Cascaded DNNs for detecting the position and orientation of Left Ventricle from 3D CT Scans

Abstract

Accurate and automated localization of the left ventricle (LV) of heart and identification of its’ orientation are crucial for effective diagnosis of various heart diseases. This work presents a fully automated deep learning based pipeline for generating a short-axis transformation of transversal 3D CT scans. The proposed novel solution consists of three independent convolutional neural networks (CNN) in cascaded sequential phases. In phase one, a binary classification network is used to determine a suitable 2D slice that clearly shows the left ventricle from the 3D image stack. In phase two, a different CNN is used to predict the location of 3 landmarks on the suitable 2D slice. From these landmarks a general orientation of the left ventricle is computed in the xy plane. In the last phase, a third CNN is used to predict the location of 3 landmarks on the orthogonal image. With all 6 points, a 3D orientation of the heart can be calculated. The proposed method performs as good as human accuracy, considering inter-observer variability, and performed equally well in terms of End-Diastolic Volume (EDV) and Left Ventricular Mass (LVM) calculation.