Background/objectives: According to the World Health Organization, the gold standard for diagnosing COVID-19 is the Reverse Transcription Polymerase Chain Reaction (RT-PCR) test. However, to confirm the diagnosis in patients who have negative results but still show symptoms, imaging tests, especially computed tomography (CT), are used. In this study, using convolutional neural networks, we compared the following topics using manual and automatic lung segmentation methods: (1) the performance of an automatic segmentation of COVID-19 areas using two strategies for data partitioning, CT scans, and slice strategies; (2) the performance of an automatic segmentation method of COVID-19 when there was interobserver agreement between two groups of radiologists; and (3) the performance of the area affected by COVID-19.
Methods: Two datasets and two deep neural network architectures are used to evaluate the automatic segmentation of lungs and COVID-19 areas. The performance of the U-Net architecture is compared with the performance of a new architecture proposed by the research group.
Results: With automatic lung segmentation, the Dice metrics for the segmentation of the COVID-19 area were 73.01 ± 9.47% and 84.66 ± 5.41% for the CT-scan strategy and slice strategy, respectively. With manual lung segmentation, the Dice metrics for the automatic segmentation of COVID-19 were 74.47 ± 9.94% and 85.35 ± 5.41% for the CT-scan and the slice strategy, respectively.
Conclusions: The main conclusions were as follows: COVID-19 segmentation was slightly better for the slice strategy than for the CT-scan strategy; a comparison of the performance of the automatic COVID-19 segmentation and the interobserver agreement, in a group of 7 CT scans, revealed that there was no statistically significant difference between any metric.
Keywords: COVID-19; computed tomography; convolutional neural networks; coronavirus; deep learning; image segmentation; lung; radiology.