ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


E2138. Detecting Pulmonary Embolism: Artificial Intelligence Versus Human Eye
Authors
  1. Ece Akduman; St. Louis University Hospital
  2. Kapil Chaudhary; St. Louis University Hospital
Objective:
Pulmonary embolism (PE) is a serious disease and can have grave consequences if missed. Artificial intelligence (AI) deep learning algorithms have shown a high degree of diagnostic accuracy for the detection of PE on CTPA. There are significant limiting factors for PE studies, resulting in suboptimal interpretation by the radiologists, yet CTPA still is the best method to diagnose PE. In our study we wanted to evaluate the role of AI in suboptimal CTPAs studies and see whether AI can improve detection of PE in the compromised studies, which may be secondary to motion, metal in the body resulting in streak artifact, suboptimal phase of contrast, or obesity. The objective of this study is to compare AI deep learning algorithm with radiologist interpretation in identifying PE in suboptimal studies.

Materials and Methods:
We identified 156 CTPAs for PE conducted at our institution from July 2020–June 2021 that were reported as suboptimal. Here we are presenting a small sample from our ongoing larger analysis on these studies. We randomly selected CTAPs (n = 16) that were interpreted as negative by a radiologist and were detected as positive on retrospective AI analysis. An experienced radiologist re-analyzed these scans to ascertain whether a study was false negative (FN: interpreted negative initially by radiologist) or if AI misinterpreted the scan as positive due to artifact (true negative, TN). We also looked for an attributable cause of the discrepancy if there were any detrimental clinical outcomes of missed cases.

Results:
On retrospective analysis, radiologist had agreement with AI on 10/16 cases which were found to have PEs and were missed by radiologist initially (FN). On 5/16 cases, no agreement was made if the filling defect was real or artifactual and remained TN. One case out of 16 was non-diagnostic for PE due to extensive streak artifacts from hardware and found to have massive bilateral PE on subsequent CTPA 10 days later. No mortality related to PE was present in these 16 cases.

Conclusion:
AI can efficiently work as a safety net to identify potentially missed Pes, specifically on suboptimal scans. Results from such analysis will advise peer-review and quality control, potentially improving patient outcomes. We will apply these observations on our larger ongoing analysis of suboptimal CTPAs (n =156) conducted at our institution in 2020–2021.