Imposing Object’s Trajectory and Dynamic Template Updates to Track ROIs in Ultrasound Image Sequences

This paper improves an existing object-tracking algorithm to track Regions of Interest (ROIs) in human liver ultrasound imaging sequences using a correlation filter-based Siamese neural network (advanced CFNet). Specifically, we impose object motion regularity to address a limitation of the baseline CFNet, which is losing the ROI when the object displaces and is deformed significantly. In addition, the proposed method uses a dynamic template update strategy to enable the recovery of the lost ROI. The dataset used in this study is publicly available, Challenge of Liver Ultrasound Tracking (CLUST 2015). It contains approximately 96 ultrasound sequences of the liver from different patients. We demonstrate that the proposed tracking method (Advanced CFNet) is robust for the dataset (CLUST- 2015) compared with the baseline CFNet.