P16 Assessing the use of an AI tool for grading coronary artery calcification on non-gated, non-cardiac chest CT. (21st September 2022)

• Record Type: Journal Article
• Authors: Shears, Ben; Meters, Rhys; Graby, John; Burnett, Tim A; Charters, Pia FP; Mody, Yashesh; Khavandi, Ali; Rodrigues, Jonathan CL
• Is Part Of: Heart. Volume 108(2022)Supplement 2
• Page Start: A8
• Page End: A8
• Publication Date: 2022-09-21
• Subjects: Heart -- Diseases -- Treatment -- Periodicals; Cardiology -- Periodicals; 616.12
• DOI: 10.1136/heartjnl-2022-BSCI.21
• Languages: English
• ISSNs: 1355-6037

Abstract

Abstract : Objective: BSCI guidelines now recommend reporting coronary artery calcification (CAC) on all CTCHEST regardless of indication. This study assessed the performance of an artificial intelligence (AI) tool for the assessment of CAC presence and severity grading on routine CTCHEST. Methods: Single centre retrospective analysis of 321 non-contrast CTCHEST across all age groups aged ≥18 years-old (January-December 2015). Scans were excluded if incomplete cardiac imaging (5), slice thickness prevented AI analysis (4) or evidence of prior coronary intervention (13). Ground truth against which AI was compared was set as experienced Radiologist re-review and grading. AI processing performance and quality assurance (QA) evaluation by Radiologist were assessed. Inter-observer variability between AI and Radiologist was measured against pre-defined levels of agreement using Cohen's κ. Results: 299 were included (mean age 64±19 years, 51% female). An AI report was produced in 97% (290/299). Of these, 78% (233/299) passed Radiologist quality assurance assessment. Reasons for failure included scoring a non-vessel (23, e.g. rib/sternum, valve, central line), a non-coronary vessel (12, e.g. pulmonary vein, aorta) or artefact (22). CAC prevalence (any degree) was 7% (17/233) on initial clinical report, versus 58% (134/233) on Radiologist re-review and 57% (133/233) by AI analysis. Inter-observer agreement for CAC presence and on QA-passed imaging was excellent (κ=0.92, p<0.001), and for severity grading was moderate (κ =0.60, p<0.001). Conclusion: AI analysis of non-contrast, non-gated CTCHEST is feasible, and may be more reliably performed than current standard practice. Reasons for failure are scan related and patient related.