A comparative study of deep learning models for COVID-19 diagnosis based on x-ray images

Shah Sufi Nesar Uddin Siddiqui, Elias Hossain, Rezowan Ferdous, Murshedul Arifeen, Wahidur Rahman, Shamsul Kabir Masum, Adrian Alan Hopgood, Alice Good, Alexander Gegov

Research output: Chapter in Book/Report/Conference proceedingConference contribution

21 Downloads (Pure)


Background: The rise of COVID-19 has caused immeasurable loss to public health globally. The world has faced a severe shortage of the gold standard testing kit known as reverse transcription-polymerase chain reaction (RT-PCR). The accuracy of RT-PCR is not 100%, and it takes a few hours to deliver the test results. An additional testing solution to RT-PCR would be beneficial. Deep learning’s superiority in image processing is characterised as the most effective COVID-19 diagnosis based on images. The small number of COVID-19 X-ray images in existing deep learning methods for COVID-19 diagnosis may degrade the performance of deep learning methods for new sets of images. Our priority for this research is to test and compare different deep learning algorithms on a dataset consisting of many COVID-19 X-ray images.

Methods: We have merged the publicly available image data into two groups (COVID and Normal). Our dataset contains 579 COVID-19 cases and 1773 Normal cases of X-ray images. We have used 145 COVID-19 cases and 150 Normal cases to test the deep learning models. Deep learning models based on CNN, VGG16 and 19, and InceptionV3 have been considered for prediction. The performance of these models is compared based on measurements of accuracy, sensitivity, and specificity. In the deep learning models, the SoftMax activation function is used along with the Adam optimiser and categorical cross-entropy loss. A customised hybrid CNN model found in literature is considered and compared to explore how the inclusion of many COVID-19 X-ray images could impact the model’s performance.

Results: The accuracy of the considered deep learning models using InceptionV3, VGG16, and VGG19 algorithms achieved 50%, 90%, and 83%, respectively, in predicting the X-ray images of COVID-19. We have shown that number of COVID-19 X-ray images does have a significant impact on the model’s performance. A customised hybrid CNN model found in the literature failed to perform well on a dataset consisting of a large number of COVID-19 X-ray images. The customised hybrid CNN model reached an accuracy of 71% on many COVID-19 X-ray images. In contrast, it achieved 98% accuracy on a small number of COVID-19 X-ray images. It is also observed from the experiments that the VGG16 performs well with an increased number of images.

Conclusions: A maximised number of COVID-19 X-ray images should be considered in building a deep learning model. The deep learning model with VGG16 performs the best in predicting from the X-ray images.
Original languageEnglish
Title of host publicationSmart and Sustainable Technology for Resilient Cities and Communities
EditorsRobert J. Howlett, Lakhmi C. Jain, John R. Littlewood, Marius M. Balas
ISBN (Electronic)9789811691010
ISBN (Print)9789811691003, 9789811691034
Publication statusPublished - 1 Jan 2022

Publication series

NameAdvances in Sustainability Science and Technology
PublisherSpringer Nature
ISSN (Print)2662-6829
ISSN (Electronic)2662-6837


  • Coronavirus (COVID-19)
  • RT-PCR
  • Machine Learning (ML)
  • Deep Learning (DL)
  • X-ray images


Dive into the research topics of 'A comparative study of deep learning models for COVID-19 diagnosis based on x-ray images'. Together they form a unique fingerprint.

Cite this