Lung Infection Detection using Progressive UNet Architecture

Abstract

The fragmentation of medical images of tissue anomalies in organs or the blood vascular system is critical for any computerized diagnostic system. Nevertheless, automated segmentation in medical image analysis is complex since it requires in-depth information about the target organ architecture. This paper presents UNet, an end-to-end deep learning segmentation technique for early recognition of COVID. The proposed UNet model is progressive and capable of diagnosing different lung infection types along with COVID-19 infection. For this, computed tomography images are considered. The XGBoost classifier is integrated with UNet for feature classification in this model. The result analysis was performed on different convolution neural network models, that is, ResNet50, Inception, ResNet101, ResNet152, DenseNet, and UNet. These models were implemented on MATLAB using the deep neural network toolbox. From the result, it has been noticed that Inception achieved a minimum accuracy of 85.2%, and UNet achieved 99% of accuracy. It has been observed that UNet achieved approx. 16% of improvement over the inception model.