Evaluation of artificial intelligence for pulmonary embolism detection on CTPA

Abstract

Pulmonary embolism (PE) is a life-threatening condition commonly diagnosed with computed tomography pulmonary angiography (CTPA). Although artificial intelligence (AI) has been applied for PE detection, its performance relative to physician-only and AI-assisted interpretation remains incompletely characterized. The study aimed to evaluate the diagnostic performance and reading time across 3 CTPA interpretation strategies: AI alone, physician-only reading, and AI-assisted physician reading. We retrospectively analyzed CTPA images from 50 patients diagnosed with PE at a single center and randomly divided them into 2 groups: one group (n = 25) was reviewed by radiologists from different levels, and the other group (n = 25) was reviewed by radiologists with AI assistance. Then, AI independently reviewed all images. The reference standard was established by consensus among 3 senior radiologists. Diagnostic performance and reading time were compared across the 3 reading paradigms. AI alone achieved an overall sensitivity of 91.24% and precision of 96.12%; sensitivity declined from 100% in grade 1 to 2 vessels to 86.32% in grade ≥6 vessels. Compared with physician-only reading, AI-assisted reading increased sensitivity and precision for junior radiologists (64.28%–75.51% and 81.82%–89.16%, respectively) and intermediate radiologists (80.67%–92.34% and 88.47%–97.31%, respectively). AI assistance also reduced mean reading time at both experience levels (both P < .05). In this small, single-center retrospective exploratory study, AI-assisted CTPA interpretation was associated with higher lesion-level sensitivity and precision and shorter reading times than physician-only interpretation. Larger prospective multicenter studies are required to validate these findings and assess their generalizability.