IEEE Systems, Man and Cybernetics Magazine - October 2023 - 48

In the process of US propagation in human tissue, diffusion
attenuation, scattering attenuation, and absorption
attenuation will occur, and ultrasound sound intensity will
attenuate with the increase in propagation distance.
Therefore, we first amplify the echo signal according to
the propagation distance through time-varying gain compensation.
Generally speaking, the attenuation of US
waves in the process of human tissue propagation can be
considered to conform to the exponential law. Therefore,
we design an exponential variable gain amplifier to amplify
the echo signal. Assuming that
signal, the signal xt
xt
sation is
xt xt et
ampraw
() () ()
=
afd
(1)
where a is the sound attenuation coefficient, f is the
center frequency of the US signal, and d is the propagation
distance of the US signal. For human skeletal muscle,
the attenuation coefficient is
slightly less than 1 dB/(cm·MHz).
Convert dB to the true value,
a.0.058/(cm·MHz). Here, we take
a 00 . 5=
/(cm·MHz).
The echo signal after the timevarying
gain compensation contains
obvious low-frequency and
high-frequency noise. To enhance
meaningful signal components, we
use a Gaussian bandpass filter to
filter out noise. Specifically, we
first perform Fourier transform on
the echo signal and then multiply it
by a Gaussian filter function, as
shown next
Gf
()=- v
e
1
2rv
f
2 2
2
raw () is the original US
amp () after time-varying gain compenMethodology
In
this article, we devise a novel strategy employing NNs
for the feature extraction and dynamic hand gesture classification
of AUS signals. We select and compare two distinct
NNs. The specific details of the experiment will be
discussed in subsequent sections.
Deep Learning Framework
In this article, we
devise a novel
strategy employing
NNs for the feature
extraction and
dynamic hand
gesture classification
of AUS signals.
(2)
where f represents frequency. v is the standard deviation
of the Gaussian function, which is calculated from the relative
bandwidth of the bandpass filter, as shown next
v=
where fc
bandwidth ·
100
% ()lg
f
c
22 2 2
(3)
is the filter center frequency. Then, we get the filtered
signal through the inverse Fourier transform. In this
section, we choose bandwidth as 100%.
Finally, we carry out logarithmic compression on the
echo signal to compress the amplitude of the signal to a
certain range for easy calculation, as follows:
xt log
c 04= ..
log compression =++1 c 1 cxenvelope t
() (· ())
(4)
where c is the logarithmic compression coefficient, we
choose
48 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE October 2023
LSTM
LSTM is a variant of the recurrent NN (RNN) primarily
designed to address the vanishing and exploding gradient
issues during the training of long sequences. The main
module of this NN structure is the LSTM module, which
includes the forget gate, input gate, and output gate. The
forget gate, crucial for accomplishing long-term memory
tasks, is the most important part of each LSTM block [27].
The LSTM network can capture the temporal dependencies
of time series and has achieved
a series of results in time-series prediction
[28]. The cyclic NN possesses
memory, shared parameters, and
complete Turing capabilities, offering
advantages in learning the nonlinear
characteristics of sequences.
RNNs are applied in natural language
processing, including speech
recognition, language modeling,
machine translation, and various
time-series predictions, catering to
our core requirements. LSTM can
group a certain number of frames
as a whole, effectively utilizing the
tempora l
character i s t ics
sequences. By employing a gesture
recognition strategy with LSTM as the core, we can provide
the AUS signal with the missing time information,
referred to as time correlation. Additionally, with the aid
of time correlation, we can accurately recognize the entire
process of a dynamic gesture, which is an issue that has
remained unsolved in previous research.
CNN
CNN is a feedforward NN typically composed of one or
more convolutional and fully connected layers. The
receptive field is integrated into CNN to simulate the
way the human brain processes information. Currently,
CNN is the most widely used network, adept at extracting
hierarchical and spatial features [13], making it suitable
for image recognition, image segmentation, target
tracking, and other applications [29]. Based on the principle
of AUS signal acquisition [18], we can infer that the
AUS signal is captured as a unit for each frame, with
each image unit representing muscle depth and signal
strength. To address the requirements of image recognition
and image segmentation, researchers initially
of
IEEE Systems, Man and Cybernetics Magazine - October 2023

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - October 2023
