Identification of images of COVID-19 from Chest Computed Tomography (CT) images using Deep learning: Comparing COGNEX VisionPro Deep Learning 1.0 Software with Open Source Convolutional Neural Networks

TL;DR

This study evaluates the performance of COGNEX VisionPro Deep Learning software in identifying COVID-19 from chest CT images, comparing it with COVIDNet-CT and other models, showing high accuracy even with reduced training data.

Contribution

It demonstrates the effectiveness of COGNEX VisionPro Deep Learning software for COVID-19 detection from CT images, highlighting its ease of use and high accuracy compared to open-source models.

Findings

VisionPro Deep Learning achieved over 99% F-score.

The software maintained high accuracy with fewer training images.

It outperformed other state-of-the-art models in COVID-19 detection.

Abstract

For testing patients infected with COVID-19, along with RT-PCR testing, chest radiology images are being used. For the detection of COVID-19 from radiology images, many organizations are proposing the use of Deep Learning. University of Waterloo and DarwinAI, have designed their own Deep Learning model COVIDNet-CT to detect COVID-19 from infected chest CT images. Additionally, they have introduced a CT image dataset COVIDx-CT, from CT images collected by the China National Center for Bioinformation. COVIDx-CT contains 104,009 CT image slices across 1,489 patient cases. After obtaining remarkable results on the identification of COVID-19 from chest X-ray images by using the COGNEX VisionPro Deep Learning Software 1.0 this time we test the performance of the software on the identification of COVID-19 from CT images. COGNEX Deep Learning Software: VisionPro Deep Learning, is a Deep Learning software that is used across various domains ranging from factory automation to life sciences. In this study, we train the classification model on 82,818 chest CT training and validation images from the COVIDx-CT dataset in 3 classes - normal, pneumonia, and COVID-19 and then test the results of the classification on the 21,191 test images are compared with the results of COVIDNet-CT and various other state of the art Deep Learning models from the open-source community. Also, we test how reducing the number of images in the training set effects the results of the software. Overall, VisionPro Deep Learning gives the best results with F-scores over 99%, even as the number of images in the training set is reduced significantly. This software is by no means a stand-alone solution in the detection of COVID-19 but can aid radiologists and clinicians in achieving faster and understandable diagnosis using the full potential of Deep Learning, without the prerequisite of having to code in any programming language.