Deep Learning Feature Extraction for COVID19 Detection Algorithm using Computerized Tomography Scan

Maisarah Mohd Sufian, Ervin Gubin Moung, Chong Joon Hou, Ali Farzamnia

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

8 Citations (Scopus)


COVID19 is a new virus that has infected over three million people worldwide and still infecting. The most used method for detecting infected persons is reverse transcriptase-polymerase chain reaction (RT-PCR). However, there is an acute scarcity of RT-PCR test kits worldwide. Therefore, Computerized Tomography (CT) scans have been used widely in hospitals to diagnose respiratory illness, among others. Many studies have been carried out where CT scan is used for COVID19 detection. Due to its nature in the extraction of image attributes, deep learning (DL) was considered a powerful method for improving the accuracy of COVID19 diagnosis with CT scans. This research compared the performance of the three most popular DL models, (i) Custom CNN, (ii) ResNet50, and (iii) VGG16, in the classification of COVID19 positive and COVID19 negative patients. The convolutional base of each model is used to obtain features of all images from the SARS-CoV-2 CT-scan dataset. The performance and effectiveness of tuned DL models are evaluated using a 10-fold cross-validation performed on a dataset consisting of training and validation images. Additionally, the tuned DL models were tested on a previously unseen test dataset. VGG16 is the best performing model in a 10-folds cross validation, with an average accuracy of 95.48%. Furthermore, VGG16 is also the best model when evaluated on an unseen testing dataset, yielding specificity and accuracy with 97% and 92.5%, respectively. Finally, the Custom CNN outperformed all other models in terms of sensitivity with 90% rate.

Original languageEnglish
Title of host publication11th International Conference on Computer Engineering and Knowledge
Subtitle of host publicationICCKE 2021
PublisherInstitute of Electrical and Electronics Engineers Inc.
Number of pages6
ISBN (Electronic)9781665402088
ISBN (Print)9781665402095
Publication statusPublished - 28 Feb 2022
Externally publishedYes
Event11th International Conference on Computer Engineering and Knowledge - Mashhad, Iran, Islamic Republic of
Duration: 28 Oct 202129 Oct 2021
Conference number: 11

Publication series

NameInternational Conference on Computer Engineering and Knowledge
ISSN (Print)2375-1304
ISSN (Electronic)2643-279X


Conference11th International Conference on Computer Engineering and Knowledge
Abbreviated titleICCKE 2021
Country/TerritoryIran, Islamic Republic of

Cite this