Purpose: The capability of providing video-rate volumetric images with zero ionizing radiation dose makes ultrasound (US) imaging suitable for intra-fraction motion tracking in radiotherapy. This study aims to develop a deep learning-based algorithm for liver motion tracking using US images.
Methods: We propose an unsupervised deformable image registration network for US liver motion tracking. A Markov-like network is used to extract features from sequential US frames (one tracked frame is followed by untracked frames) to estimate a set of deformation vector fields (DVFs) through the registration of the tracked frame and the untracked frames. This network is optimized in an unsupervised manner, namely, without the need for ground truth DVFs. The positions of the landmarks in the untracked frames are finally determined by transforming the landmarks in the tracked frame according to the estimated DVFs. The performance of the proposed method was evaluated on the testing dataset by calculating the tracking error between the predicted and ground truth landmarks on each frame.
Results: The proposed method was assessed on the MICCAI CLUST 2015 dataset, which contains 63 2D and 22 3D US image sequences respectively from 42 and 18 subjects collected using 7 US scanners with 8 types of transducers. The experimental results show that our proposed method achieved a mean tracking error of 0.70±0.38 mm on the 2D sequences and 1.71±0.84 mm on the 3D sequences.
Conclusion: We have proposed and investigated a new deep learning-based algorithm for liver motion tracking using US images and demonstrated its feasibility for real-time motion tracking. This imaging tool has strong potential for US-guided motion tracking during liver radiotherapy.