AI Boosts Incidental Pulmonary Embolism Detection on CT

ARTIFICIAL intelligence (AI) could significantly improve the detection of incidental pulmonary embolism (iPE) on routine contrast-enhanced CT scans, according to new real-world research.

A new study found that introducing a commercial AI detection tool was associated with an increase in reported iPE, particularly smaller emboli at the segmental level. However, researchers also identified important limitations in the proportion of eligible scans analysed by the technology.

AI in Routine CT Imaging

Pulmonary embolism can be detected incidentally when patients undergo contrast-enhanced CT for reasons unrelated to suspected embolism. Unlike dedicated CT pulmonary angiography (CTPA), these examinations are not optimised specifically for identifying pulmonary emboli, potentially making subtle cases more difficult to detect.

Researchers evaluated 11,690 contrast-enhanced CT examinations performed at a large tertiary hospital. They compared scans conducted between September and December 2022, before implementation of dedicated AI software, with those conducted during the same months in 2023, following its introduction.

Of the examinations analysed, 10,411 were eligible non-CTPA studies.

Detection Rates Rise After AI Introduction

Following AI implementation, the reported iPE detection rate increased significantly, from 0.35% before deployment to 0.63% afterwards.

The difference was particularly marked for segmental pulmonary emboli. Detection at this level increased from 0.08% before implementation to 0.42% following the introduction of AI.

Researchers said the findings suggest that AI could help radiologists identify incidental emboli that might otherwise be overlooked during routine imaging.

However, the results also revealed limitations in real-world AI performance. The algorithm failed to analyse 21.3% of eligible non-CTPA examinations during the post-implementation period.

Among examinations that were successfully processed, AI sensitivity for iPE was 65.6%. When all eligible examinations were considered, including those not analysed by the software, sensitivity fell to 60.0%.

Performance also varied according to contrast phase. Sensitivity reached 71.4% for early arterial scans and 77.8% for late arterial scans, compared with 56.2% for venous-phase examinations.

Real-World Monitoring Remains Essential

The researchers concluded that deploying dedicated AI software was associated with a meaningful increase in iPE reporting, predominantly through improved detection of segmental emboli.

Nevertheless, incomplete scan coverage and differences in performance between CT protocols highlight the challenges of translating AI performance into everyday clinical practice.

The findings emphasise the importance of evaluating AI tools after clinical deployment rather than relying solely on pre-implementation performance data. Ongoing, protocol-specific monitoring could help healthcare teams understand where AI provides the greatest benefit and where additional radiologist oversight remains particularly important.

Reference

O’Herlihy F et al. Incidental Pulmonary Embolism Detection on Routine Contrast-Enhanced CT before and after Deployment of a Commercial AI Tool: A Real-World Evaluation. Br J Radiol. 2026.

Featured image: Markoff on AdobeStock