Echocardiogram View Conversion using Clinically Constrained Conditional GAN

Abstract

Echocardiography (echo) is a primary imaging modality to study cardiovascular conditions. The echo examination involves capturing a set of standardized echo views, and accurate acquisition needs the expertise of an experienced sonographer. In this paper, we aim to facilitate such acquisition by developing a predictive echo view conversion model to provide the novice sonographer with an estimate of what the next standard view may look like, given a captured view. We propose a novel constrained conditional generative adversarial network (CGAN) architecture to predict phase-synchronized echos seen from an alternate echo view to the input. The proposed training framework imposes a structured regularization to the translation informed by heart ventricle and atrium size. Quantitative assessment of the images shows an 84% correlation between the segmentation mask area of generated and ground-truth images, demonstrating the framework’s ability to produce accurate predictions for a multitude of cardiac geometries.