IEEE Robotics & Automation Magazine - December 2016 - 123

into the view-planning platform.
Based on feature structures (chambers,
valves, vessels) shown in either longaxis view or short-axis view, five views
were defined as follows:
● view 1 (short axis): right and left
atria, pulmonary valve, and aortic
arch
● view 2 (long axis): pulmonary
artery, left atrium, left ventricle, and
pulmonary valve
● view 3 (long axis): right atrium,
right ventricle, and tricuspid valve
● view 4 (long axis): left atrium, left
ventricle, mitral valve, aortic arch,
and aortic valve
● view 5 (short axis): right and left
ventricles at midchamber.

Link
Mechanism

TEE Robot

Articulation Link for
the Translation Axis

X7-2t TEE Probe
Robot Control Laptop

iE33 Ultrasound
Machine
Custom HeartEsophagus Phantom

Experimental Setup

Figure 5. The experimental setup for the phantom experiment with the TEE robot,
ultrasound machine, and heart-esophagus phantom. An enlarged photo of the link
mechanism is shown in the top right.

Experimental and Postprocessing Methods
The TEE ultrasound probe was inserted into the robotic system with an ultrasound machine (iE33, Philips Healthcare,
The Netherlands) connected. The 3-D image data were
streamed out to a PC in real time via Transmission Control
Protocol/Internet Protocol. Two experiments using this
setup were designed to test the automatic acquisition of
TEE. The first experiment aimed to test the feasibility and
accuracy of automatic acquisition with open-loop control.
During the experiment, the probe head was manually
inserted into the heart-esophagus phantom with a random
starting position. A full-volume acquisition was performed
and the acquired 3-D ultrasound image was registered to the
MR segmentation. As described in the "Image-Based Probe
Position Tracking" section, the image-based probe tracking
method was then employed and the current probe pose was
obtained. From this starting position, the robotic movement
needed to obtain each target view was calculated. This step
was defined as the initialization and the robot was then
actuated to obtain each view relative to the known initial
pose. Full-volume 3-D images were acquired at each view
position. This was repeated five times with different initial
positions. For each of these initial positions, each view was
acquired two times with the probe moving through the
sequence from view 1 to view 5, then from view 5 to view 1.
The second experiment aimed to test the improvement in
the accuracy of the automatic acquisition using the imagebased probe tracking method to provide feedback adjustments. At each adjustment iteration, the actual position was
determined by the image-based tracking and a new movement was calculated to obtain the required view. If the registration failed during the image-based tracking method,
feedback adjustments were not performed and the current
view was skipped. In this study, we ran the experiments and
quantified the accuracy improvement over three iterations
of feedback adjustments. This experiment used four different initial poses and five views for each.

For analysis of the performances of the automatic acquisition, accurate measures of the probe positions at each
view were obtained using the automatic registration method described in the "Image-Based Probe Position Tracking"
section, followed by manual corrections as needed. Errors
in the probe positioning and image space were calculated
by comparing the positions actually obtained to the preplanned views. For the probe pose error, transformations
from probe coordinates to the MR coordinates for planned
and acquired views were calculated and decomposed to
give the transformation parameters: the Cardan angles
referring to rotations about the x, y, and z axes and the
translation distances in the x, y, and z axes. The distance
and the orientation error were defined separately as the
root sum square (RSS) of the differences between the x, y,
and z axes components. For the open-loop experiment, the
distance and the orientation errors of each view were calculated. For the closed-loop experiment, the distance and the
orientation errors after each feedback adjustment were calculated to quantify the improvement in probe positioning
accuracy after different numbers of feedback adjustments.
To quantify the error in the image space, we then defined a
number of marker points in the ultrasound image coordinates (90° × 90° cone, 10-cm depth). Ten image planes were
selected within the TEE field of view, each parallel to the
transducer of the TEE probe. The interval between two
planes was 10 mm. The marker points were defined on the
four corners of the image plane [Figure 6(a)]. From the initial preplanned views in the view-planning platform and
the acquired ultrasound views, the locations of corresponding marker points were obtained and compared in MR
coordinates. As shown in Figure 6(b), this was done by
transforming the marker point locations from the ultrasound image coordinates to the MR coordinates based on
the probe positions defined by the view-planning platform
for planned views and by the registration result for real
acquired views. Mean position errors between corresponding
DECEMBER 2016

IEEE ROBOTICS & AUTOMATION MAGAZINE

123

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2016