Deep Learning Based Non-rigid Device Tracking in Ultrasound Image

Cardiac interventional procedure is an important type of minimally invasive surgery that deals with catheter-based treatment of cardiovascular disease, it is usually accompanied by X-ray fluoroscopy imaging as a typical setup, this is not desirable because the operator will be exposed to a significant radiation dose for a long time. Motivated by reducing the radiation hazard for cardiologists and sonographers during intervention, this research is aiming to develop a fully automatic and robust non-rigid device tracking and segmentation platform on pure ultrasound images. Based on the dataset collected on a tissue mimicking cardiac phantom, this research focuses on building up the automatic non-rigid device tracking model via deep learning method. Before training, the target tip is localized with an external EM tracking device, then the tracking framework is realized by automatic tip segmentation using U-net on both 2D and 3D TOE images.