DETERMINATION OF THORACIC COMPUTERIZED TOMOGRAPHY FINDINGS IN COVID-19
CASES BY DEEP LEARNING
Abstract
Aims: The effects of the COVID-19 pandemic continue around the world.
Imaging methods play an important role in the diagnosis of COVID-19. The
aim of this study was to develop a system that would allow for the
distinguishing of lesions at different stages of the disease based on
similar signs of other viral diseases and monitoring the emergence,
progression, and/or remission of lesions in different areas of the
lungs. Methods: For the deep learning (DL) system, the thoracic CT
images from 1,382 images were reviewed. These belonged to patients whose
SARS-CoV-2 RT-PCR tests turned out positive, were diagnosed with
COVID-19, and had signs of lung involvement. Of 1,382 images in the
dataset, 180 were assigned for testing and 1,202 were assigned for
training. Apart from our dataset, 131 images for internal testing and
1,365 images for external testing were used. The trainings were
continued to cover 316,000 steps. Results: Internal and external
analyses were used to assess the developed model. The internal analysis
success rate was 93.12%. For first external analysis we used 85 images.
In the first external analysis we assessed a single CT image of each
patient who was in the mixed image lists, and the success rate was found
to be 70.31%. In the second external analysis, 645 thoracic CT images
of patients diagnosed with COVID-19 and 635 images of another patients
who had signs of non-COVID-19 diseases were used. We assessed the
thoracic CT images with both COVID-19 and non-COVID-19 disease signs.
The success rate in the identification of COVID-19 patients was 88.4%.
Conclusion: Special modeling systems developed using DL may help
accelerate workflow and making the process easier. This is especially
important in cases in which fast and accurate assessment is essential
for of a large number of patients, as happens in a pandemic.