Automated Segmentation of the Right Ventricle from 3D Echocardiography Using Labels from Cardiac Magnetic Resonance Imaging

Segmentation of the right ventricle (RV) from 3D echocardiography (3DE) is a challenging task. In comparison to the left ventricle (LV), the complex geometry of the RV hinders accurate and reproducible volume quantification. While more accessible, 3DE falls short of gold-standard cardiac magnetic resonance (CMR) imaging for volume quantification due to its low spatial resolution and poor contrast-to-noise ratio. The use of machine learning can overcome these challenges to improve 3DE RV segmentation. This study assessed this approach by leveraging segmentations derived from CMR as ground truth labels, and including LV labels as contextual information. Forty subjects (20 females; 20 with cardiac diseases of mixed origin; 20 healthy controls) were imaged with transthoracic 3DE and cine CMR <1 h apart. Biventricular segmentations from CMR were spatially registered to corresponding end-diastolic and end-systolic 3DE images. Paired 3DE images and CMR labels from 32 subjects were used to train deep-learning models for RV segmentation from 3DE. One model was trained with RV labels only, and a second was trained with both RV and LV labels. Using the 8 test cases, the model trained with biventricular labels predicted an end-diastolic volume of 158 ± 36 ml, end-systolic volume of 105 ± 40 ml, and ejection fraction of 36 ± 11 %, which were not statistically significantly different to values measured using CMR (165 ± 30 ml, 115 ± 32 ml and 31 ± 8 %, respectively; P=NS). Inclusion of LV labels improved segmentation accuracy in cases with RV free wall signal dropout. These results indicate that leveraging CMR-derived labels for deep-learning can facilitate reliable clinical assessment of RV function from 3DE.