<s>
U-Net	B-Application
is	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
that	O
was	O
developed	O
for	O
biomedical	O
image	B-Algorithm
segmentation	I-Algorithm
at	O
the	O
Computer	O
Science	O
Department	O
of	O
the	O
University	O
of	O
Freiburg	O
.	O
</s>
<s>
The	O
network	O
is	O
based	O
on	O
the	O
fully	O
convolutional	O
network	O
and	O
its	O
architecture	O
was	O
modified	O
and	O
extended	O
to	O
work	O
with	O
fewer	O
training	O
images	O
and	O
to	O
yield	O
more	O
precise	O
segmentations	B-Algorithm
.	O
</s>
<s>
Segmentation	B-Algorithm
of	O
a	O
512×512	O
image	O
takes	O
less	O
than	O
a	O
second	O
on	O
a	O
modern	O
GPU	B-Architecture
.	O
</s>
<s>
The	O
U-Net	B-Application
architecture	O
stems	O
from	O
the	O
so-called	O
“	O
fully	O
convolutional	O
network	O
”	O
first	O
proposed	O
by	O
Long	O
,	O
Shelhamer	O
,	O
and	O
Darrell	O
.	O
</s>
<s>
The	O
main	O
idea	O
is	O
to	O
supplement	O
a	O
usual	O
contracting	O
network	O
by	O
successive	O
layers	O
,	O
where	O
pooling	O
operations	O
are	O
replaced	O
by	O
upsampling	B-Algorithm
operators	O
.	O
</s>
<s>
One	O
important	O
modification	O
in	O
U-Net	B-Application
is	O
that	O
there	O
are	O
a	O
large	O
number	O
of	O
feature	O
channels	O
in	O
the	O
upsampling	B-Algorithm
part	O
,	O
which	O
allow	O
the	O
network	O
to	O
propagate	O
context	O
information	O
to	O
higher	O
resolution	O
layers	O
.	O
</s>
<s>
The	O
network	O
only	O
uses	O
the	O
valid	O
part	O
of	O
each	O
convolution	B-Language
without	O
any	O
fully	O
connected	O
layers	O
.	O
</s>
<s>
This	O
tiling	O
strategy	O
is	O
important	O
to	O
apply	O
the	O
network	O
to	O
large	O
images	O
,	O
since	O
otherwise	O
the	O
resolution	O
would	O
be	O
limited	O
by	O
the	O
GPU	B-Architecture
memory	O
.	O
</s>
<s>
U-Net	B-Application
was	O
created	O
by	O
Olaf	O
Ronneberger	O
,	O
Philipp	O
Fischer	O
,	O
Thomas	O
Brox	O
in	O
2015	O
and	O
reported	O
in	O
the	O
paper	O
“	O
U-Net	B-Application
:	O
Convolutional	O
Networks	O
for	O
Biomedical	O
Image	B-Algorithm
Segmentation	I-Algorithm
”	O
.	O
</s>
<s>
"	O
Fully	O
convolutional	O
networks	O
for	O
semantic	B-Algorithm
segmentation	I-Algorithm
"	O
.	O
</s>
<s>
The	O
contracting	O
path	O
is	O
a	O
typical	O
convolutional	O
network	O
that	O
consists	O
of	O
repeated	O
application	O
of	O
convolutions	B-Language
,	O
each	O
followed	O
by	O
a	O
rectified	B-Algorithm
linear	I-Algorithm
unit	I-Algorithm
(	O
ReLU	B-Algorithm
)	O
and	O
a	O
max	O
pooling	O
operation	O
.	O
</s>
<s>
The	O
expansive	O
pathway	O
combines	O
the	O
feature	O
and	O
spatial	O
information	O
through	O
a	O
sequence	O
of	O
up-convolutions	O
and	O
concatenations	O
with	O
high-resolution	O
features	O
from	O
the	O
contracting	O
path	O
.	O
</s>
<s>
There	O
are	O
many	O
applications	O
of	O
U-Net	B-Application
in	O
biomedical	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
such	O
as	O
brain	O
image	B-Algorithm
segmentation	I-Algorithm
(''	O
BRATS''	O
)	O
and	O
liver	O
image	B-Algorithm
segmentation	I-Algorithm
(	O
"	O
siliver07	O
"	O
)	O
as	O
well	O
as	O
protein	O
binding	O
site	O
prediction	O
.	O
</s>
<s>
Variations	O
of	O
the	O
U-Net	B-Application
have	O
also	O
been	O
applied	O
for	O
medical	O
image	O
reconstruction	O
.	O
</s>
<s>
Here	O
are	O
some	O
variants	O
and	O
applications	O
of	O
U-Net	B-Application
as	O
follows	O
:	O
</s>
<s>
Pixel-wise	O
regression	O
using	O
U-Net	B-Application
and	O
its	O
application	O
on	O
pansharpening	O
;	O
</s>
<s>
3D	O
U-Net	B-Application
:	O
Learning	O
Dense	O
Volumetric	O
Segmentation	B-Algorithm
from	O
Sparse	O
Annotation	O
;	O
</s>
<s>
TernausNet	O
:	O
U-Net	B-Application
with	O
VGG11	O
Encoder	O
Pre-Trained	O
on	O
ImageNet	O
for	O
Image	B-Algorithm
Segmentation	I-Algorithm
.	O
</s>
<s>
U-Net	B-Application
source	O
code	O
from	O
Pattern	O
Recognition	O
and	O
Image	O
Processing	O
at	O
Computer	O
Science	O
Department	O
of	O
the	O
University	O
of	O
Freiburg	O
,	O
Germany	O
.	O
</s>
