Landmark tracking in 4D ultrasound using generalized representation learning

In this study, we present and validate a novel concept for target tracking in 4D ultrasound. The key idea is to replace image patch similarity metrics by distances in a latent representation. For this, 3D ultrasound patches are mapped into a representation space using sliced-Wasserstein autoencoders.

A novel target tracking method for 4D ultrasound is presented that performs tracking in a representation space instead of in images space. Sliced-Wasserstein autoencoders are trained in an unsupervised manner which are used to map 3D ultrasound patches into a representation space. The tracking procedure is based on a greedy algorithm approach and measuring distances between representation vectors to relocate the target . The proposed algorithm is validated on an in vivo data set of liver images. Furthermore, three different concepts for training the autoencoder are presented to provide cross-patient generalizability, aiming at minimal training time on data of the individual patient.

Eight annotated 4D ultrasound sequences are used to test the tracking method. Tracking could be performed in all sequences using all autoencoder training approaches. A mean tracking error of 3.23 mm could be achieved using generalized fine-tuned autoencoders. It is shown that using generalized autoencoders and fine-tuning them achieves better tracking results than training subject individual autoencoders.

It could be shown that distances between encoded image patches in a representation space can serve as a meaningful measure of the image patch similarity, even under realistic deformations of the anatomical structure. Based on that, we could validate the proposed tracking algorithm in an in vivo setting. Furthermore, our results indicate that using generalized autoencoders, fine-tuning on only a small number of patches from the individual patient provides promising results.