Feature Description using Autoencoders for Fast 3D Ultrasound Tracking

Abstract 3D ultrasound imaging is a promising modality for therapy guidance, e.g. in radiation therapy. It is able to provide volumetric soft tissue images in real-time. However, due to low image quality, high noise ratio and high data dimensionality, real-time capable US image processing methods like target tracking are challenging. In this study, a feature-based tracking approach is investigated. The FAST feature detector is used to detect local image features in 3D ultrasound images. Two different feature descriptors are tested and evaluated in terms of target tracking: The BRIEF descriptor as well as a slicedwasserstein autoencoder. On the basis of a feature matching algorithm, tracking experiments are executed and evaluated using eight labeled 3D US sequences. The mean tracking error measured is 2.08±1.50mm and 2.29±1.59mm using the autoencoder and the BRIEF descriptor, respectively. The results indicate that using an autoencoder for feature description improves the tracking performance compared to a binary descriptor. The proposed tracking method could be executed in fast runtimes of 137 ms and 256 ms per image on average making it real-time capable.