Evaluation of artificial intelligence for pulmonary embolism detection on CTPA: A single-center retrospective study comparing AI, radiologists, and AI-assisted interpretation.

AI alone achieved an overall sensitivity of 91.24% and precision of 96.12% for pulmonary embolism detection on CTPA.

• AI independently reviewed all 50 CTPA images in the study.
• The reference standard was established by consensus among 3 senior radiologists.
• AI overall sensitivity was 91.24% and precision was 96.12%.
• AI performance declined with more distal vessels: sensitivity was 100% in grade 1 to 2 vessels but declined to 86.32% in grade ≥6 vessels.

AI assistance improved sensitivity and precision for junior radiologists compared to physician-only reading.

• 25 patients were reviewed by radiologists without AI assistance and 25 with AI assistance.
• For junior radiologists, sensitivity increased from 64.28% (physician-only) to 75.51% (AI-assisted).
• For junior radiologists, precision increased from 81.82% (physician-only) to 89.16% (AI-assisted).
• Radiologists were categorized by experience level (junior, intermediate, and implicitly senior).

AI assistance improved sensitivity and precision for intermediate radiologists compared to physician-only reading.

• For intermediate radiologists, sensitivity increased from 80.67% (physician-only) to 92.34% (AI-assisted).
• For intermediate radiologists, precision increased from 88.47% (physician-only) to 97.31% (AI-assisted).
• Improvements were observed at the lesion level of analysis.

AI assistance reduced mean reading time for both junior and intermediate radiologists.

• Reading time reduction was statistically significant for both junior and intermediate radiologists (both P < .05).
• The reduction in reading time was observed across both experience levels when AI assistance was provided.
• Specific mean reading times were not reported in the abstract but statistical significance was confirmed.

The study was a small, single-center retrospective exploratory study with a limited sample size of 50 patients.

• 50 patients diagnosed with PE at a single center were retrospectively analyzed.
• Patients were randomly divided into two groups of 25 each: one reviewed by radiologists alone and one reviewed with AI assistance.
• Authors explicitly noted the study as 'small, single-center retrospective exploratory' and called for 'larger prospective multicenter studies' to validate findings.
• The study design limits generalizability of the findings.

AI sensitivity for pulmonary embolism detection declined as vessel grade increased, indicating reduced performance in more distal vasculature.

• Sensitivity was 100% for grade 1 to 2 vessels.
• Sensitivity declined to 86.32% for grade ≥6 vessels.
• This gradient suggests AI performance is better in larger, more proximal pulmonary vessels than in smaller, more distal ones.