A new approach based on transfer learning methods and ensemble learning in the diagnosis of lower respiratory tract infections from x-ray images

Today, diseases caused by viruses cause major epidemics, affecting the lives of millions of people. Some of the diseases caused by these viruses are lower respiratory tract infections, and the COVID-19 pandemic caused by the Corona virus is one of the most acute and severe viruses of recent years. Despite the development of virus vaccines, the rates of COVID-19 cases are increasing rapidly around the world. It is seen that artificial intelligence techniques are also used for the diagnosis of COVID-19 and other lower respiratory tract diseases. Especially deep learning techniques produce faster and more successful results than classical PCR testing and manual interpretation of X-ray images. Methods such as deep structured learning, interpretive learning, and transferred learning, also known as deep learning, are artificial neural network-based methods. In this study, COVID-19 and other lower respiratory tract infections were diagnosed using X-ray images with an ensemble classifiers hybrid model based on 9 different deep learning methods. The hybrid approach called DeepFeat-E performs diagnosis using deep features obtained from transfer models and classifiers consisting of classical machine learning methods. A dataset of 21,165 X-ray images in total of 10,192 Normal, 6012 Lung Opacity (Non-COVID lung infection), 1345 Viral Pneumonia and 3616 COVID-19 (Patients) were used to test the proposed approach. With the proposed approach, it was seen that the highest success was obtained with the deep features and Stacking ensemble learning method of DenseNet201 TL(Transfer Learning) models. In experimental studies with datasets with four, three and two classes, the test accuracy was 90.17%, 94.99% and 94.93%, respectively. In addition, it was observed that the system increased the accuracy values obtained in all DR models by varying amounts. According to the results obtained within the scope of this thesis study, it has been shown that the proposed DeepFeat-E hybrid system can be used quickly and reliably in the diagnosis of lower respiratory tract infectious diseases.