DEEP LEARNING APPROACHES FOR OBJECT TRACKING AND MOTION ESTIMATION OF ULTRASOUND IMAGING SEQUENCES

Abstract

In recent decades, object tracking and motion estimation in medical imaging have gained importance. It is a powerful tool that can be used to improve diagnostic accuracy and therapy efficiency. This importance has led researchers to search for faster and more accurate algorithms for object tracking. Different approaches have been used to perform object tracking, such as object detection, motion estimation, and similarity matching, which are the focus of this study.

Different avenues can be used to address similarity matching. First, the classical method, which takes an object and searches for a similar object in the subsequent frame (because it is an object tracking in a video sequence) by examining all the sub-windows in the subsequent frame and measuring a cost function between the reference object and the sub-window. This approach is inefficient and cannot achieve real-time tracking. The deep learning method for similarity matching utilizes twin convolutional networks that produce a feature map that is later combined using a correlation layer. This layer provides a score map that points to a high-similarity area.

This study examined and developed object tracking algorithms to track objects of interest in the human liver using a correlation filter-based neural network (CFNet). The dataset used in this study was CLUST-2D, which was provided by the Swiss Federal Institute of Technology in Zürich (ETH). It contains approximately 96 ultrasound sequences of the liver from different patients. Three versions of the CFNet network were tested in this study. First, baseline-CFNet was used for training. Baseline-CFNet struggled to track objects under significant displacements and deformations. To address this limitation of the baseline-CFNet, a second version was developed. Advanced-CFNet is the second version of CFNet implemented in this study. This is the main contribution of this study. This version incorporates a dynamic template update and motion prediction modules, which improve object tracking by preventing tracker drift and maintaining the template from being polluted with inappropriate appearances of the tracked object. The third version implemented in this study is Kalman-CFNet, which utilizes a linear Kalman filter to estimate an object's motion and enhance its robustness against unexpected motions. The comparative analysis demonstrated the superiority of Advanced-CFNet, as it achieved lower root mean square error (RMSE) values than the other methods, particularly in challenging scenarios. These findings highlight the effectiveness of the advanced CFNet for object tracking in liver ultrasound imaging.