COVID-19 detection in radiological images with deep learning

New Coronavirus Disease (COVID-19) is an RNA type virus that is spreading worldwide. COVID-19, which was first seen in Wuhan, China, in December 2019, spread rapidly and soon began to be seen in all countries of the world. Many countries were negatively affected by this situation in terms of health, economy, sociology and psychology. Symptoms such as respiratory tract infections, fever, cough, and shortness of breath are important in diagnosing the disease. Many methods have been worked on to minimize the negativities experienced since 2019. The most important precaution to be taken to minimize the damage caused by the disease to people is to diagnose those who show symptoms as soon as possible. PCR is usually the first procedure performed to diagnose patients who apply to the hospital with the mentioned suspicions. Laboratory examinations, X-ray and Computed Tomography images take a long time, which has led researchers to other diagnostic methods. In this thesis study, a model that can help physicians detect the disease through Computed Tomography (CT) images has been designed. This model, based on deep learning, aims to detect the disease by classifying COVID-19 positive (infected) and COVID-19 negative (healthy) chest CT images. Seeing ground glass pneumonia on CT images speeds up the diagnosis process. The data set used in the thesis study was carried out as a result of the permissions obtained according to the ethics committee decision of Tokat Gaziosmanpaşa University project number 23-KAEK-033 and number 83116987-092. The data set consists of CT images of people examined at Tokat Gaziosmanpaşa University Research and Application Hospital. The data set was created from CT images of patients and healthy people diagnosed as COVID-19 positive by the relevant polyclinic. In the data set created from the obtained BT images, the images in dcm format were first converted to jpg format, which is more suitable for the study. With the images whose extensions were edited, classes were created in two folders: Covid-19 positive and Covid-19 negative. Images with complex and meaningless names have been renamed. In the dataset prepared for data processing, comparisons were made on deep learning models such as AlexNet, Densenet201, GoogleNet, ResNet-50, Vgg16, EfficientNet and the success rates obtained by training on the proposed model. The result in the test data; This rate was 88.81 percent with the Alexnet model, 86.16 percent with the DenseNet201 model, 98.31 percent with the EfficientNet model, 95.52 percent with the GoogleNet model, 94.02 percent with the Resnet-50 model and 94.02 percent with the Resnet-50 model. A success rate of 92.53% was achieved with the VGG-16 model. When the confusion matrix of the proposed ESA model was examined, it was seen that 372 of 375 COVID-19 positive patients were classified correctly and 3 were classified incorrectly. Of the 375 healthy samples, 372 were correctly classified and 3 were incorrectly classified as COVID-19 positive. A 99.20% success rate was achieved with the proposed ESA model, and an effective and successful model was proposed for COVID-19 detection.