A combined multi-scale deep learning and random forests approach for direct left ventricular volumes estimation in 3D echocardiography

Estimation of left ventricular (LV) volumes from 3D echocardiography (3DE) is a popular clinical approach in accurate assessment of left ventricular function for the diagnosis of cardiac disease. The segmentation of 3DE volumes is a crucial step in traditional methods. Nevertheless, segmentation itself is an extremely challenging problem due to the presence of speckle noise and discontinuous edges. Therefore, direct left ventricular volumes estimation methods without the segmentation become attractive in cardiac function analysis. The aim of this paper is to present a fully learning framework to estimate the left ventricular volume in 3DE. The proposed method combined unsupervised multi-scale convolutional deep network and random forests. The multi-scale convolution deep network adopted multi-scale convolutional filters to represent features of unlabeled end-diastolic and end-systolic 3DE volumes (EDV and ESV). And then we formulated left ventricular volume estimation as a regression problem and used random forests for efficient volume estimation. The experiments results suggested that our proposed method is feasible and can achieve higher accuracy, even in case of echocardiography images with irregular geometry.