Implementation and Comparison of U-net networks for Automatic COVID-19 Lung Infection Segmentation

Abstract

With the big number of COVID-19 patients, efficient detection tools are necessary. In this work, we proposed an automatic detection and quantification tool based on deep learning model. The architecture used is U-Net architecture, one of the most known for medical applications. This network was introduced as a binary semantic segmentation tool. It uses a dataset of 100 images, seventy-two of them for training, ten for validation, and eighteen for testing. The model will be compared with other deep learning models and evaluated using several evaluation metrics. The results have shown an accuracy of 0.958, sensitivity of 0.900, dice coefficient of 0.726, and a specificity of 0.962. Compared with other related works, our network has the best accuracy and specificity. The obtained results show the ability of the model as a binary segmentation tool to separate infection tissue and healthy tissue.