Conexiant’s news site is now MDSpire News. Learn more
Advertisement
AI may aid segmental pulmonary embolism detection
Incidental pulmonary embolism detection increased following AI deployment on routine contrast-enhanced CT, with the clearest increase occurring in segmental emboli.
Assess whether deployment of a commercial AI tool was associated with changes in incidental pulmonary embolism (iPE) reporting on routine contrast-enhanced CT.
Approach:
Design: Single-center retrospective cross-sectional comparison of two 4-month periods before and after AI implementation.
Population and imaging: Included 11,690 contrast-enhanced CT examinations in patients aged 16 years or older; the primary analysis covered non-CT pulmonary angiography examinations with complete or partial chest coverage.
Assessment: Researchers searched radiology reports for pulmonary embolism, manually confirmed cases, and assessed reporting rates, embolus location, and AI performance.
Key Findings:
Among eligible non-CT pulmonary angiography examinations, iPE reporting increased from 0.35% (17/4,888) before implementation to 0.63% (35/5,523) afterward.
Segmental iPE reporting increased from 0.08% (4/4,888) to 0.42% (23/5,523); changes in main, lobar, and subsegmental iPE were not significant.
AI analyzed 4,344 of 5,523 eligible postimplementation examinations. Sensitivity was 65.6% among processed examinations and 60% across all eligible examinations.
AI sensitivity varied by contrast phase: 77.8% for late arterial, 71.4% for early arterial, and 56.2% for venous-phase examinations.
Immediate management changed in 77% (40/52) of patients with iPE, primarily through starting or intensifying anticoagulation or placing an inferior vena cava filter.
Interpretation:
The authors reported increased iPE detection after AI deployment, particularly for segmental emboli. They could not determine how much of the increase was directly attributable to AI versus increased radiologist attention, and the clinical significance of additional predominantly peripheral emboli remains uncertain.
Limitations:
Retrospective, single-center design may limit generalizability.
Report keyword searches may have missed cases; the original radiology report was the reference standard.
Some examinations were not routed to the AI algorithm, and detailed comorbidities and venous thromboembolism risk factors were not systematically collected.
Contrast-phase and chest-coverage performance estimates were based on small numbers of iPE cases.
Conclusion:
The investigators called for prospective multicenter studies to assess the prognostic implications and appropriate management of additional iPE detected after AI implementation. The authors reported no study funding.
