0022-1554/79/2701-0128$02.00/0 THE
JOURNAL
Copyright
OF HISTOCHEMISTRY
AND
Iterative
Vol.
CYTOCHEMISTRY
© 1979 by The Hiatochemical
Society,
Image
1, pp.
128-135,
Printed
Transformations
for an Automatic Smears FERNAND
Centre
27, No.
Inc.
de Morphologie
Mathematique,
Ecole
Received
for
Screening
1979
in U.S.A.
of Cervical
MEYER des
Mines
de
publication
June
Paris, 19,
77305
Fontainebleau,
France
1978
The new generation of image analysis systems permits the use of iterative image transformations. It is now possible to construct algorithms where the elementary steps are not arithmetic operations but image transformations. This will be illustrated by two examples. In the first, the absorption image of Feulgen Stained nuclei is processed by contrast algorithms in order to detect suspect cells. In the second, free lying cells are separated from overlapping cells and other artefacts by the use of skeletonization procedures.
Through the presentation ofa program ofcervical screening we want to introduce the reader to a new methodology pattern recognition. The whole process is done automatically by
a real
time
T.V.
image
analysis
system
described
defmed; the eroded set is the unburnt part of at time t = X, (in gray in Fig. 1A). Dilating a the same thing as eroding its complementary If we want to restore the eroded set A’ to
of
elsewhere
we can
(13). The contrast of the nuclei in the absorption image of a Feulgen stained slide is used for the detection of suspect cells and the elimination of many types of artefacts. At a higher resolution the fluorescence image gives a mask for the nuclei; shape analyzing of them eliminates the artefacts and automatically
separates
method,
no
the
clustering
measurements
cells.
are
made
With
in the
regard
to
whole
present But
the
in a field at the main interest
same time lies in the
basic transformations. Since they hardware, high speed and processing And only under that very condition be of any practical interest.
and not simplicity
one
the
agnosed,
comes
details
about
(5)).
material,
The
stained staining
from
the
all
Dr.
apparatus
analysis
system
scribed
in (14).
we with
after of the
other.
We
will
introduce
morphology
an Acriflavine double staining group
in
is bit
can
the
classical
planes,
Feulgenand di-
Leiden
be
2, 6-10,
12,
13).
Imagine
that
to analyze
is composed
background Suppose
of the figure a fire is started
the
boundary.
the
middle
the
fire
The
of dry
fire
of the
at times
important
texture
extensively
figure. t
=
interior
grass
is composed simultaneously will
propagate Figure
0, 2, 4. The
tools but
the
size
used
rigorous
These
points points
then of
skeleton
eliminating
of the
and
that
the
at uniform
erosion
figure
of unburnable at all
1A shows of size
the
(see we
wet points
have
speed
toward lines
X is then
will
or grass. along
front
of
of
128
size
followed
comparison
size
X works
of like
a
called
skeleton
constant.
intersect fronts
points
skeleton
advancing
the
fire
center
of
an
velocity can
the
thus
skeleton
is less than the branches
application
than and
of
the
a
the
from each
fire;
figure.
lines
will
circle.
the If the
describe
radius
the
in Figure the
of the
of
decreasing
We
from
figure
the fire front will extinguish
quench the
the
at
original
until
For
the
Figure
1
1D. of the
define all
fire
is higher
a new
branches
than
1
transformation along
which
the
a given value k > 1. In Figure of the skeleton along which 4. The each
opening
skeleton of a connected of its components marks binary In this to
contrast
images but in last section we analysis
in
gray tone images. A gray tone image may be represented a chain of mountains, where the highest points stand for
as the
darkest
the
gray tains
easily
see
erosion
The
(Fig. 1C). from the
a convex part of the original figure. Until now we have dealt only with reality we meet only gray tone images.
also
exterior
will two
is represented the
by
An
opening. opening
continuously
extinguished
is not
1C).
aptly called a skeleton. The fire middle of the figure with uniform points the advancing line of fire from
defines
circles
but
an
remain extracts
are
is a circle, is
(Fig.
the
boundary and the
propagation velocity 1E we retained only
de-
in mathe-
way
is
that
the velocity was higher figure is now disconnected,
in an intuitive
A”.
set to
X. The small particles passing through the sieve The big particles are cut into two parts: small through the sieve vanish and only the coarsest
quench
Along
in (1)
Leitz
more
fire the
(more
found
same shows
of the region
concentric
technique
several
of
figure
METhODS
matical
of A”
one region some other set
Ploem’s
use
eight
A and
stick-figure representation propagates toward the speed; however, at some
can easily be realized in power can be combined. can automated cytology
in SITS
preparation
a dilation
sets
we get
parts not passing through The next transformation
MATERIAL Cytological Stilbenisothiocyanate
by
it and
sieve of mesh are eliminated. parts passing
program.
The processing is made solely by iterative image transformations. The absence of measurement permits us to process objects another.
dilate
the original set amounts set. its original
points
tone and
image. peaks
and
the
The may
represent
contrast
discrimination be done with
peaks
between a topper
zones
smooth (hat).
in
mounIf it is
AUTOMATED
CYTOLOGY
AND
MATHEMATICAL
possible
A
top
to force
of
the
2. If we have
hat
detected
on the
mountain
in such
h2
a, by the
of contrast
when
there
h1 + h such as: threshold something
=
upper
details
Low
may
be found
Resolution
exist
that
two
is detected
h is burst); b, the lower threshold of size R (The peak enters into
More
a way
the
is
a peak
h1 and
height opening
hat
burst, then the mountain was a peak, it was not. This may be clearly understood in Figure come back to the gray tone image, it means that we
otherwise
olds
129
MORPHOLOGY
thresh-
(the
hat
of
disappears by an the hat of radius R).
in (10).
Detection
of Suspect
B
Cells
in Cervical
Smears
An
intensive
use
of the
top
hat
transformation
is outlined
here. Description Figures
of
3A
and
they two
appear successive
on
field
of 250
criteria see
light
3
Now
we
slide: normal
2
The
staining
With
some
have
seen
by
D
4A
a leukocyte,
an
and
shows The
may
thus
be
and
erosion.
B, Grassfire
D, skeleton.
and
dilation.
E, conditional
C,
they
are
represented no suspect
the the
or
dust
of on a
or
carcinoma
a
staining
with
reduced cells in a 4C will be
in situ
corresponding detection
cells
a narrow by
B detects leukocytes
eliminated.
Finally
carcinoma the results of
and
alarms
in Figures cells were
3A and present,
between
detected
On
hat
the
C,
only
The and
6A
and the
next
can dys-
section
preliminary
show fields.
that The
comparison cells
and
contrary,
in situ these
themselves
top
remain. in reality.
6B) permit intermediate
positive.
the carcinoma leukocytes. Since
peak
results: Figures 5A and been found in two different
obviously
procedure.
theoretically
cells
cannot detects
highly contrasted and particles coy-
upside-down and
in situ obtained
the
only
white an
peak also
stage
by an elimination
hat all
lution views (Figs. SB and that are found with normal
dilation:opening.
present
profiles of all these top hats of Figure
there remain of degenerative
detected
Description statistical alarm has
top-hat
types
in situ cell, cell covered
leukocytes
be followed
present be
events
of cells
of the
and narrower If we eliminate
by leukocytes, and the groups
they
the
various
a carcinoma a squamous
the three
because A. Since
it must
plastic and will present
bisectrix
in
of all other cells and artefacts. The hat A cells and the granulocytes, but ignores the cells are low-contrasted. It also ignores
artefact the hat
leukocytes
E
by
as
between a square
j
considered
that
different
(clump
recognition
The higher leukocytes.
“tools”
ofdegenerative
sufficient
granulocytes,
A, Grassfire
one
us to detect
of the
artefact
a group
for the
section
to the right, a granulocyte,
4B
the large enter into
followed
In fields
is Acriflavine-Feulgen imagination
enables
contrast. Figure relief representation.
ered cells
the
is a drawing
the setting aside detects the suspect normal intermediate
&
in
from the left intermediate,
artefact),
lA-I).
illustration:
microscope
at a resolution of 1 TV screen represents
(5).
precisely
peaks. Figure
FIG.
an
different
the darkness of the points as their altitude. Thus, the cells would appear as mountains of various heights and shapes. A normal intermediate cell would be a small hill, a granulocyte a group of three sharp peaks, and overlapping cells would appear as a chain of mountains. transformation
erosion
and
two
the TV screen TV lines. The
x 250
transmitted
C
the
B we
in
a positive high resoof the show the
B, where many leukocytes nothing has been detected.
cells that fields but
130
MEYER
FIG.
FIGS.
3A
and
2.
Microscope
B.
Top
hat
transformation.
Fields
at the
TV.
screen
(250iL
x 250i).
1’ #{149}1#{149}
Afl C
A
n FIG.
4.
Three
top hats
for the detection
of cells.
AUTOMATED
FIGS.
5A and
B.
A, a cell
CYTOLOGY
has
been
found
:
AND
(in white).
:
#{149}.
at a higher
resolution.
The
cell
is positive.
B, confirmation
at a higher
resolution.
The
cell
is positive.
L,
.
and
B, confirmation
,.
..
6A
131
MORPHOLOGY
.1
J
FIGS.
MATHEMATICAL
B.
A, a cell
.
has
been
found
(in
white).
4r.. 44’’
FIG.
7.
Detected
small
cell
carcinoma
cell
(magnified
view).
FIG.
8.
Detected
false
positive
two
overlapping
normal
nuclei.
132
MEYER
Most
detection
DNA
content,
and
cell carcinoma. of all cases. high
work
completely
on
the
size miss
of cells
or
case
of small
the
And yet, this form of disease represents Dr. Ploem and his group has shown that
DNA
missing
methods therefore
cells
them
exist is
even
not
in
these
negligible.
cases,
but
Fortunately
risk
of
method
of
detecting unenlarged but highly contrasted cells perfectly detects these cells of small cell carcinoma. This is illustrated on Figure 7 where one of these cells has been found near three normal intermediate cells. Let
us
consist as
now
speak
merely
such.
two
in
One
of
of
some
failures
artefacts
these
overlapping
the
of
which
our cannot
is represented
cases
normal
nuclei
method. be
They
recognized
in Figure
were
recognized
Conclusion: many
30% a few
the
our
on the
validity
will
be safer
would
slides of
be
the
if we an
first
must
results
be
method
can
add
seem
screened be
miss
matic
cells with high DNA content so typical would provide us with a more comfortable detected cells with regard to the false negative probably cause a slight increase in the number tives.
This
is
not
them. The next or how to separate Shape
suspect
serious
since
we
Recognition
cells. The
evaluation we
of the
method
treated
can
1000
only
fields
in
be done
each
of
statistically. 18 slides.
We
stopped before 1000 fields were viewed when, after a few hundred fields, a sufficient number of suspect cells were found. The slides chosen at random, with the only condition being to have a suitable staining. The results are shown in
ofArtefacts
Table
I.
The
machine
fields in available,
has
been
each slide. refocusing
programmed
to
go
through
1000
stops and asks if it has found a true I shows the results. The slides have
been
beforehand
screened
manually
According
normal;
“2”
is for
stage;
and
estimation We point
inflammation; “4”
means
slides, was
suspect
reasonably
“
1” means
“3”
for
dysplasia
carcinoma
cells
Slide
Diagnosis
18452 349
in
were
has
has is zero
found.
The
number
o f 18 Slides Investigated Number Cells
for
or Fal
se Positi
of
Nrlr
of s
Ic
True positive
Alarms 18
60,000
1,000
33
3b
10,000
1,000
17
10
1,000
54
41
0
86 0
3912
2
15,000
1,000
0
9
17484
3b
70,000
1,000
2
7
R 3512 R 2582
4
2,000
1,000
14
2
3a-b
150,000
1,000
6
22
R3l32
2
100,000
500
0
57
R2182
1
1,000
0
0
1,000
6
55
6,000
3a
100,000
2
30,000
0
8
R2592
5
3,000
200
39
2
R2662
4
6,000
41
2
R 2072
3a-b
600 500
31
51
R
3a
50,000
1,000
1
8
2
100,000
1,000
0
27
2082
within
is necessary
of a chromosome Since the chromo-
a mitosis,
before
is
a reorientation
visual
recognition.
This is not the case in cytology: the cytologist is able to recognize a cell from an artefact whatever their orientation or position within the field of view may be. He may even work at different magnifications. We would like our recognition
False Positive
Cells
29
R 3162
recognition is correct.
ye Cells
1,000
3772
of
I True
1,000
R
Separation
Cells
of false
3,000
17545
artefacts
made. in all
4
H
eliminate
A rough been since
2
197
random
to
or precansitu.
18312
R
at
how
been
3a-b
3b-4
located
chromosome
of dysplasia. number of rate. It would of false posi-
completely
60,000 100,000
18382
are
It
hypchro-
low.
TABLE of analysis
Results
R
a diagnosis
code,
of the number of cells measured out that the false negative rate
positive alarms
and
to a classical
cerous
of each
method.
large
to eliminate
and
results
Since an automatic focus is not yet was sometimes necessary. At each posi-
tive alarm, the machine or a false positive. Table given.
Choice of a method: The only possible if its orientation somes
know
chapter explains how clustering cells.
Clustering
Therefore
very
on
the
to our
the
but
assessment
Probably
criterion
to not
promising
any
niade.
a detection
advantage
very
before
This
8 where as
The
more
60,000
1,000
FIGS.
940,000
isolated
9 and
cells,
10. several
The conditional in artefacts.
hisectrix
marks
one
kernel
in
AUTOMATED
CYTOLOGY
AND
A
MATHEMATICAL
133
MORPHOLOGY
B
hA
FIGS.
and
B.
with
the
kernels
removed,
isolated
A
cells
have
one
hole;
artefacts
have
several.
B ‘S
I I
I
I I ‘S
...-..-
‘5,.
12A
FIGS.
and
B.
#{248},,
The
skeleton:
FIGS.
procedure
to have
the
the
clustering
This
constraint
is
on size
features
we
in the
Methods
constant
along
the In
objects Figures
are are
point
is higher
out
only
for
isolated
cells.
Triple
points
A and
B in artefacts.
B
very
than
strong
an and
In particular, transformations
section
that
skeleton
from
the and
the
FIGS.
whatever
in
Therefore for
image.
the
should
from
prohibited.
and not 9 and
on these
or
tools. all
the
Initial
flexibility:
program
looking
represented in gray. called ro-conditional fire
B.
objects
cells
the range of possible on the magnification, based
same
of the used,
from
and
13A
the
orientation
magnification cell
loop
B
A
or
one
the
field,
position
whatever
recognize
the
the
pictures
that
considerably
since
Dilation.
A
we
restrains
the size depends or measurements had
skeleton.
speed
of the only
minimal
B.
an isolated
to derive We
grassfire on
the
have
the
seen
is not shape
of
on their size position or orientation. 10 various shapes with various sizes
six times
and
artefact.
depends
White represents bisectrix, along
14A
speed
of the
cells
are
fire.
iSA
and
B.
Skeleton
of
the
dilated
part;
triple
points
for
are
the part of the skeleton which the speed of the
all isolated
FIGs.
artefacts.
We
marked
by
one
by
several
How cessing:
white
kernel kernels.
inside, That
skeletonization We
now
while
will
all
be our
permits have
two
other
shapes
are
marked
criterion. further
categories
parallel of
objects:
pro1)
cell
134
MEYER
characterized We
by
could
after particle,
continue
another, and
one with
counting
kernel;
2) artefacts
our
procedure
the
number
eliminating
all
by
several
by taking of
particles
kernels.
one
particle
parallel.
one
kernels
inside
each
The
containing
more
than
kerneLs
FIGs.
17A
and
But
then
set difference
our
processing
between
produces
one
hole
the in
16A,
B.
B and
A, cells
C.
Two
skeleton
in white;
artefacts
lines
for
in grey.
18A
and
B.
A, a cluster
of cancerous
cells.
isolated
C
each
object.
B, two
Triple
skeleton
points
lines
for
generated
artefacts.
by the
A
FIGS.
flO
original
the
B
A
FIGS.
kernel.
B, the
cells
are
automatically
separated.
program.
longer
image cells
would
and and
the
he
white several
AUTOMATED
holes
in
clustering
Figures
1 la
cells
and
changed; the holes. If we
b.
cells now
preserves
the
The
observation
recognition
can
easily
be achieved
cells
skeleton
of
the
or
more
having
no
points
of the
image, reduction
points,
have
remain.
vanished,
border. The too irregular. 13b
original zation
an
we
take
the
border For
(Figs.
iSa
leads
all
difference
and
b).
Each
to a small
of triple
points
appendices important
artefacts,
characterizing by
between
the
the
shorter irregularities
than
of
previous only by an
dilated
are
to
be
Note
can
the
procedure
points
be
and done
in
by a Golay
(in white)
could
step
in
and
the
if we
set
detection
being
and
the can
by the
(Fig.
of the
sets
process, 16c).
And
the
consecutive how
the
Finally, sures
Figure in
their
Each
the
17b
shows each
skeletons
object,
the
two
particle. are
which
different convex parts. the resulting skeleton. and works effectively
skeleton
The
triple
eliminaisolated
overlapping
isolated
is not
lines
particles
cells
cells
cells,
“round”
generated
without the
others
enough
and
by the
triple are
artefacts.
is segmented
18a and b. Figure l8a In Figure l8b these
or in
distribution.
and
we could
artefacts
and
we required
to the
position
to the magnification only a few tools may
insensitive
to the
experimental
reliability
base
of
used. be used.
fluctuations procedure
of
the
en-
results.
ACKNOWLEDGMENTS
ings
of
mated 2.
the
report
represents cells have
of the Engineering
VIth
Cytology,
Digabel
in
CITED Leiden Automation group, Foundation Conference
J Histochem
H, Lantuejoul
Ch:
Cytochem, Iterative Vol.
Pattern
Proceedon auto-
in press
algorithms,
8, Riederer
Special Verlag
Classification
Wiley, New York, 1973 4. Golay MJE: Hexagonal Pattern Vol. C 18 N#{176} 8, August 1969
Issues
Gmbh,
and
Scene
Analysis,
lEE
Trans
on Comp,
Transforms,
of
Stutt-
5.
Cornelisse CJ, Ploem JS: A new type of two-color fluorescence staining for cytology specimens. J Histochem Cytochem 24: 72, 1976 6. Klein JC, Serra J: The texture analyzer, J Micros 95:349, 1972 7. Matheron G: Elements pour une th#{233}oriedes milieux poreux, Masson et Cie, Paris, 1967 8. Matheron G: Random Sets and Integral Geometry, Wiley, New York, 1975
9. Meyer
F: Automated
10.
Meyer
Paris,
F: Contrast
Metallography,
points
Therefore the triple points appear in But the same procedure can be used for the separation of clustering cells.
This is illustrated in Figures a cluster of cancerous cells.
value
fluctuations
or in the
reproducibility
Symposium
for
and
of tools,
object
of different
illumination
gray
is insensitive
the field constraints
choice
in the
insensitive
be
These results could not have been obtained without the friendly collaboration of Mrs A. van Driel. She adjusted the parameters of the contrast program for an optimal detection of suspect cells and gave her diagnosis on all the alarms described in Table I.
artefacts. program
the
inherent
such
algorithms. of cells which
within strong
a method
in the the
about
When one inherent
where only nuclei were detected, Since the contrast of an object
under
Practical Metallography, gart, 1977 3. Duda Ro, Hart E:
leads two
these
a parallel
particle.
shows
stable
procedure
the objects Under these
all
analysis 16b). These
detection
from
almost
remarks
may
variation
program on contrast For the shape recognition
1. Al I: Progress
eliminate
Since
cell
crossing
we detect only In this way we
particle.
each
17a
be recognized
original
lines,
smoothness B in Fig.
parallel
Figure
remains
a few
to chose
changes
at a gray value is now apparent.
the
we wish to characterize. and illustration: To sumto one or more holes inside
each
for
transform
artefacts.
the
of each
procedure. the instabilities
fluctuations
a slight
completely
For example the cytoplasm
and
These
example
staining
to make
LITERATURE
that
skeletonization
one
ends
of the
for
For
like
recognition to analyze
analyzes
and
set
and
a given length of the borders.
skeletonized
disconnected,
this
of
chromatin
border
dilation
on these
each particle (set A in Fig. 16b), the to a ring around each particle (set
the
one
fluctuations.
types.
the
these
a small
irregularity
appendix
again.
objects
the
our
analysis
point”. of skeletonization
appears
to
we would
of a pattern study one has
a recognition
For
choose the degree of smoothness Combination ofboth criteria marize, the angular analysis leads
tion
the
particles
border
lateral
main line at a “triple Thus the same procedure
are
inside
a few binary
of a cell is always smooth or at example, Figure 13a represents a
artefact.
dilate
set
B).
one. We thus get Figures 14a and b. The skeletoniof these new sets leads to a closed line for each particle
now
sets
only
a complex
Most artefacts are detected by those that are not are characterized
Figure
of the
A and
and
with
the robustness starts a new in the
loop,
lines
points
criterion-smoothness
irregularities,
because one
three
the
Starting
ending with a few points of information is important.
irregular least not
the
12B,
completely
artefacts
the borders: method but
we
where
Figure
the
Conversely,
than
The
CONCLUSION
quickly detected by a simple Golay the treatment is completed, the cells,
A complementary
cell,
points,
(see
These points can be transform. Thus, after
more
segmented. separately.
As a conclusion
now
line.
been automatically can then be analyzed
b.
in parallel
having
triple
crossing
in
135
MORPHOLOGY
several which
that
closed
MATHEMATICAL
particles
12a and
shows
is a simple
artefacts,
our
have algorithm
Figures
pictures
the
are
we obtain
AND
represented of
artefacts
these
of the two
as
of
skeleton
presents
artefacts, connectivity
one hole while a skeletonization
connectivity, of cells
in
the
have apply
the
skeleton
or
Thus
CYTOLOGY
Cytology,
Proceedings,
20-24 June, 1977 features extraction,
Vol. 8. Riederer
Buffon
Special
Verlag
Gmbh,
bi-Centenary
Issues
Stuttgart,
of Practical
1977
1 1. Preston K: Feature extraction by Golay hexagonal pattern transforms, IEEE Trans on Comp, Vol C20, N#{176} 2, Sept, 1971 12. Serra J: One, two, three . . . infinity: Special Issues of Practical Metallography, Vol. 8, Riederer Verlag Gmbh, Stuttgart, 1977
13. Serra Press, 14.
Vrolijk cytology,
ence
J: Image London, J:
Analysis in press
A real Presented
on Automated
time
by Mathematical
TV at the
image analysis VIth Engineering
Cytology.
J Histochem
Morphology, system for Foundation
Cytochem,
Academic automated Confer-
in press