Robust Landmark Detection of the Left Ventricle Based on 2D Ultrasound Image Data

Abstract

Echocardiography, or cardiac ultrasound, is a convenient imaging modality for bedside examination of the heart in real-time. However, acquiring ultrasound images of good quality is challenging, and currently ultrasound imaging of the heart is limited to expert echocardiographers. In order for ultrasound to be available for inexperienced ultrasound users, an aiding tool based on augmented reality has been developed. Even so, the current prototype requires 3D ultrasound volume as input. This feature is not available on the portable ultrasound scanners which inexperienced users are more likely to use. In order for the current prototype to work on portable ultrasound scanners, a fan-shaped volume acquired of 2D ultrasound images with its known orientation is believed to be able to substitute the 3D ultrasound image volume input needed today.   The aim of the master thesis is to investigate the possibility of changing the 3D ultrasound volume input of the current prototype with 2D ultrasound images with known orientation. Important anatomic landmark locations as well as the performance of edge detection are investigated. Several acquisition strategies are tested in order to investigate whether the 3D ultrasound image volume can be changed to 2D ultrasound images with known orientation. Firstly, 2D ultrasound image slices with known orientation are extracted from a pre-recorded 3D ultrasound image volume with known ground truth. Secondly, is to observe what happens when the effect of electrocardiogram (ECG) triggering is removed. Thirdly, free-hand scanning in-vivo lab where the fan-shaped volume are created with acquiring 2D ultrasound images with known position and orientation are tracked by an optical tracker.   The thesis concludes that the 3D ultrasound volume can be substituted with 2D ultrasound images with known orientation. Important factors to achieve satisfactory performance include symmetric probe movement, adjusting the sector azimuth and depth such that the surrounding left ventricular wall and septum are shown clearly in the ultrasound image as well as voiding the elevation rotation since this can create noisy ultrasound images.