Rebecca A Jonas1, Emil Barkovich2, Andrew D Choi2, William F Griffin2, Joanna Riess2, Hugo Marques3, Hyuk-Jae Chang4, Jung Hyun Choi5, Joon-Hyung Doh6, Ae-Young Her7, Bon-Kwon Koo8, Chang-Wook Nam9, Hyung-Bok Park10, Sang-Hoon Shin11, Jason Cole12, Alessia Gimelli13, Muhammad Akram Khan14, Bin Lu15, Yang Gao15, Faisal Nabi16, Ryo Nakazato17, U Joseph Schoepf18, Roel S Driessen19, Michiel J Bom19, Randall C Thompson20, James J Jang21, Michael Ridner22, Chris Rowan23, Erick Avelar24, Philippe Généreux25, Paul Knaapen19, Guus A de Waard19, Gianluca Pontone26, Daniele Andreini26, Marco Guglielmo26, Mouaz H Al-Mallah16, Robert S Jennings27, Tami R Crabtree27, James P Earls28. 1. Department of Internal Medicine, Thomas Jefferson University Medical Center, Philadelphia, PA, USA. 2. Department of Radiology and Division of Cardiology, The George Washington University School of Medicine, Washington, DC, USA. 3. Nova Medical School - Faculdade de Ciências Médicas, Lisboa, Portugal. 4. Division of Cardiology, Severance Cardiovascular Hospital and Severance Biomedical Science Institute, Yonsei University College of Medicine, Yonsei University Health System, Seoul, South Korea. 5. Ontact Health, Inc., Seoul, South Korea. 6. Division of Cardiology, Inje University Ilsan Paik Hospital, South Korea. 7. Kang Won National University Hospital, Chuncheon, South Korea. 8. Department of Internal Medicine, Seoul National University Hospital, Seoul, South Korea. 9. Cardiovascular Center, Keimyung University Dongsan Hospital, Daegu, South Korea. 10. Division of Cardiology, Department of Internal Medicine, International St. Mary's Hospital, Catholic Kwandong University College of Medicine, Incheon, South Korea. 11. National Health Insurance Service Ilsan Hospital, Goyang, South Korea. 12. Mobile Cardiology Associates, Mobile, AL, USA. 13. Department of Imaging, Fondazione Toscana Gabriele Monasterio, Pisa, Italy. 14. Cardiac Center of Texas, McKinney, TX, USA. 15. State Key Laboratory of Cardiovascular Disease, Fuwai Hospital, Beijing, China. 16. Houston Methodist Hospital, Houston, TX, USA. 17. Cardiovascular Center, St. Luke's International Hospital, Tokyo, Japan. 18. Medical University of South Carolina, Charleston, SC, USA. 19. Amsterdam University Medical Center, VU University Medical Center, Amsterdam, the Netherlands. 20. St. Luke's Mid America Heart Institute, Kansas City, MO, USA. 21. Kaiser Permanente San Jose Medical Center, San Jose, CA, USA. 22. Heart Center Research, LLC, Huntsville, AL, USA. 23. Renown Heart and Vascular Institute, Reno, NV, USA. 24. Oconee Heart and Vascular Center at St Mary's Hospital, Athens, GA, USA. 25. Gagnon Cardiovascular Institute at Morristown Medical Center, Morristown, NJ, USA. 26. Centro Cardiologico Monzino, IRCCS, Milan, Italy. 27. Cleerly Inc, New York, NY, USA. 28. Department of Radiology and Division of Cardiology, The George Washington University School of Medicine, Washington, DC, USA; Cleerly Inc, New York, NY, USA. Electronic address: jearls@mfa.gwu.edu.
Abstract
OBJECTIVES: To determine whether coronary computed tomography angiography (CCTA) scanning, scan preparation, contrast, and patient based parameters influence the diagnostic performance of an artificial intelligence (AI) based analysis software for identifying coronary lesions with ≥50% stenosis. BACKGROUND: CCTA is a noninvasive imaging modality that provides diagnostic and prognostic benefit to patients with coronary artery disease (CAD). The use of AI enabled quantitative CCTA (AI-QCT) analysis software enhances our diagnostic and prognostic ability, however, it is currently unclear whether software performance is influenced by CCTA scanning parameters. METHODS: CCTA and quantitative coronary CT (QCT) data from 303 stable patients (64 ± 10 years, 71% male) from the derivation arm of the CREDENCE Trial were retrospectively analyzed using an FDA-cleared cloud-based software that performs AI-enabled coronary segmentation, lumen and vessel wall determination, plaque quantification and characterization, and stenosis determination. The algorithm's diagnostic performance measures (sensitivity, specificity, and accuracy) for detecting coronary lesions of ≥50% stenosis were determined based on concordance with QCA measurements and subsequently compared across scanning parameters (including scanner vendor, model, single vs dual source, tube voltage, dose length product, gating technique, timing method), scan preparation technique (use of beta blocker, use and dose of nitroglycerin), contrast administration parameters (contrast type, infusion rate, iodine concentration, contrast volume) and patient parameters (heart rate and BMI). RESULTS: Within the patient cohort, 13% demonstrated ≥50% stenosis in 3 vessel territories, 21% in 2 vessel territories, 35% in 1 vessel territory while 32% had <50% stenosis in all vessel territories evaluated by QCA. Average AI analysis time was 10.3 ± 2.7 min. On a per vessel basis, there were significant differences only in sensitivity for ≥50% stenosis based on contrast type (iso-osmolar 70.0% vs non isoosmolar 92.1% p = 0.0345) and iodine concentration (<350 mg/ml 70.0%, 350-369 mg/ml 90.0%, 370-400 mg/ml 90.0%, >400 mg/ml 95.2%; p = 0.0287) in the context of low injection flow rates. On a per patient basis there were no significant differences in AI diagnostic performance measures across all measured scanner, scan technique, patient preparation, contrast, and individual patient parameters. CONCLUSION: The diagnostic performance of AI-QCT analysis software for detecting moderate to high grade stenosis are unaffected by commonly used CCTA scanning parameters and across a range of common scanning, scanner, contrast and patient variables. CONDENSED ABSTRACT: An AI-enabled quantitative CCTA (AI-QCT) analysis software has been validated as an effective tool for the identification, quantification and characterization of coronary plaque and stenosis through comparison to blinded expert readers and quantitative coronary angiography. However, it is unclear whether CCTA screening parameters related to scanner parameters, scan technique, contrast volume and rate, radiation dose, or a patient's BMI or heart rate at time of scan affect the software's diagnostic measures for detection of moderate to high grade stenosis. AI performance measures were unaffected across a broad range of commonly encountered scanner, patient preparation, scan technique, intravenous contrast and patient parameters.
OBJECTIVES: To determine whether coronary computed tomography angiography (CCTA) scanning, scan preparation, contrast, and patient based parameters influence the diagnostic performance of an artificial intelligence (AI) based analysis software for identifying coronary lesions with ≥50% stenosis. BACKGROUND: CCTA is a noninvasive imaging modality that provides diagnostic and prognostic benefit to patients with coronary artery disease (CAD). The use of AI enabled quantitative CCTA (AI-QCT) analysis software enhances our diagnostic and prognostic ability, however, it is currently unclear whether software performance is influenced by CCTA scanning parameters. METHODS: CCTA and quantitative coronary CT (QCT) data from 303 stable patients (64 ± 10 years, 71% male) from the derivation arm of the CREDENCE Trial were retrospectively analyzed using an FDA-cleared cloud-based software that performs AI-enabled coronary segmentation, lumen and vessel wall determination, plaque quantification and characterization, and stenosis determination. The algorithm's diagnostic performance measures (sensitivity, specificity, and accuracy) for detecting coronary lesions of ≥50% stenosis were determined based on concordance with QCA measurements and subsequently compared across scanning parameters (including scanner vendor, model, single vs dual source, tube voltage, dose length product, gating technique, timing method), scan preparation technique (use of beta blocker, use and dose of nitroglycerin), contrast administration parameters (contrast type, infusion rate, iodine concentration, contrast volume) and patient parameters (heart rate and BMI). RESULTS: Within the patient cohort, 13% demonstrated ≥50% stenosis in 3 vessel territories, 21% in 2 vessel territories, 35% in 1 vessel territory while 32% had <50% stenosis in all vessel territories evaluated by QCA. Average AI analysis time was 10.3 ± 2.7 min. On a per vessel basis, there were significant differences only in sensitivity for ≥50% stenosis based on contrast type (iso-osmolar 70.0% vs non isoosmolar 92.1% p = 0.0345) and iodine concentration (<350 mg/ml 70.0%, 350-369 mg/ml 90.0%, 370-400 mg/ml 90.0%, >400 mg/ml 95.2%; p = 0.0287) in the context of low injection flow rates. On a per patient basis there were no significant differences in AI diagnostic performance measures across all measured scanner, scan technique, patient preparation, contrast, and individual patient parameters. CONCLUSION: The diagnostic performance of AI-QCT analysis software for detecting moderate to high grade stenosis are unaffected by commonly used CCTA scanning parameters and across a range of common scanning, scanner, contrast and patient variables. CONDENSED ABSTRACT: An AI-enabled quantitative CCTA (AI-QCT) analysis software has been validated as an effective tool for the identification, quantification and characterization of coronary plaque and stenosis through comparison to blinded expert readers and quantitative coronary angiography. However, it is unclear whether CCTA screening parameters related to scanner parameters, scan technique, contrast volume and rate, radiation dose, or a patient's BMI or heart rate at time of scan affect the software's diagnostic measures for detection of moderate to high grade stenosis. AI performance measures were unaffected across a broad range of commonly encountered scanner, patient preparation, scan technique, intravenous contrast and patient parameters.