Echocardiography View Classification Using Quality Transfer Star Generative Adversarial Networks

2D echocardiography (echo) is the most widely used imaging technique to identify cardiac disease. In addition to anatomical variability in patients, the quality of acquired echo image can vary significantly depending on the ultrasound (US) machine and the experience level of the operator, where a poor image quality can affect the diagnosis. This variability can also result in reduced performance of machine learning models trained on these data. With the recent advances in generative adversarial networks (GAN), we demonstrate that it is possible to transfer the image quality of echo images to a user-defined quality level with the use of a multi-domain transfer approach referred as StarGAN. The proposed quality transfer StarGAN (QT-StarGAN) requires no pairs of low-and high-quality echo images and incorporates the temporal information of echo images during the training phase. We evaluate the proposed approach using 16,612 echo cine series obtained from 3,157 patients. Using a standard echo view classification task, we demonstrate that the accuracy of classification is significantly improved using QT-StarGAN.