Performance of artificial intelligence for biventricular cardiovascular magnetic resonance volumetric analysis in the clinical setting

• Published in: The international journal of cardiovascular imaging Vol. 38; no. 11; pp. 2413 - 2424
• Main Authors: Hatipoglu, Suzan, Mohiaddin, Raad H., Gatehouse, Peter, Alpendurada, Francisco, Baksi, A. John, Izgi, Cemil, Prasad, Sanjay K., Pennell, Dudley J., Krupickova, Sylvia
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.11.2022; Springer Nature B.V
• Subjects: Artificial Intelligence; Automation; Cardiac Imaging; Cardiology; Cardiomyopathy; Cardiovascular disease; Coefficient of variation; Contours; Decision making; Dilated cardiomyopathy; Heart diseases; Humans; Imaging; Ischemia; Magnetic resonance; Magnetic Resonance Spectroscopy; Medicine; Medicine & Public Health; Original Paper; Predictive Value of Tests; Radiology; Segmentation; Stroke Volume; Surveys; Ventricle; Ventricular Function, Left; Volumetric analysis; Ventricular volumes; Cardiovascular magnetic resonance; Ventricular function; Myocardial segmentation; Artificial intelligence; Myocardial disease; Ejection fraction
• ISSN: 1875-8312; 1569-5794; 1875-8312; 1573-0743
• DOI: 10.1007/s10554-022-02649-1

Cardiovascular magnetic resonance (CMR) derived ventricular volumes and function guide clinical decision-making for various cardiac pathologies. We aimed to evaluate the efficiency and clinical applicability of a commercially available artificial intelligence (AI) method for performing biventricular volumetric analysis. Three-hundred CMR studies (100 with normal CMR findings, 50 dilated cardiomyopathy, 50 hypertrophic cardiomyopathy, 50 ischaemic heart disease and 50 congenital or valvular heart disease) were randomly selected from database. Manual biventricular volumetric analysis (CMRtools) results were derived from clinical reports and automated volumetric analyses were performed using short axis volumetry AI function of CircleCVI 42 v5.12 software. For 20 studies, a combined method of manually adjusted AI contours was tested and all three methods were timed. Clinicians` confidence in AI method was assessed using an online survey. Although agreement was better for left ventricle than right ventricle, AI analysis results were comparable to manual method. Manual adjustment of AI contours further improved agreement: within subject coefficient of variation decreased from 5.0% to 4.5% for left ventricular ejection fraction (EF) and from 9.9% to 7.1% for right ventricular EF. Twenty manual analyses were performed in 250 min 12 s whereas same task took 5 min 48 s using AI method. Clinicians were open to adopt AI but concerns about accuracy and validity were raised. The AI method provides clinically valid outcomes and saves significant time. To address concerns raised by survey participants and overcome shortcomings of the automated myocardial segmentation, visual assessment of contours and performing manual corrections where necessary appears to be a practical approach.