Endocardium tracking by fusing optical flows in straightened images with learning based detections

The endocardium tracking in ultrasound images is challenging due to large shape variations and the signal dropout. In this paper, we present a method to fuse multiple information sources to robustly track the endocardium. The first novelty of the method is to perform tracking in a straightened shape space, to minimize the image pattern changes caused by cardiac motions. Straightened images are used in an optical flow based tracking method to accurately estimate endocardium motions. The second novelty is to fuse the optical flow tracking method with learning based detections to improve tracking accuracy. We demonstrate through experiments that the presented method can achieve robust and accurate tracking of endocardial boundary.