2055. Diagnostic Accuracy of Artificial Intelligence in Detecting Pulmonary Embolism on Contrast Chest CT in Hospitalized Patients with COVID-19
Authors * Denotes Presenting Author
  1. Karim Zaazoue; Mayo Clinic
  2. Matthew McCann; Mayo Clinic
  3. Ahmed Ahmed; Mayo Clinic
  4. Isabel Cortopassi; Mayo Clinic
  5. Brent Little; Mayo Clinic
  6. Justin Stowell; Mayo Clinic
  7. Charles Ritchie *; Mayo Clinic
Identify the sensitivity and specificity of an Artificial Intelligence (AI)-powered algorithm for the detection of pulmonary embolism (PE) on contrast-enhanced chest computed tomography (CECCT) in hospitalized patients with COVID-19.

Materials and Methods:
Using a prospectively maintained database of patients hospitalized with COVID-19, a retrospective analysis for pulmonary embolism CECCTs examinations from March 2020 - December 2021 was performed. Based upon the original radiology reports, all studies positive for PE were included (n = 524) and a randomly selected control group was created using studies negative for PE (n = 980). Of the 1504 studies, 1404 (93.4%) were CT pulmonary angiographies (CTPAs) and 100 (6.6%) were contrast-enhanced CTs. The studies were evaluated using an AI detection software with a solution based on a proprietary deep convolution neural network that was trained on tens of thousands of prior studies in patient without COVID-19. Cases were anonymized, uploaded for analysis and reported as a binary classification for PE (+/-). Using a semi-quantitative scoring system total severity score (TSS), the pulmonary airspace disease from COVID-19 was scored 0 - 25 and accordingly were classified as mild (0 - 8), moderate (8 - 16), or severe (17 - 25). When discrepancies between the AI and original radiologist’s results were found, two blinded cardiothoracic radiologists reviewed the studies and rendered a final interpretation of positive, negative, or indeterminate. When their results did not align, a third cardiothoracic radiologist rendered a final interpretation.

There were 83 examinations that were discrepant, 13 of which were deemed as indeterminate by our blinded radiologists and were excluded. Of the 526 examinations reported positive for PE, the AI software identified 490 exams as positive (sensitivity 93.2%; 95% confidence interval [CI] 90.6 – 95.2%). Of the 965 examinations reported negative for PE, the AI software identified 961 as negative (specificity 99.6%; 95% CI 98.9–99.9%). Accuracy was 97.3%; 95% CI 96.4 - 98.1%. After reinterpretation, 30 of the original radiologist’s reports were identified to be inaccurate;10 were false positive and 20 were false negative. Accuracy for the mild, moderate, and severe TSS groups was 98.4%, 96.7%, and 97.2%, respectively (p value = 0.4). The median TSS for positive and negative PE study groups was 15 interquartile range (IQR) 9 - 22, and 17 (IQR 10 - 24), respectively. AI detected 92.5% of central PEs, 90.9% of the segmental PEs, and 96.3% of the subsegmental PEs (p value= 0.097). Accuracy of AI in the CTAs performed was 97.8%; CI 95% 96.8 - 98.5% compared to 90.9%; CI 95% 83.4 - 95.8% in the CECT group (p value < 0.001).

Regardless of the degree of airspace disease from COVID-19, AI solutions remain to be an invaluable tool for radiologists in evaluating CECCTs for PE. However, AI was found to be more accurate when CTPA, rather than CECT, was used.