected test data. The main contribution of this paper is a new Deep Convolutional Neural Network model using Deep Transfer Learning through the Xception architecture for the assisted diagnosis of COVID-19, pneumonia or healthy patients, trained on COVID-19 Chest X-Ray Database and evaluated through two external databases, which give the model novelty within the lack of external validation in all the literature reviewed.