Deep learning based physician decision support system design for COVID-19 diagnosis on computed tomography images

COVID-19 pandemic negatively affects the whole world in many ways. Since its emergence, it has led to the development of various methods and approaches for the solution of its negativities. The common goal of these search for solutions is to minimize the damages of COVID-19. In this thesis, a deep learning-based system was designed to assist radiologists in the process of detecting COVID-19 disease from chest computed tomography images. This deep learning-based system is built on two different models capable of classification and semantic segmentation. Members of the EfficientNet model family were used to gain the classification capability of the computer and to distinguish the chest computed tomography images with COVID-19 positive or COVID-19 negative findings. Eight different EfficientNet model family members were trained with the EFSCH-19 dataset samples, which we created with real patient images, thanks to the permission given by the Ministry of Health of the Republic of Turkey. EfficientNet-B2 model, which reached 99.75% accuracy, 99.50% sensitivity, 100% precision, 100% specificity, 99.75% F1-Score and 99.50% MCC rate, was chosen as the classifier deep learning model in the test phase carried out as a result of the training process. The chest computed tomography image, which is given as input to the classifier deep learning model, is estimated as COVID-19 positive or COVID-19 negative by the binary classification method at the output. In order to gain the semantic segmentation capability of the computer, the information of which class the pixels in the chest computed tomography images belong to must be defined. In this context, the mask images were manually marked by the radiologist with 200 chest computed tomography images in the EFSCH-19 dataset, and the second dataset, which we named EFSCH-19-Seg, was created. In the mask images created, there are three different pixel values to represent lung areas, infection areas and the remaining areas. U-Net models trainings was carried out with the help of the EFSCH-19-Seg dataset. In these trainings, instead of the default encoder network of the U-Net model, five different models in the ResNet model family were used. Compared to other models, the ResNet-50-based U-Net model was the model that best completed the test phase with 93.76% Jaccard index and 96.61% Dice score values. In this thesis, as a result of the successful results obtained for classification and semantic segmentation tasks, a physician decision support system that can be quickly integrated into the field in the fight against the COVID-19 pandemic has been proposed.