Incidental pulmonary embolism in abdominal CT: detection rate and characteristics with artificial intelligence

Abstract

Abstract Background Abdominal CT is a mainstay in the evaluation of abdominal infections, trauma, oncology, and postoperative complications. Pulmonary embolism is a common complication, but there is a risk that these ancillary findings are overlooked. In addition, data on detection rate and characteristics of incidental pulmonary embolism (iPE) on abdominal CT are lacking. Purpose The current study compared the period before and after implementing an artificial intelligence (AI) algorithm for iPE detection regarding detection rate and characteristics. Material and Methods A retrospective cross-sectional study was performed on abdominal CTs between August 1, 2019, and January 31, 2021 (before AI implementation, 8026 studies) and August 1, 2021, and January 31, 2023 (after AI implementation, 8765 studies). iPE cases were identified through text search and manually confirmed. Study indication and urgency were recorded for iPE patients, and the most proximal iPE level was assessed. A total of 1000 cases after AI implementation were randomly selected and manually reviewed for AI accuracy analysis. Results A total of 5876 patients with a mean age of 63.6 ± 17.7 years were included before AI implementation, and 6310 patients with a mean age of 63.2 ± 18.3 years after AI implementation. The iPE detection rate was higher after AI implementation, 0.57% (50/8765 studies) vs 0.12% (10/8026), P < .001. The most common study indications were abdominal pain (25%, 15/60 cases) and infection (30%, 18/60 cases). There were no differences in CT pulmonary angiography usage or the most proximal extent of the iPE between the periods before or after AI implementation, P > .05. AI identified 46/50 of the reported iPE with 7 AI false-positive cases for a positive predictive value of 87% (95% confidence interval: 75-93%). In the manually reviewed randomly selected subset, iPE prevalence was 1.7% (15/874, 95% confidence interval: 1.0-2.8%) with AI having 40% sensitivity (95% CI, 16-68) and 100% specificity (95% CI, 99.5-100). Conclusion Implementing AI for iPE detection and triage increased the iPE detection rate in abdominal CT. The AI sensitivity was moderate, with very few AI false positives.