Deep convolutional neural networks for detection of abnormalities in chest X-rays trained on the very large dataset

Kadir Aktas; Vuk Ignjatovic; Dragan Ilic; Marina Marjanovic; Gholamreza Anbarjafari

doi:10.1007/s11760-022-02309-w

Deep convolutional neural networks for detection of abnormalities in chest X-rays trained on the very large dataset

Signal Image Video Process. 2023;17(4):1035-1041. doi: 10.1007/s11760-022-02309-w. Epub 2022 Jul 20.

Authors

Kadir Aktas^{1

2}, Vuk Ignjatovic³, Dragan Ilic³, Marina Marjanovic³, Gholamreza Anbarjafari^{1

2

4

5}

Affiliations

¹ iCV Research Lab, Institute of Technology, University of Tartu, 51009 Tartu, Estonia.
² iVCV OÜ, 51011 Tartu, Estonia.
³ Singidunum University, Belgrade, 11010 Serbia.
⁴ PwC Advisory, Helsinki, Finland.
⁵ Yildiz Technical University, Istanbul, Turkey.

Abstract

One of the main challenges in the current pandemic is the detection of coronavirus. Conventional techniques (PT-PCR) have their limitations such as long response time and limited accessibility. On the other hand, X-ray machines are widely available and they are already digitized in the health systems. Thus, their usage is faster and more available. Therefore, in this research, we evaluate how well deep CNNs do when it comes to classifying normal versus pathological chest X-rays. Compared to the previous research, we trained our network on the largest number of images, 103,468 in total, including 5 classes such as COPD signs, COVID, normal, others and Pneumonia. We achieved COVID accuracy of 97% and overall accuracy of 81%. Additionally, we achieved classification accuracy of 84% for categorization into normal (78%) and abnormal (88%).

Keywords: Automatic diagnosis of COVID-19; Optimization of convolutional neural network; Swarm intelligence; X-ray.