Automated Segmentation of COVID-19 Lesion from Lung CT Images Using U-Net Architecture

Abstract

Pneumonia caused by the novel Coronavirus Disease (COVID-19) is emerged as a global threat and considerably affected a large population globally irrespective of their age, race, and gender. Due to its rapidity and the infection rate, the World Health Organization (WHO) declared this disease as a pandemic. The proposed research work aims to develop an automated COVID-19 lesion segmentation system using the Convolutional Neural Network (CNN) architecture called the U-Net. The traditional U-Net scheme is employed to examine the COVID-19 infection present in the lung CT images. This scheme is implemented on the benchmark COVID-19 images existing in the literature (300 images) and the segmentation performance of the U-Net is confirmed by computing the essential performance measures using a relative assessment among the extracted lesion and the Ground-Truth (GT). The overall result attained with the proposed study confirms that, the U-Net scheme helps to get the better values for the performance values, such as Jaccard (>86%), Dice (>92%) and segmentation accuracy (>95%).

Deep learning‐based COVID‐19 diagnosis using CT scans with laboratory and physiological parameters

“…For example, some models like VGG, AlexNet, ResNet, and GoogleNet [59] are classified by linking fully connected layers to multiple classifiers. Other models, such as Faster RCNN, R-CNN, and YOLO [60], are typically applied for object detection for another model used for a segmentation task, represented by U-Net [61]. As a feature extractor, all models use the convolutional part of the CNN core architecture, VGG, AlexNet, ResNet, and GoogleNet.…”