Reliability of Machine Learning in functional assessment in cardiac magnetic resonance imaging

Abstract Purpose Our study aims at evaluating Machine Learning’s reliability to calculate ventricular volumes and functions in cardiac magnetic resonance imaging (CMR). Material and methods Eighty-five patients underwent CMR: forty healthy subjects, fifteen affected by myocarditis, seventeen dilated cardiomyopathy patients (DCM), and thirteen hypertrophic cardiomyopathy patients (HCM). Left ventricle (LV) and right ventricle (RV) analyses were performed both manually by operators and using Cvi42 software, which automates the identification of telesystolic and telediastolic phases. Statistical analysis assessed the concordance between measurements obtained manually and through the software, considering the following parameters: end-diastolic volume (EDV-BSA), end-systolic volume (ESV-BSA), stroke volume (SV), ejection fraction (EF), and detection of telesystolic and telediastolic phases. Results Intraclass correlation coefficient (ICC) analysis for LV volumes showed high concordance between manual and automatic measurements (ESV-BSA 0.97, EDV-BSA 0.98, SV 0.87, EF 0.93). ICC analysis for RV volumes presented high concordance between ESV-BSA and EDV-BSA measurements as well (ICC 0.90 and 0.91, respectively), whereas SV and EF measurements showed lower values (0.65 and 0.67, respectively). Statistical analysis also exhibited good concordance between manual and automated methods in the detection of telesystolic and telediastolic phases (ICC 0.80 and 0.84, respectively). Conclusion The LV and RV analyses conducted using the automated tool provide non-inferior performance to manual analyses, in particular for LV volumes.