Professional Documents
Culture Documents
223877.
Corresponding author
Email addresses: yh2475@columbia.edu (Yrj o H ame),
mika.pollari@tkk.fi (Mika Pollari)
URL: http://users.tkk.fi/yhame/ (Yrj o H ame)
1
Present address: Dept. of Biomedical Engineering, Columbia
University, New York, NY, USA
ve cancers causing the most deaths worldwide, and
metastatic lesions are also common in the liver (Fried-
man et al., 2003).
Contrast-enhanced computed tomography (CECT) is
most commonly used for liver lesion evaluation and
staging after initial ultrasound imaging (Hann et al.,
2000). The imaging is commonly performed in two
or three phases that correspond to the dierent times
at which the contrast agent arrives to the liver through
the dual blood supply of the organ (Baron, 1994). In
addition, native computed tomography (CT) imaging is
commonly performed.
The correct timing of the CECT imaging phases is
dicult due to variability of patients, making the im-
age data often sub-optimal. Typical CT data also has
a relatively high level of noise, and as the contrast be-
tween the tumor and parenchyma is often low, the tumor
may be dicult to detect, and even more so to reliably
delineate. In addition to the limitations of the imaging
method, liver tumor segmentation is also complicated
by tumor variability in size and structure, and they may
appear practically anywhere within the organ.
State-of-the-art segmentation methods oer reduc-
tions in the amount of required user interaction, repeat-
able results and accuracy comparable with manual seg-
mentations. For the overall treatment process, these
traits reduce expenses and increase the process relia-
bility. Segmentation methods that require only mini-
mal initial user interaction, i.e. semi-automatic meth-
ods, have been the recent focus of research. They have
proved to be able to provide reliable results with ac-
curacy similar to interactive methods (Deng and Du,
2008). Fully automatic methods generally suer from
lower accuracy and robustness, as well as a signicantly
higher computational cost.
The semi-automatic method by Smeets et al. (2009)
is based on a level set method tted on a fuzzy classi-
cation of the image data. The method performed well
in the 3D Liver Tumor Segmentation Challenge 2008
(LTS08) (Deng and Du, 2008) but its accuracy declines
if the tumor has a low-contrasted edge. In addition,
since the classication assumes normal distributions for
the classes, it does not perform so well if the tumor is
adjacent to other structures than healthy liver tissue.
Another semi-automatic method by Moltz et al.
(2008) estimates typical tumor and parenchyma intensi-
ties based on input from the user, and denes thresholds
for region growing based on these estimates. The result
is post-processed with morphological operations. The
method also performed well with the LTS08 data, but it
encounters diculties with tumors that have inhomoge-
neous intensity distributions.
A more general approach to lesion segmentation pre-
sented by Jolly and Grady (2008) also estimates the in-
tensity distribution based on user-given points, and the
segmentation is based on a fuzzy connectedness algo-
rithm that nds a cost value for every image point. The
method is able to segment various kinds of tumors as
proved by an extensive evaluation. However, the seg-
mentation accuracy leaves room for improvement and it
does not perform well with heterogeneous tumors.
Other related work includes semi-automatic method
by Li et al. (2006), where tumor boundaries are lo-
cated with a machine learning-based classier, and the
liver structure segmentation method by (Freiman et al.,
2008), which uses a multi-class Bayesian classier and
morphological operations for adjustments. In addition,
a recent publication includes a benchmark study of three
semi-automatic methods (Zhou et al., 2010).
Tumors with low contrast are challenging for these
tumor segmentation methods, especially if the image
has a high level of noise. Ambiguous borders cause dif-
culties in particular for boundary-based segmentation
methods. This has created a need for a method that is
able to perform reliably and accurately with these char-
acteristics present, without increasing the amount of re-
quired user interaction.
The target application of this work was radiofre-
quency ablation (RFA) treatment planning, where reli-
able tumor segmentations are needed for accurate nee-
dle placement. Tumors treated with RFA are typically
relatively small in size, with diameters of less than 5 cm,
and they have a generally spherical shape (Gazelle et al.,
2000). A spherical shape is typical also for metastatic
lesions in the liver (Halvorsen et al., 1982) and for single
nodular hepatocellular carcinoma (Kanai et al., 1987).
A novel semi-automatic method was developed for
segmenting liver tumors from low-quality CT data. The
method is based on non-parametric intensity distribu-
tion estimation and the hidden Markov measure eld
(HMMF) model (Marroquin et al., 2003). The HMMF
model adds a continuous-valued measure eld estima-
tion step to the classical Markov Random Field (MRF).
The application of the measure eld provides a smooth
cost function that is simple and ecient to optimize,
improving the MRF by removing diculties with lo-
cal minima and oscillating behavior. Also, the measure
eld captures the classication uncertainty by reducing
the weight for points with uncertain classications in
the eld cliques.
The method assumes a roughly spherical shape for
tumors, and that in general, the intensity distributions of
tumors and the adjacent tissue do not necessarily follow
any particular statistical distribution. Amultivolume ap-
proach is also presented for using all the available image
data.
The developed method was evaluated using two sets
of patient data, the publicly available data set of LTS08
and a data set that consisted of pre-operative images of
patients treated with RFA. Also, a novel framework of
creating articial data with ground truth segmentations
was developed. The articial data is used for analyzing
performance with dierent levels of contrast.
Following this introduction, the developed method
and the data used for training and evaluation are de-
scribed in Section 2. The evaluation results are reported
in Section 3, and Section 4 concludes the paper with a
discussion.
2
2. Methods
2.1. Segmentation task formulation
This general formulation follows the outline of the
original HMMF model (Marroquin et al., 2003). Some
signicant modications have also been introduced,
most importantly in the probability distribution P(q) of
the measure eld q, and in the non-parametric intensity
distribution estimates that are kept static in the segmen-
tation process.
Let the observed image be I. The segmentation task
consists of nding the label eld f that maximizes the
posterior probability P( f |I). Let represent the image
domain, with r representing image points (voxels).
Then f (r) Z
M
= {1, ..., M}, where M is the number
of dierent classes for the segmentation task. For the
purposes of liver tumor segmentation, M = 2 (see dis-
cussion in Section 2.2).
The label eld f is found in two steps. The rst step
consists of generating a Markov random vector eld q
with distribution
P(q) =
Q(q)
K
exp
_
_
S
D
(q)
C
W
C
(q)
_
_
, (1)
where Q(q) is a class-dependent prior probability func-
tion, K is a positive normalizing constant, S
D
(q) is a
shape prior dependent on the input data D, C are the
cliques of a given neighborhood system, and W
C
are
potential functions. The used shape prior S
D
is simi-
lar to the approach introduced in Flach and Schlesinger
(2008). The M-dimensional vector q(r) is also con-
strained by
M
k=1
q
k
(r) = 1, q
k
0, (2)
where q
k
(r) is the kth component of q(r). Here the con-
straint becomes q
1
(r) + q
2
(r) = 1; q
1
, q
2
0.
The label eld f is generated from q in the second
step, each f (r) being an independent sample from the
distribution q(r), with:
P( f |q) =
_
r
q
f (r)
(r),
where the component q
f (r)
(r) of vector q(r) corresponds
to class f (r).
For nding the optimal estimator q
= arg max
q
P(q|I),
with the constraint (2) applied. Using the Bayes rule,
the posterior distribution P(q|I) is dened as:
P(q|I) =
1
R
P(I|q)P(q), (3)
where R is a positive normalizing constant. The con-
ditional distribution is dened as (see Marroquin et al.
(2003) for proof):
P(I|q) =
_
r
M
k=1
P(I(r)| f (r) = k)q
k
(r). (4)
For brevity, we denote the observation likelihood func-
tions as P(I(r)| f (r) = k) = v
k
(r) and P(I(r)| f (r)) = v(r).
The sum term in (4) can then be expressed as v(r) q(r),
or v
1
(r)q
1
(r) + v
2
(r)q
2
(r) for M = 2.
Combining (1), (3), and (4) results in
P(q|I) =
1
KR
exp
_
U(q)
_
,
where
U(q) =
r
log(v(r) q(r)) log(Q(q))
S
D
(q) +
C
W
C
(q). (5)
As 1/KR > 0, q
, the optimal
estimator f
(r):
f
k
(r). (6)
With M = 2, this is equal to f
(r) = max(q
1
(r), q
2
(r)).
2.2. Method overview
The segmentation is performed in four stages:
1. Preprocessing and user input
2. Estimation of observation likelihood functions
3. HMMF segmentation
4. Post-processing
The rst stage involves input from the user, after which
all the subsequent stages are performed automatically.
Here, the stages are described briey, with a more
thorough presentation in the following sections. The
process is illustrated in Fig. 1.
Given an image or multiple images of a patient as in-
put to the method, the user selects two points indicating
3
Figure 1: Stages of the segmentation method
the location of the tumor. The method then performs the
preprocessing steps based on the user input.
The second stage uses the image data and regions de-
ned in the previous stage to estimate intensity distri-
butions for the segmentation classes. Using these esti-
mates, each image point r is then assigned a likelihood
value v
k
(r), indicating how probable such an observa-
tion I(r) is for each class k. The result is passed on to
the third stage of the method.
The third stage performs the actual segmentation us-
ing the formulation presented in Section 2.1.
The nal post-processing stage modies the segmen-
tation objects by removing overown sections. This is
done by comparing the tumor object shape with the cen-
ter of the tumor as dened by user input. The output of
the post-processing stage is the nal segmentation.
A segmentation in two classes was selected in the im-
plementation, since the number of actual tissue types
around the tumor is unknown in general. Two classes
simply indicate whether a point is part of the tumor or
not. The used nonparametric intensity distribution esti-
mates are very useful in cases with several tissue types
around the tumor.
In the following, k = 1 represents the class corre-
sponding to the tumor. The stages of the method are
illustrated with an example in Fig. 2.
All of the parameter values presented here were se-
lected based on results from training data and used for
all evaluation data. The used training data was provided
by the LTS08 competition.
2.3. Preprocessing and user input
From image I, the user selects the axial slice view
where the tumor appears the largest. Then, the user se-
lects two points on opposite edges of the tumor, so that
if a line was drawn between the points, it would pass ap-
proximately through the center of the tumor as observed
in the slice view. Let the selected points be l
1
and l
2
.
The next step is to construct a ROI
R
. The
segmentation is performed only for points r
R
. In
addition, training samples T
k
R
need to be deter-
mined for each class k for estimating the observation
likelihood functions in the next step.
Using the user-dened points, the following variables
are determined:
1. ROI center r
c
=
1
2
(l
1
+ l
2
)
2. tumor radius d
T
=
1
2
|l
1
l
2
|
3. ROI radius d
ROI
= max(1.5d
T
, d
min
)
The ROI radius is limited to at least d
min
= 8 mm, to pre-
vent it from becoming too small for very small tumors.
In addition, the width of the ROI edge d
e
is assigned a
value of 1.5 mm.
To simplify notation, let x = {x
1
, x
2
, x
3
} represent co-
ordinate points with origin at r
c
, so that x = r r
c
. The
terms x
1
, x
2
, and x
3
are coordinates along the sagittal,
coronal, and axial axis, respectively. Then
R
is dened
as a sphere, with center at r
c
and radius d
ROI
:
R
=
_
x | x
2
1
+ x
2
2
+ x
2
3
d
2
ROI
.
_
The training samples T
1
are determined as the set of
points within an ellipsoidal area centered at r
c
:
T
1
=
_
_
x
R
|
x
2
1
a
2
1
+
x
2
2
a
2
2
+
x
2
3
a
2
2
1
_
_
,
where the coordinates x
1
and x
2
are obtained by rotating
x
1
and x
2
around the x
3
axis with center of rotation at
r
c
, so that the long axis of the ellipsoid passes through
the input points l
1
and l
2
. The variables a
1
and a
2
are
assigned values depending on the tumor radius: a
1
=
2d
T
and a
2
= 0.8d
T
.
The training samples T
2
are determined as the set of
points at the ROI edge as dened by d
e
:
T
2
= {x
R
| d(x, r
c
) d
ROI
d
e
} ,
4
(a) (b) (c)
(d) (e) (f)
Figure 2: Main stages of the method illustrated with an example: a) User input and ROI construction, with the following markers: outer ring for
ROI border, ellipsoid for sampling area of tumor training data, x-markers for input points, small circle for r
c
, b) observation likelihood function for
class 1, c) observation likelihood function for class 2, d) measure eld MAP estimate q
i=1
K
_
I(r) T
k
(t
i
)
h
_
, (7)
where h is a free parameter controlling the smoothness
of the estimate, and K is a Gaussian kernel:
K(x) =
1
2
2
exp
_
(x )
2
2
2
_
.
5
The variables
2
and are the variance and the mean of
the kernel, respectively. Here we use values
2
= 1 and
= 0, so (7) can be written as
v
k
(r) =
1
hn
k
2
n
k
i=1
exp
_
(I(r) T
k
(t
i
))
2
2h
2
_
.
An empirically chosen value h = 2.2 is used here.
The above likelihood estimation is used only for
points having intensity values higher than 0, since lower
values do not usually occur in tumors in CT images
(Dowsett et al., 1998). Points r with intensities of 0
of less, i.e. I(r) 0 are assigned likelihood values of
v
1
(r) = 0 (tumor class) and v
2
(r) = v
m
, where v
m
is the
highest likelihood value assigned to any point for class
2.
2.5. Segmentation with HMMF model
To nd the MAP estimate q
k=1
k
q
k
(r),
where the weights
k
control the prior probability for
each class k. The used values were
1
= 1 and
2
=
1.08 0.55, where is a measure of separation of the
two intensity distribution estimates:
=
1
2
_
r
| v
1
(r) v
2
(r) | . (8)
The value of varies from 0 for identical distributions,
to 1 for completely separated distributions.
Next, the shape prior S
D
is dened. It takes advan-
tage of the location information provided by the user,
essentially indicating that an area around the center r
c
should be segmented as the tumor and the region at the
ROI edge should not be a part of the tumor. The shape
prior is dened as:
S
D
(q) =
s
r
R
kM
s
k
(r, D)q
k
(r),
Figure 3: Example of shape prior s
1
, with input points shown with
x-markers
where
s
= 3.0 is a weighting constant. The function s
for class k = 1 is dened inside the ROI as:
s
1
(r, D) =
_
_
1 (1 + a
D
(r))
1
, if d(r, r
c
) < d
ROI
d
e
1, otherwise
with
a
D
(r) = exp
_
_
d(r, r
c
)
d
T
d
s
__
,
where = 20 controls the slope of the function and d
s
is a modier parameter controlling the size of the center
region with respect to tumor radius d
T
, chosen here as
d
s
= 0.55. For class 2, s
2
(r) = s
1
(r).
The function s takes the form of a logistic function
scaled with the tumor radius d
T
. For class k = 1, s
is close to a value of one at the ROI center r
c
, and ap-
proaches zero when the distance to r
c
is larger than d
T
d
s
.
At the ROI edge, it is given a value of 1. An example
of the shape prior is shown in Fig. 3.
The potential functions W
C
enforce the smoothness
of q. Here, pairwise cliques C in the 26-neighborhood
of each point r are used to compute the potentials. The
potential functions are dened as:
W
r
1
r
2
=
W
exp
_
d(r
1
, r
2
)
2
/(2
2
W
)
_
M
k=1
(q
k
(r
1
)q
k
(r
2
))
2
,
where d(r
1
, r
2
) is the Euclidian distance between the two
points,
W
= 1.5 is the standard deviation of the expo-
nential term used to modify the weights of the neighbor-
ing points and
W
= 20 is a weighting constant.
The function U is minimized using the gradient de-
scent optimization method. After this, the label eld f
is found as described in (6).
2.6. Post-processing
Structures adjacent to the tumor with similar intensity
distributions may cause the segmentation to overow
6
outside the tumor. This often happens through a narrow
passage, and results in a segmentation object that has a
handle attached to the spherical tumor mass. These
handles are removed in the post-processing stage, along
with any objects that were classied as tumor, but not
attached to the actual tumor object.
The handle removal is based on comparing two dif-
ferent distance values of points from the ROI center r
c
.
The rst value is the Euclidian distance d
E
(r, r
c
), and the
second one is a weighted distance value d
w
(r, r
c
), which
approximates the distance to be traveled inside the seg-
mentation object to connect the two points. After ob-
taining the two values, the dierence is observed. If the
dierence is large, it can be deduced that r belongs to a
handle, since there is no direct path from r
c
to r when
advancing inside the segmentation object.
The implementation of this is done using the Fast
Marching Method (Sethian, 1996), so that a front with a
spatially varying speed function F(r) is advanced start-
ing from the center point r
c
. The passing time t
r
of the
front at each point r is assigned as its distance value
d(r, r
c
).
To do this, the Eikonal equation is solved:
F(r)|d(r, r
c
)| = 1, where d(r, r
c
) is the gradient of the
distance function. To compute the Euclidian distance,
F(r) = 1, r
R
. For computing the weighted dis-
tance, F(r) = 1 if f (r) = 1 and F(r) = 0.1 otherwise. In
the case of the weighted distance, the advance of the
front is signicantly slower outside the segmentation
object. Using the distance values, a probability volume
is then computed: (r) = exp[(d
E
(r, r
c
) d
W
(r, r
c
))
2
]
if f (r) = 1, and (r) = 0 otherwise. The nal seg-
mentation f
(r) = 1, otherwise f
(r) = 0.
The operations of the post-processing stage are illus-
trated in Fig. 4 using an articial example. The gure
shows that the handles are removed in the process, while
leaving the spherical object and the protrusions in the
bottom part intact.
Fig. 5 illustrates the post-processing stage using a
real CT volume example. The gure shows a large
tumor, where the segmentation has slightly overown.
The overown region and a disconnected object are
clearly seen in the 3D object of Fig. 5(b). The post-
processing operation removes the handle and the dis-
connected object while keeping the tumor segmentation
intact.
2.7. Multiphase segmentation
To incorporate information from multiple CECT
phase images, all of the available images were regis-
tered to a common coordinate system. The registration
was performed with an in-house registration program,
which was similar to the IRTK software (Rueckert et al.,
1999) with a few modications.
The native CT image of each patient was selected as
a common target and all CECT images of the same pa-
tient were registered pair-wise with the target. The po-
sitional dierences were corrected with a rigid transfor-
mation model and local deformations with a B-spline
model (Rueckert et al., 1999; Rohlng et al., 2003).
During the non-rigid registration, the control point dis-
tance was hierarchically rened from isotropic 40 mm
to 20 mm. These values were chosen based on a previ-
ous study (Rohlng et al., 2004), where similar transfor-
mation models were successfully used to estimate res-
piratory motion in the liver.
As the energy function, the inverse of normalized mu-
tual information (Studholme et al., 1999) was used:
E =
H(I
S
) + H(I
T
)
H(I
S
, I
T
)
,
where H(I) is the marginal entropy and H(I
S
, I
T
)
the joint entropy for source I
S
and target I
T
images.
For non-rigid registration, smoothness (Wahba, 1990;
Rueckert et al., 1999) and incompressibility (Rohlng
et al., 2003) constraints were tested but not used, since
they did not result in any increase of registration accu-
racy.
The energy in rigid registration was minimized with
the downhill simplex method (Nelder and Mead, 1965),
and in the non-rigid case with the conjugate gradient
method (Press, 2007). For both optimization methods
an implementation from Press (2007) was used. To
speed up the computation, two-level multiresolution op-
timization was used. We used isotropic voxel dimen-
sions of 1.0 and 2.0 mm
3
for the ne and coarse reso-
lution levels, respectively. After registration, the CECT
images were transferred and resampled to the target do-
main with the computed transformations. Resampling
was performed with trilinear interpolation.
The segmentation process mostly remains the same
as in the single-volume case, with the only dierence
being how the observation likelihood functions are esti-
mated. The user can use any image for providing input,
but this should be done using the one with the highest tu-
mor contrast. The subsequent steps in the preprocessing
and user input stage remain the same as in the single-
volume case.
The estimation of observation likelihood functions is
initially done separately for each image. Then, a sepa-
ration measure of the intensity distribution estimates (8)
for each image is computed. The separation measure is
used as a weighting factor for a joint estimate. With this
7
(a) (b) (c) (d)
Figure 4: Post-processing operation illustrated with an articial example, where tumor center r
c
indicated with x-marker: a) spherical binary
object with protrusions at the bottom and an overown section at the top, b) weighted distance d
W
(r, r
c
), c) probability volume (r), and d) nal
segmentation f