COVID-Rate: An Automated Framework for Segmentation of COVID-19 Lesions from Chest CT Scans

TL;DR

This paper introduces COVID-Rate, an automated deep learning framework for accurate segmentation of COVID-19 lung lesions from chest CT scans, aiding diagnosis and resource allocation during the pandemic.

Contribution

The paper presents a novel open access dataset and a deep neural network framework for autonomous COVID-19 lesion segmentation in CT images, demonstrating high accuracy and generalization.

Findings

Dice score of 0.802 on test data

High specificity of 0.997 in lesion detection

Effective segmentation on external datasets

Abstract

Novel Coronavirus disease (COVID-19) is a highly contagious respiratory infection that has had devastating effects on the world. Recently, new COVID-19 variants are emerging making the situation more challenging and threatening. Evaluation and quantification of COVID-19 lung abnormalities based on chest Computed Tomography (CT) scans can help determining the disease stage, efficiently allocating limited healthcare resources, and making informed treatment decisions. During pandemic era, however, visual assessment and quantification of COVID-19 lung lesions by expert radiologists become expensive and prone to error, which raises an urgent quest to develop practical autonomous solutions. In this context, first, the paper introduces an open access COVID-19 CT segmentation dataset containing 433 CT images from 82 patients that have been annotated by an expert radiologist. Second, a Deep Neural Network (DNN)-based framework is proposed, referred to as the COVID-Rate, that autonomously segments lung abnormalities associated with COVID-19 from chest CT scans. Performance of the proposed COVID-Rate framework is evaluated through several experiments based on the introduced and external datasets. The results show a dice score of 0:802 and specificity and sensitivity of 0:997 and 0:832, respectively. Furthermore, the results indicate that the COVID-Rate model can efficiently segment COVID-19 lesions in both 2D CT images and whole lung volumes. Results on the external dataset illustrate generalization capabilities of the COVID-Rate model to CT images obtained from a different scanner.