Trackable-speckle detection using a dual-path convolutional neural network for nodes selection in speckle tracking echocardiography

Speckle tracking echocardiography (STE) is widely used to quaantify regional motion and deformation of heart tissues. Before tracking, a segmentation step is first carried out, and only a set of nodes in the segmented model are tracked. However, a random selection of the nodes even after tissue segmentation could lead to an inaccurate estimation. In this paper, a convolutional neural network (CNN)-based method is presented to detect trackable speckle spots that have important properties of the texture for speckle tracking. The proposed CNN was trained and validated on 29500 ultrasound manually labelled image patches extracted from the echocardiography of 65 people. Using the proposed network, in silico experiments for automatic node selection were conducted to investigate the applicability of the proposed method in speckle tracking. The results were statistically highly significant (P<0.001) and demonstrated that the proposed method has the least tracking error among various existing methods.