How does artificial intelligence effect on the assessment of lung damage in COVID-19 on chest CT scan?

Abstract

BACKGROUND: During the pandemic, computed tomography (CT) was one of the most important tools for assessing COVID-19-related lung changes. In COVID-19 patients, radiologists in Moscow used the adapted CT0-4 scale to visually as-sess the dependence of the severity of the general condition on the nature and severity of radiological signs of changes in the lungs based on computed tomography. In a large stream of scans, the doctor may miss findings and make errors in assessing the volume of lung damage, so the use of AI services in outpatient healthcare during a pandemic can be beneficial. AIM: The goal of this study is to compare the distribution of CT0-4 categories designed by radiologists with the results of AI services processing and categories formed without AI services. Methods: We used retrospective study design, full study protocol is registered on ClinicalTrials.gov (NCT04489992). The results of primary CT scans with the CT0-4 categories were analyzed in outpatient medical institutions of the Health Department from April 08, 2020, to December 01, 2020, and separately for November (from November 01, 2020, to December 01, 2020). CT was performed on 48 computed tomographs in accordance with standard protocols, and the data was processed by the single radiology information systems. CTs in the test group received AI services, while CTs in the control group did not. The analysis includes five AI services: RADLogics COVID-19 (RADLogics, USA), COVID-IRA (IRA labs, Russia), Care Mentor AI, COVID (Care Mentor AI, Russia), Third Opinion. CT-COVID-19 (Third Opinion, Russia), and COVID-MULTIVOX (Gammamed, Rus-sia). Moreover, AI services are encoded at random. Results: The CT scan results of 260,594 patients were examined (m/f % = 44/56, mean age = 49.5). The test group consisted of 115,618 CT scans, while the control group consisted of 144,976 CT scans. Depending on the specific AI service, CT0 was established by 2.3–18.5% less than the control group for different subgroups of categories. The categories CT3-4 were established by 4.7–27.6% less than without AI, and the categories CT4 by 40–60% less than without AI (p <0.0001). For November (from November 01, 2020, to December 01, 2020), the CT scan results of 41,386 patients were analyzed (m/f % = 44/56, average age = 53.2 years). The test group consisted of 28,881 CT scans, while the control group included 12,505 CT scans. Depending on the specific AI service, CT0 was established by 1–2.6% less than the control group for different subgroups of categories. Further, the categories CT3–CT4 were established by 0.2–15.7% less than without AI, and the categories CT4 were established by 25% less than without AI (p = 0.001). Conclusion: The use of AI services for primary CT scans on an outpatient basis reduces the number of CT0 and CT3–CT4 results, which can influence the therapeutic approach for COVID-19 patients.