<s>
Deep	B-Algorithm
learning	I-Algorithm
in	I-Algorithm
photoacoustic	I-Algorithm
imaging	I-Algorithm
combines	O
the	O
hybrid	O
imaging	O
modality	O
of	O
photoacoustic	B-Algorithm
imaging	I-Algorithm
(	O
PA	O
)	O
with	O
the	O
rapidly	O
evolving	O
field	O
of	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Photoacoustic	B-Algorithm
imaging	I-Algorithm
is	O
based	O
on	O
the	O
photoacoustic	O
effect	O
,	O
in	O
which	O
optical	O
absorption	O
causes	O
a	O
rise	O
in	O
temperature	O
,	O
which	O
causes	O
a	O
subsequent	O
rise	O
in	O
pressure	O
via	O
thermo-elastic	O
expansion	O
.	O
</s>
<s>
Photoacoustic	B-Algorithm
imaging	I-Algorithm
has	O
applications	O
of	O
deep	B-Algorithm
learning	I-Algorithm
in	O
both	O
photoacoustic	O
computed	O
tomography	O
(	O
PACT	O
)	O
and	O
photoacoustic	O
microscopy	O
(	O
PAM	O
)	O
.	O
</s>
<s>
The	O
one	O
of	O
the	O
first	O
applications	O
of	O
deep	B-Algorithm
learning	I-Algorithm
in	O
PACT	O
was	O
by	O
Reiter	O
et	O
al	O
.	O
</s>
<s>
After	O
this	O
initial	O
implementation	O
,	O
the	O
applications	O
of	O
deep	B-Algorithm
learning	I-Algorithm
in	O
PACT	O
have	O
branched	O
out	O
primarily	O
into	O
removing	O
artifacts	O
from	O
acoustic	O
reflections	O
,	O
sparse	O
sampling	O
,	O
limited-view	O
,	O
and	O
limited-bandwidth	O
.	O
</s>
<s>
There	O
has	O
also	O
been	O
some	O
recent	O
work	O
in	O
PACT	O
toward	O
using	O
deep	B-Algorithm
learning	I-Algorithm
for	O
wavefront	O
localization	O
.	O
</s>
<s>
There	O
have	O
been	O
networks	O
based	O
on	O
fusion	O
of	O
information	O
from	O
two	O
different	O
reconstructions	O
to	O
improve	O
the	O
reconstruction	O
using	O
deep	B-Algorithm
learning	I-Algorithm
fusion	O
based	O
networks	O
.	O
</s>
<s>
In	O
Reiter	O
et	O
al.	O
,	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
similar	O
to	O
a	O
simple	O
VGG-16	O
style	O
architecture	O
)	O
was	O
used	O
that	O
took	O
pre-beamformed	O
photoacoustic	O
data	O
as	O
input	O
and	O
outputted	O
a	O
classification	O
result	O
specifying	O
the	O
2-D	O
point	O
source	O
location	O
.	O
</s>
<s>
This	O
utilization	O
of	O
deep	B-Algorithm
learning	I-Algorithm
trained	O
on	O
simulated	O
data	O
produced	O
in	O
the	O
MATLAB	B-Language
,	O
and	O
then	O
later	O
reaffirmed	O
their	O
results	O
on	O
experimental	O
data	O
.	O
</s>
<s>
Traditionally	O
these	O
artifacts	O
were	O
removed	O
with	O
slow	O
iterative	O
methods	O
like	O
total	B-Algorithm
variation	I-Algorithm
minimization	I-Algorithm
,	O
but	O
the	O
advent	O
of	O
deep	B-Algorithm
learning	I-Algorithm
approaches	O
has	O
opened	O
a	O
new	O
avenue	O
that	O
utilizes	O
a	O
priori	O
knowledge	O
from	O
network	O
training	O
to	O
remove	O
artifacts	O
.	O
</s>
<s>
In	O
the	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
that	O
seek	O
to	O
remove	O
these	O
sparse	O
sampling	O
,	O
limited-bandwidth	O
,	O
and	O
limited-view	O
artifacts	O
,	O
the	O
typical	O
workflow	O
involves	O
first	O
performing	O
the	O
ill-posed	O
reconstruction	O
technique	O
to	O
transform	O
the	O
pre-beamformed	O
data	O
into	O
a	O
2-D	O
representation	O
of	O
the	O
initial	O
pressure	O
distribution	O
that	O
contains	O
artifacts	O
.	O
</s>
<s>
Then	O
,	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
CNN	B-Architecture
)	O
is	O
trained	O
to	O
remove	O
the	O
artifacts	O
,	O
in	O
order	O
to	O
produce	O
an	O
artifact-free	O
representation	O
of	O
the	O
ground	B-General_Concept
truth	I-General_Concept
initial	O
pressure	O
distribution	O
.	O
</s>
<s>
Another	O
technique	O
was	O
proposed	O
using	O
a	O
simple	O
CNN	B-Architecture
based	O
architecture	O
for	O
removal	O
of	O
artifacts	O
and	O
improving	O
the	O
k-wave	O
image	O
reconstruction	O
.	O
</s>
<s>
Prior	O
to	O
deep	B-Algorithm
learning	I-Algorithm
,	O
the	O
limited-view	O
problem	O
was	O
addressed	O
with	O
complex	O
hardware	O
such	O
as	O
acoustic	O
deflectors	O
and	O
full	O
ring-shaped	O
transducer	O
arrays	O
,	O
as	O
well	O
as	O
solutions	O
like	O
compressed	O
sensing	O
,	O
weighted	O
factor	O
,	O
and	O
iterative	O
filtered	O
backprojection	O
.	O
</s>
<s>
The	O
result	O
of	O
this	O
ill-posed	O
reconstruction	O
is	O
imaging	O
artifacts	O
that	O
can	O
be	O
removed	O
by	O
CNNs	B-Architecture
.	O
</s>
<s>
The	O
deep	B-Algorithm
learning	I-Algorithm
algorithms	O
used	O
to	O
remove	O
limited-view	O
artifacts	O
include	O
U-net	O
and	O
FD	O
U-net	O
,	O
as	O
well	O
as	O
generative	B-Algorithm
adversarial	I-Algorithm
networks	I-Algorithm
(	O
GANs	O
)	O
and	O
volumetric	O
versions	O
of	O
U-net	O
.	O
</s>
<s>
The	O
typical	O
method	O
to	O
remove	O
artifacts	O
and	O
denoise	O
limited-bandwidth	O
reconstructions	O
before	O
deep	B-Algorithm
learning	I-Algorithm
was	O
Wiener	O
filtering	O
,	O
which	O
helps	O
to	O
expand	O
the	O
PA	O
signal	O
's	O
frequency	O
spectrum	O
.	O
</s>
<s>
The	O
primary	O
advantage	O
of	O
the	O
deep	B-Algorithm
learning	I-Algorithm
method	O
over	O
Wiener	O
filtering	O
is	O
that	O
Wiener	O
filtering	O
requires	O
a	O
high	O
initial	O
signal-to-noise	O
ratio	O
(	O
SNR	O
)	O
,	O
which	O
is	O
not	O
always	O
possible	O
,	O
while	O
the	O
deep	B-Algorithm
learning	I-Algorithm
model	O
has	O
no	O
such	O
restriction	O
.	O
</s>
<s>
Photoacoustic	O
microscopy	O
differs	O
from	O
other	O
forms	O
of	O
photoacoustic	B-Algorithm
tomography	I-Algorithm
in	O
that	O
it	O
uses	O
focused	O
ultrasound	O
detection	O
to	O
acquire	O
images	O
pixel-by-pixel	O
.	O
</s>
<s>
PAM	O
images	O
are	O
acquired	O
as	O
time-resolved	O
volumetric	O
data	O
that	O
is	O
typically	O
mapped	O
to	O
a	O
2-D	O
projection	O
via	O
a	O
Hilbert	B-Algorithm
transform	I-Algorithm
and	O
maximum	O
amplitude	O
projection	O
(	O
MAP	O
)	O
.	O
</s>
<s>
The	O
first	O
application	O
of	O
deep	B-Algorithm
learning	I-Algorithm
to	O
PAM	O
,	O
took	O
the	O
form	O
of	O
a	O
motion-correction	O
algorithm	O
.	O
</s>
<s>
The	O
two	O
primary	O
motion	O
artifact	O
types	O
addressed	O
by	O
deep	B-Algorithm
learning	I-Algorithm
in	O
PAM	O
are	O
displacements	O
in	O
the	O
vertical	O
and	O
tilted	O
directions	O
.	O
</s>
<s>
used	O
a	O
simple	O
three	O
layer	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
,	O
with	O
each	O
layer	O
represented	O
by	O
a	O
weight	O
matrix	O
and	O
a	O
bias	O
vector	O
,	O
in	O
order	O
to	O
remove	O
the	O
PAM	O
motion	O
artifacts	O
.	O
</s>
