Impact of fully automated assessment on interstudy reproducibility of biventricular volumes and function in cardiac magnetic resonance imaging

Cardiovascular magnetic resonance (CMR) imaging provides reliable assessments of biventricular morphology and function. Since manual post-processing is time-consuming and prone to observer variability, efforts have been directed towards novel artificial intelligence-based fully automated analyses. Hence, we sought to investigate the impact of artificial intelligence-based fully automated assessments on the inter-study variability of biventricular volumes and function. Eighteen participants (11 with normal, 3 with heart failure and preserved and 4 with reduced ejection fraction (EF)) underwent serial CMR imaging at in median 63 days (range 49-87) interval. Short axis cine stacks were acquired for the evaluation of left ventricular (LV) mass, LV and right ventricular (RV) end-diastolic, end-systolic and stroke volumes as well as EF. Assessments were performed manually (QMass, Medis Medical Imaging Systems, Leiden, Netherlands) by an experienced (3 years) and inexperienced reader (no active reporting, 45 min of training with five cases from the SCMR consensus data) as well as fully automated (suiteHEART, Neosoft, Pewaukee, WI, USA) without any manual corrections. Inter-study reproducibility was overall excellent with respect to LV volumetric indices, best for the experienced observer (intraclass correlation coefficient (ICC) > 0.98, coefficient of variation (CoV, < 9.6%) closely followed by automated analyses (ICC > 0.93, CoV < 12.4%) and lowest for the inexperienced observer (ICC > 0.86, CoV < 18.8%). Inter-study reproducibility of RV volumes was excellent for the experienced observer (ICC > 0.88, CoV < 10.7%) but considerably lower for automated and inexperienced manual analyses (ICC > 0.69 and > 0.46, CoV < 22.8% and < 28.7% respectively). In this cohort, fully automated analyses allowed reliable serial investigations of LV volumes with comparable inter-study reproducibility to manual analyses performed by an experienced CMR observer. In contrast, RV automated quantification with current algorithms still relied on manual post-processing for reliability.