IEEE Robotics & Automation Magazine - September 2023 - 87

either reselect the target or reposition the robot by using
the hand guidance functionality.
■ FS (tracking error monitoring): A safety mechanism
continuously estimates the error between the available
target position,
rbtn ,
r^h from the image-based tracking
and the one estimated by trajectory learning,
rbtn .
u^h If
this error is unacceptably larger than a user-defined
threshold (in our study, 2.5 mm), the HIFU sonication
functionality is disabled.
■ FS (acoustic coupling verification before sonication): The
verification of the coupling between the HIFU transducer
and patient's body is a necessary step for enabling HIFU
sonication. This is carried out by analyzing the radio frequency
data recorded by the confocal imaging probe during
a safe (i.e., 1 W in power and 1 s in duration) HIFU
sonication before the actual sonication, as described in a
previous work [25].
SOFTWARE ARCHITECTURE
This section provides a description of the overall software
architecture of the HIFUSK platform. The different modules
of the HIFUSK platform are controlled through dedicated
software developed in the Robot Operating System (ROS)
framework. Specifically, HIFUSK encompasses the following
six ROS nodes:
1) GUI node: This was developed using Qt Creator.
2) Robot control node: This involves a User Datagram
Protocol connection with the KUKA controller to trigger
the robot control states (hand guidance, position and force
control, and the brake). The KUKA controller runs an
application implemented with the KUKA Sunrise operating
system.
3) Echographic imaging node: This involves an ROS bridge
with the echographic probe, using the Open Image-Guided
Therapy Link protocol to stream the 500 × 500-pixel
B-mode US images as an ROS topic at 30 Hz.
4) Image-based tracking node: This uses an open source ROS
version of the OpenTLD tracker and provides continuous
tracking at 20 Hz on the US images (which is adequate to
track 0.2-0.3-Hz breath-induced motions).
5) Trajectory learning node: This uses the Python package
optimize.curve_fit to perform sinusoidal regression.
6) HIFU sonication node: This involves a TCP/IP connection
to direct the HIFU transducer and sonication parameters.
VALIDATION OF HIFUSK
This section describes the tests carried out to validate the
HIFUSK platform. This validation moved from controlled
laboratory conditions (when testing the technology on phantoms
and ex vivo animal tissues) to an in vivo clinical trial
on a pig.
DRY LAB VALIDATION
We first evaluated the HIFUSK platform in a dry lab (i.e.,
on phantoms) in terms of motion tracking and sonication
accuracy.
EXPERIMENTAL SETUP
The experimental setup to achieve this goal is illustrated in
Figure 7(a). Tests were conducted on a thermochromic silicone-based
phantom [15]: the phantom changes its color
according to temperature. Specifically, it turns white from
red when the local temperature exceeds 65 °C. This condition
can be achieved by thermal sonication. When the HIFU sonication
stops, the phantom cools down and recovers the original
red color. The phantom was designed in a cylindrical
shape with a central peg (1 cm in radius) as a shooting target.
It was attached to an acoustic absorber that was connected to
a 1-DoF motorized slide. The slide simulated the motion due
to respiration that has to be compensated throughout HIFU
therapy. Then, the motorized slide performed a breathing-like
sinusoidal motion with an amplitude of 10 mm and a frequency
of 0.2 Hz (i.e., 12 breaths per minute). All the setup elements
were immersed in a tank containing degassed water.
MOTION TRACKING AND LEARNING ANALYSIS
EXPERIMENTAL PROTOCOL
The results of the motion compensation experiments were evaluated
by comparing the target position from the encoder of the
motorized slide and the one estimated from the tracking and
learning modules. The target (i.e., the centroid of the phantom
peg) and a surrounding ROI were selected on the user interface.
The estimation of the target position from the tracking and
learning modules was observed for 10 s (after the initialization
of each module). This procedure was iterated 10 times.
RESULTS
Figure 7(b) displays the target position estimated by the tracking
module (red) and learning module (blue) together with the
signal from the motorized slide. Both the average error and
the maximum error across each repetition were computed.
The average error was always less than 1 mm for each repetition.
The average error from the learning module was lower
with respect to the one from the tracking module, although it
was characterized by higher variability across repetitions. In
terms of the maximum error, the learning module featured
lower values with respect to tracking, and it facilitated keeping
the median maximum error around 1.6 mm.
SONICATION ACCURACY ANALYSIS
EXPERIMENTAL PROTOCOL
The accuracy in performing HIFU sonication was evaluated
both in static and dynamic conditions, namely, simulating
the target motion due to breathing. The experimental protocol
included the execution of a 60-W, 10-s, continuous
HIFU sonication [26]. Through the GUI, the centroid of the
thermochromic phantom peg was selected as the shooting
target. During sonication, a 4K, 2,160-pixel, 30 frames-persecond
video was simultaneously recorded. Ten repetitions
were performed. A semiautomatic video segmentation algorithm
was implemented in MATLAB [Figure 7(c)]. For each
SEPTEMBER 2023 IEEE ROBOTICS & AUTOMATION MAGAZINE
87
IEEE Robotics & Automation Magazine - September 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2023
