Neural-Network-Based Heart Motion Prediction for Ultrasound-Guided Beating-Heart Surgery

Abstract

A neural-network-based heart motion prediction method is proposed for ultrasound-guided beating-heart surgery to compensate for time delays caused by ultrasound (US) image acquisition and processing. Such image processing is needed for tracking heart tissue in US images, which is itself a requirement for beating-heart surgery. Once the heart tissue is tracked in US images, a recurrent neural network (NN) is employed to learn how to predict the motion of the tracked heart motion in order to compensate for the delays introduced in the initial US image processing step. To verify the feasibility of predicting both simple and complex heart motions, the NN is tested with two types of heart motion data: (i) fixed heart rate and maximum amplitude, and (ii) varying heart rate and maximum amplitude. Also, the NN was tested for different prediction horizons and showed effectiveness for both small and large delays. The heart motion prediction results using NN are compared to the results using an extended Kalman filter (EKF) algorithm. Using NN, the mean absolute error and the root mean squared error between the predicted and the actually tracked heart motions are roughly 60% smaller than those achieved by using the EKF. Moreover, the NN is able to predict the heart position up to 1000 ms in advance, which significantly exceeds the typical US image acquisition/processing delays for this application (160 ms in our tests). Overall, the NN predictor shows significant advantages (higher accuracy and longer prediction horizon) compared to the EKF predictor.