Fast Ultrasound to Ultrasound Auto-Registration for Interventional Cardiology

Many interventional procedures rely on 3D trans-esophageal echocardiography (TEE) imaging to guide the interventionist during instrument/device manipulation inside the heart. Manual registration of ultrasound to ultrasound or other modalities is commercially available. However, auto-registration is highly desirable, to avoid manual re-labelling and to estimate probe motion for motion compensation during surgery. There are a number of published auto-registration methods, but they are not suitable for interventional setups as they are not fast enough. In this work we present a close to real-time 3D ultrasound volume to 3D ultrasound volume auto-registration method that is implemented on the GPU. The auto-registration method performed on the clinical data was successful and good visual alignment was observed in 5 out of 7 cases. Decomposing a volume into its polynomial coefficients took 73ms and was performed once for each pyramid level. One iteration took on an average of 53ms. The proposed registration method is fast enough to allow real-time usage on the ultrasound scanner. When the method converges to a good result it is close to the expert's manual registration with discrepancies of 0.48 ± 2.1mm for translation in the X direction, -0.8 ± 2.0mm for translation in the Y direction, 0.86 ± 1.8mm for translation in the Z direction, -1.82 ± 1.54° for rotation about the X axis, 1.47 ± 2.77° for rotation about the Y axis, and -0.63 ± 0.76° for rotation about the Z axis.