Deep learning-based algorithm for automatic detection of incidental pulmonary embolism on contrast-enhanced CT: a...

Incidenal pulmonary embolism (iPE) is increasingly detected on contrast-enhanced computed tomography (CECT) performed for non-PE indications, reflecting the growing volume and complexity of cross-sectional imaging. These findings, although unexpected, carry important clinical implications and may be underreported due to the primary diagnostic focus of the study. Artificial intelligence (AI) applications offer the potential to augment radiologist workflow bt training exams and highlighting cases suspicious for iPE, thereby improving detection accuracy and timeliness in routine clinical practice.

Likelihood of incidental pulmonary embolism (iPE) increases with increased body computed tomography CT) imaging. This study evaluates the diagnostic performance and effectiveness of triage of a standalone AI solution for detecting iPE in contrast-enhanced CT (CECT) exams obtained for non-PE clinical indications.

A commercially available deep learning-based software, CINA-iPE (Avicenna.AI, La Ciotat, France), analyzes CECT images to highlight suspected incidental PE cases. Consecutive retrospective CECTs from 5 clinical centers, not performed for PE evaluation, were collected until a selected balanced dataset between positive and negative cases was obtained. The reference standard was established by three independent U.S. board-certified radiologists reviewing the same images. Diagnostic performance and the time-to-notification (from data acquisition to processing of results) were computed.

A total of 381 anonymized CECT cases were acquired on 39 different scanner models from GE, Philips, Siemens, and Canon. The algorithm correctly identified 159/181 exams positive for PE (sensitivity 87.8% [95% CI: 82.2%-92.2%]) and 184/200 exams negative for PE (specificity 92.0% [95% CI: 87.3%-95.4%]), yielding an accuracy of 90.0% [95% CI: 86.6%-92.8%]. Of 16 detected false positive cases, 50% were complex CECTs subject to disagreement among the reference read radiologists. The device missed 22 pulmonary embolisms, with 45.5% of them being complex cases and subject to disagreement among reviewers. The time from data acquisition to processing results was 1.5 ± 0.5 (mean ± SD, 95% CI: 1.4%-1.5%) minutes.

The CINA-iPE application accurately identified incidental PE in studies not performed specifically for evaluation of PE with high sensitivity and specificity. Automatically processed results were available to interpreting physicians within minutes, which could be used to prioritize interpretation of studies. This may be useful for increasing the accuracy or speed of detection of iPE.