Deep Pose Tracking Estimator to Facilitate Ultrasound-guided Percutaneous Interventions

Ultrasound-guided, percutaneous interventions of the liver such as tumor biopsies and ablations have high clinical significance. Accurate localization of the target lesion in ultrasound (US) images and proper needle aiming can be challenging based on lesion location, echogenicity, and breathing-induced motion. Moreover, a frequently encountered challenge for the operator arises when the target and the needle cannot be jointly visualized within the ultrasound’s field of view. To facilitate the targeting procedure for manually executed interventions, we propose a deep pose tracking estimator for the needle and the target relying on ultrasound images only. Our tracking algorithm exploits deep neural network (DNN) based detections of target anatomy and needle tip in 2D ultrasound images in a statistical estimator to continuously and accurately estimate needle position and orientation and target position. With the estimated information, real time feedback can be provided on the correct needle insertion angle to reach a target lesion. We tested the proposed method in two different scenarios, in-vitro data collected from a 3D printed phantom and an anthropomorphic phantom and in-vivo clinical data collected by radiologists during a biopsy intervention. The results show that our approach facilitates needle aiming in dynamic situations, is robust towards non-optimal performance of the DNN detector and helps to address significant shortcomings of commercial, clinical US sensors.