Optical Flow Networks for Heartbeat Estimation in 4D Ultrasound Images

Congenital heart defects is one of the most common neonatal diseases and has a very low survival rate. The fetal heart is generally smaller and possesses a faster than normal beating rate, thus making medical diagnosis difficult. The efficiency and accuracy of diagnosis of congenital heart disease can be improved by computer-aided diagnostic methods. Optical flow is a robust algorithm for object recognition and motion detection, and has potential in early detection of congenital heart defects. In this paper, an end-to-end deep learning system is proposed for obtaining the optical flow information from 4D fetal cardiac ultrasound images. The optical flow network model is trained by using gradients of image sequences obtained from a virtual data set. Subsequently, the trained model is used to detect the cardiac motion. Experimental results and performance evaluation demonstrate the effectiveness of the proposed network. Apart from the efficacy of the proposed method, a visualization of the fetal cardiac motion using pseudo-color is provided. It is envisaged that the proposed method can be used in clinical applications requiring automatic detection of congenital fetal heart defects.