Interactive Translation in Echocardiography Training System With Enhanced Cycle-GAN
Long Teng, Zhongliang Fu, Yu Yao
Abstract
Interactive translation in echocardiography training system refers to the pixel-wise translation between ultrasound cardiac and theoretical sketch images in the course of hand-on operation. It is capable of efficiently gaining more insights into clinical ultrasound anatomy. However, major studies on the synthesis of ultrasound cardiac image primarily discuss the physical model simulation, while studies on cardiac image segmentation place an emphasis on image processing. Thus, they cannot be easily integrated into one pipeline for interactive translation. This paper presents an enhanced Cycle-GAN for interactive translation. Perceptual loss is introduced to enhance the quality of synthetic ultrasound texture, while Cycle-GAN translates between two modalities. The proposed method is trained on 300 pair images and tested on 68 pair images. As revealed from the experiment results, the proposed method is feasible in interactive translation, and it is superior over Cycle-GAN for ultrasound image synthesis.