Assessment of Artificial Intelligence Technology for Pulmonary Embolism Detection

Zehavit Kirshenboim Daniel Raskin Eli Konen Gal Yaniv Elio DiSegni Orly Goitein

Departments of Diagnostic Imaging, Sheba Medical Center, Tel-Hashomer, affiliated to the Sackler School of Medicine, Tel-Aviv University, Israel



Purpose: To compare artificial intelligence (AI) deep learning technology with radiology interpretation in identifying pulmonary embolism (PE) on CT of Pulmonary Angiography (CTPA) scans.

Method and Materials: Consecutive CTPA scans obtained for the clinical suspicion of PE between December 2018 and February 2019 were subjected to a dedicated AI algorithm (Aidoctm), specifically designed for the detection of PE on CTPA (AI-PE). AI-PE results were compared to a senior radiologists` interpretation, defined as the “gold standard” for PE detection. True positive, true negative, false positive and false negative were defined according to the “gold standard”. In cases of discrepancies a third reading by a senior radiologist was performed. The sensitivity, specificity, positive and negative predictive values and accuracy of AI PE for PE detection were calculated.

Results: The study cohort included 374 CTA scans performed for suspected PE; one scan was excluded due to severe motion artifacts.

True positive and true negative results were documented in 57/373 (15%) and 292/373 (78%) of studies, respectively. False positive and false negative cases were documented in 10 cases each.

Sensitivity, specificity, positive and negative predictive values and accuracy of AI-PE as compared with the “gold standard“ were, 85.07%, 96.7%, 85.07%, 96.7% and 94.6% `{` P<0.05`}` respectively.

Conclusion: AI-PE demonstrated high sensitivity, specificity and accuracy for PE detection as compared with the “gold standard“- the radiologists` interpretation. Thus, AI-PE might provide solutions for quicker diagnosis, lowering misdiagnosis and serve as a peer review platform.

Clinical Relevance: The dedicated AI-PE demonstrated high specificity and NPV when compared to the “gold standard“ for PE diagnosis. Such technology holds great promise in aiding the radiologist, especially in the face of the constant increasing workload. Integration of AI-PE into daily practice offers new solutions for quicker diagnosis, lowering misdiagnosis, prioritize reporting and peer reviewing.

KLAS Research Webinar AI in Clinical Practice
Recent Posts
Latest Tweet
See Demo Tailored for your PACS