Motion tracking in 3D ultrasound imaging based on Markov-like deep-learning-based deformable registration

Abstract

Ultrasound imaging during or prior to radiation therapy offers a great potential in terms of safety, cost and real time imaging capacity. However, this task is challenging for tumors of abdominal such as liver cancer due to respiratory motion. In this work, we proposed an unsupervised deep-learning-based method to track the respiratory motion for 3D ultrasound (US) liver imaging. A Markov-like network, which extract features from consistent 3D US frames, was utilized to estimate a sequence of deformation vector fields (DVFs) that register tracked frame with landmark to match the coming untracked frames without landmark. Then, the landmarks of the coming frames were tracked by moving landmark position of tracked frame to the coming frames based on DVFs. The Markov-like network aims to consider motion consistency between each two frames and is implemented via a generative adversarial network (GAN). The proposed method was evaluated on the MICCAI CLUST 2015 challenge dataset. A retrospective study was performed with a total of 8 sets of 3D US sequence, and each 3D sequence has 4-96 frames marked with landmarks. We used a leave-one-out experiment to test our tracking method. Quantitative evaluation was performed by calculating the tracking error between estimated landmarks and the ground truth landmarks on each frame. Results of the proposed method showed a mean tracking error of 2.01±1.06 mm for 3D liver US images. We proposed a deep learning-based approach for 3D US liver motion tracking. This method reduces the processing time for tracking respiratory motion significantly, which can reduce the delivery uncertainty.