Incidental Pulmonary Embolism Detection on Routine Contrast-Enhanced CT before and after Deployment of a Commercial AI Tool: A Real-World Evaluation.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Mater Misericordiae University Hospital, Eccles St, Dublin 7, Dublin, Ireland.
- Department of Haematology, Mater Misericordiae University Hospital, University College Dublin, Belfied, Dublin 4, Dublin, Ireland.
- School of Medicine, University College Dublin, University College Dublin, Belfied, Dublin 4, Ireland.
- School of Medicine, University College, Mater Misericordiae University Hospital, Eccles St, Dublin 7, Dublin, Ireland.
Abstract
To assess the impact of deploying AI software for PE detection on reporting rates of incidental pulmonary embolism (iPE) on routine contrast‑enhanced CT. Consecutive adult patients undergoing contrast-enhanced CT in a large tertiary care hospital were included. Two time periods were compared; September 1st to December 31st 2022 (before the introduction of dedicated AI tools for PE detection) and September 1st to December 31st 2023 (following introduction of AI). Cases positive for PE were identified retrospectively. Data on baseline demographics, imaging variables and immediate clinical management were gathered from the electronic health record. 11,690 contrast-enhanced CT studies were analysed, 10,411 of which were eligible non-CTPA studies. iPE detection rate increased from 0.35% to 0.63% (p = 0·0389) following AI implementation. The segmental iPE detection rate increased significantly (0.08% vs. 0.42%, p < 0.001). In the post‑AI period, the iPE algorithm did not analyse 21.3% of non‑CTPA studies. AI sensitivity for iPE detection was 65.6% among analysed studies and 60.0% when all eligible examinations were included. AI sensitivity was 71.4%, 77.8% and 56.2% for early arterial, late arterial and venous phase scans respectively. The introduction of a dedicated AI tool for iPE detection significantly increased iPE detection, particularly segmental iPE. This occurred despite incomplete AI analysis of eligible scans and a modest AI algorithm sensitivity. The proportion of eligible examinations processed by AI and algorithm sensitivity varied by CT protocol and contrast phase, highlighting the importance of evaluating and monitoring AI tools under real-world clinical conditions following implementation. Deployment of a commercial PE AI detection system in routine practice was associated with a significant increase in reported iPE on non‑CTPA CT, predominantly at the segmental level. However, the AI analysed only 78.7% of eligible non-CTPA examinations, highlighting incomplete examination coverage as an important real-world implementation limitation. AI coverage and performance also varied across CT protocols and contrast phases, underscoring the importance of protocol-specific evaluation following clinical deployment.