Machine Learning the Right Heart: Leveraging CMR for RV Volume Quantification From 3D Echocardiography

Abstract

Right ventricular (RV) volume indices are important prognostic markers in several cardiac diseases. While cardiac magnetic resonance (CMR) imaging remains the gold-standard for volume quantification, echocardiography is more accessible. Unlike two-dimensional echocardiography (2DE), three-dimensional echocardiography (3DE) enables volume quantification without geometric assumptions. However, manual RV segmentation from 3DE is slow and hindered by speckle noise and poor spatial resolution. Machine learning can be utilised to overcome these challenges for efficient and accurate 3DE RV assessment. Thirty participants (18 with cardiac disease, 12 healthy controls) were imaged with transthoracic 3DE and cine CMR <1 hour apart. The RV segmentations from CMR were spatiotemporally aligned to the corresponding 3DE for each participant across one cardiac cycle. Paired 3DE images and CMR labels were used to train a deep-learning model (80/20 training/testing split) that predicts segmentations from 3DE only. Resulting end-diastolic volume (EDV), end-systolic volume (ESV), and ejection fraction (EF) were compared with CMR for each case. Machine learning segmentations predicted lower EDV (–17±21 mL), ESV (–8±24 mL), and EF (–3±9%) compared with CMR, but there were no statistically significant differences (p>0.05, n=6). These biases were smaller than those previously found for manual 3DE RV segmentations (–33±25 mL, –30±20 mL, and 6±10%). Machine learning enabled accurate RV volume quantification from 3DE, which was not possible with 2DE due to complex RV geometry. Automated segmentations of 3DE produced comparable volumes to gold-standard CMR, which could facilitate the use of 3DE for clinical RV assessment.