Detecting COVID-19 in Computed Tomography Images: A Novel Approach Utilizing Segmentation with UNet Architecture, Lung Extraction, and CNN Classifier

Abstract

Our study introduces an innovative framework tailored for COVID-19 diagnosis utilizing a vast, meticulously annotated repository of CT scans (each comprising multiple slices). Our framework comprises three key Parts: the segmentation module (based on UNet and optionally incorporating slice removal techniques), the lung extraction module, and the final classification module. The distinctiveness of our approach lies in augmenting the original UNet model with batch normalization, thereby yielding lighter and more precise localization, essential for constructing a comprehensive COVID-19 diagnosis framework. To gauge the efficacy of our framework, we conducted a comparative analysis of other possible approaches. Our novel approach segmenting through UNet architecture, enhanced with Batch Norm, exhibited superior performance over conventional methods and alternative solutions, achieving High similarity coefficient on public data. Furthermore, at the slice level, our framework demonstrated remarkable validation accuracy and at the patient level, our approach outperformed other alternatives, surpassing baseline model. For the final diagnosis decisions, our framework employs a Convolutional Neural Network (CNN). Utilizing the COV19-CT Database, characterized by a vast array of CT scans with diverse slice types and meticulously marked for COVID-19 diagnosis, our framework exhibited enhancements over prior studies and surpassed numerous alternative methods on this dataset.