Adjunctive AI Leads to 16 Percent Increase in CT Sensitivity for Incidental Pulmonary Embolism

Can artificial intelligence (AI) enhance the detection of incidental pulmonary embolism (IPE) on chest and abdominal contrast-enhanced computed tomography (CECT) scans?

In a new prospective study, recently published in the American Journal of Roentgenology, researchers reviewed radiologist interpretation of 1,467 chest and abdominal CECT images drawn from 1,434 patients (mean age of 53.8) for IPE in the first phase. In the second phase of the study, radiologist had access to adjunctive AI (BriefCase for iPE Triage, versions 8.0 and 8.1, Aidoc) in interpreting 3,182 CECT exams in 2,886 patients (mean age of 55.4), according to the study.

While the study authors saw no significant difference with specificity (99.1 percent for radiologists alone vs. 99.9 percent for adjunctive AI), they noted a 16.2 percent higher increase in sensitivity for IPE with adjunctive AI (96.2 percent vs. 80 percent).

“ … A radiology department’s implementation of AI assistance for routine CECT examinations of the chest or abdomen resulted in increased sensitivity for IPE detection in comparison with interpretation by radiologists alone although (there was no) significant difference in other diagnostic performance metrics,” noted lead study author Cody H. Savage, M.D., a radiology resident affiliated with the University of Maryland Medical Intelligent Imaging Center within the Department of Diagnostic Radiology and Nuclear Medicine at the University of Maryland School of Medicine, and colleagues.

The researchers noted that adjunctive AI demonstrated higher accuracy in cases of clinically significant IPE (92.3 percent vs. 75 percent) and clinically uncertain IPE (100 percent vs. 83.3 percent), but they acknowledged these differences in small numbers of patients lacked statistical significance. There were also no statically significant differences with respect to overall positive predictive value (PPV) (92.6 with AI vs. 94.1 without AI) and negative predictive value (NPV) (99.9 percent with AI vs. 99.7 percent without AI), according to the study authors.

While there was no statistically significant difference in the mean report turnaround time for IPE-positive cases (64.6 minutes with AI vs. 78.3 minutes without AI), the researchers noted significant differences in the emergency department (ED) setting with mean report turnaround time (48.4 minutes with AI vs. 73.6 minutes without AI) and mean interpretation time (34 minutes with AI vs. 59.2 minutes without AI).

The study authors also pointed out that emergency radiologists assessed 42.6 percent of all CECT exams (1,979/4,649) reviewed in the study.

“We posit that the AI system caused these time reductions by prompting ER radiologists to more quickly review resident interpretations for these cases,” added Savage and colleagues.

(Editor’s note: For related content, see “FDA Clears CT-Based AI Tools for PE Detection and Stroke Assessment,” “Study Shows Mixed Results with AI Software for PE Detection on CTPA Scans” and “Expediting the Management of Incidental Pulmonary Emboli on CT.”)

In regard to study limitations, the authors acknowledged that the cohort was entirely drawn from one academic medical center, which may limit broader extrapolation of the results. They noted that staffing changes and fluctuating patient loads may have affected the results from the two phases of the study. The researchers also acknowledged the subsequent radiology reviews were limited to positive results for IPE in CT interpretation via the original radiology report or AI assessment.