Incidental pulmonary embolism (iPE) represents a potentially life-threatening condition frequently detected in contrast-enhanced CT (CECT) scans conducted for unrelated clinical reasons. Early diagnosis is critical, especially given the disparity in mortality rates between untreated and promptly managed cases. Conventional CT pulmonary angiography (CTPA) remains the gold standard for PE diagnosis, but growing use of non-PE-targeted CT scans has underscored the need for supplementary detection tools. A deep learning-based solution, CINA-iPE, was developed to triage suspected iPEs in routine CECTs. A multicentre study sought to assess its diagnostic performance, evaluating accuracy, time-to-notification and robustness across diverse imaging environments.
Study Design and Data Collection
To assess CINA-iPE’s standalone performance, researchers conducted a retrospective analysis across five clinical centres, involving 381 anonymised CECT scans. These scans, originally not intended for PE evaluation, adhered to specific inclusion criteria, such as patient age above 18 and axial acquisition of contrast-enhanced images including part of the lungs. Cases were excluded if they presented severe artefacts, poor opacification or limited lung field coverage.
Three US board-certified radiologists independently reviewed each case to determine the presence of iPE, focusing only on occlusions in the main, lobar, interlobar or segmental arteries. Discrepancies between readers were resolved by majority consensus. Notably, subsegmental and chronic PEs were not considered within the target detection range of the software. This approach enabled the creation of a balanced dataset with 181 positive and 200 negative cases for final validation.
Diagnostic Performance and Technical Evaluation
CINA-iPE correctly identified 159 out of 181 positive iPE cases, yielding a sensitivity of 87.8%. It also classified 184 of 200 negative cases accurately, resulting in a specificity of 92.0%. The overall accuracy reached 90.0%, and the area under the receiver operating characteristic curve was calculated at 0.90. These results were consistent across imaging parameters, scanner types, patient age groups and sex, confirming the generalisability of the software.
Must Read: Detecting Pulmonary Embolism with Non-Contrast Imaging
False negative cases (22 in total) frequently involved image quality issues or complex findings, including severe artefacts, high noise or borderline subsegmental emboli. Conversely, the 16 false positive detections often arose from anatomical or pathological confounders, such as masses, artefacts or subsegmental clots excluded from the reference standard. Still, the system demonstrated strong localisation ability, with 89.5% of detected emboli correctly aligned with expert-defined anatomical locations.
The average time from image acquisition to result availability was 1.5 minutes, with a range of 0.3 to 2.7 minutes. This rapid turnaround supports its use in triaging studies for early iPE detection, enabling faster interpretation prioritisation in clinical settings.
Clinical Implications and Limitations
The software’s ability to flag incidental emboli shortly after scan acquisition offers the potential to reduce missed diagnoses and expedite treatment. Unlike other deep learning models trained and validated on dedicated CTPA studies, CINA-iPE was tested across diverse clinical sites, scanner types and imaging protocols. This provides stronger evidence for its robustness in real-world scenarios where PE is not the primary concern.
Nevertheless, certain limitations remain. The study was retrospective and included only cases with sufficient image quality, which may bias results towards higher performance. Radiologist assessments were used as the reference standard despite variability in inter-rater agreement. Additionally, no comparison was made between software output and unaided radiologist performance, and the study did not assess changes in clinical workflow or diagnostic turnaround.
Subsegmental PEs and chronic thromboembolic findings were not targeted by the software, due to their less urgent clinical relevance and the difficulty in achieving consensus diagnosis. Although CINA-iPE may assist in reducing oversight of critical incidental findings, over-reliance without clinician validation risks alert fatigue from false positives.
The multicentre evaluation of CINA-iPE demonstrated strong standalone diagnostic performance in detecting incidental pulmonary embolism on contrast-enhanced CT scans not intended for PE assessment. High sensitivity, specificity and speed of detection indicate that the software can effectively support radiologists in prioritising review of unsuspected embolic events. Its consistent performance across varied acquisition protocols and patient demographics reinforces its potential as a practical triage tool in diverse clinical environments. Further prospective studies are warranted to assess its impact on diagnostic workflows and patient outcomes.
Source: Radiology Advances
Image Credit: iStock
