Two-stage Semi-automatic Organ Segmentation

Framework using Radial Basis Functions and

Level Sets

Andreas Wimmer1,2 , Grzegorz Soza2 , and Joachim Hornegger1

1
Institute of Pattern Recognition, University of Erlangen-Nuremberg, Erlangen,
Germany
[email protected]
2
Siemens Medical Solutions, Forchheim, Germany

Abstract. The automatic segmentation of complex anatomical struc-

tures often fails due to low-contrast or missing edges, pathologic alter-
ations, or high noise. As an alternative, we propose a novel two-stage
semi-automatic algorithm that is able to segment complex structures
like the liver shape with moderate user interaction. The first stage of
our algorithm is the manual delineation of cross-sections of the anatom-
ical structure in 2-D multi-planar reconstruction views. From this set of
contours, an initial 3-D surface is reconstructed using radial basis func-
tions. In a second step, the surface is evolved using a level set algorithm
incorporating a new combination of both image information and shape
information, the latter being derived from the initial contours. The al-
gorithm has been evaluated for 10 Computed Tomography scans of the
liver and has shown promising results.

1 Introduction
The automatic segmentation of complex anatomical structures is still a challeng-
ing task. Algorithms relying solely on image intensities or features derived thereof
often fail in case of low-contrast or missing edges, as shown in Fig. 1(b). In addi-
tion, pathologic structures and high noise, for example due to reduced radiation
dose in Computed Tomography (CT), can introduce false edges and irregular
feature statistics, hindering edge- or region-based approaches, respectively. In
order to cope with missing or ambiguous low-level information, high-level infor-
mation has been successfully employed, for example in the form of shape priors
[1]. This, however, requires modeling anatomical characteristics and demands
for a training phase. Considering complex structures with high variations, for
example the shape of the liver, it may be difficult or even impossible to capture
the set of admissible instances, as illustrated in Fig. 1(a) and 1(c).
A segmentation algorithm incorporating shape information is presented in [2]
as part of a framework for the automatic anatomical, pathological, and functional
segmentation of the liver in CT scans. The anatomical segmentation involves
aligning a liver template followed by the deformation of a shape constrained
deformable model using global and local forces.

T. Heimann, M. Styner, B. van Ginneken (Eds.): 3D Segmentation in The Clinic:

A Grand Challenge, pp. 179-188, 2007.
 (a) (b) (c)

Fig. 1. (a) Sample CT slice with low-contrast boundary between liver and muscle
tissue. (b) Enlarged view of the region marked in (a), segmentation obtained for edge-
based Geodesic Active Contours [7] expanding from within the liver. (c) Slice containing
the largest extent of another liver, illustrating the high variation in liver shape.

Shape information is also employed in [3], where an Active Shape Model

(ASM) approach for the fully automated 3-D segmentation of the liver in CT
scans is evaluated for a large number of datasets and different gray-level appear-
ance models guiding the deformation. In order to cope with the huge variety of
liver shapes, restrictions of the classical ASM are loosened by incorporating a
spring-model deformable mesh.
In [4], a semi-automatic segmentation method for 3-D objects in medical im-
ages is presented. The proposed method consists in manually delineating several
cross-sections of the anatomical structure in parallel slices using the live wire
method [5]. For intermediate slices, contours are determined by shape-based in-
terpolation of contours in slices above and below.
The Random Walker algorithm [6] is a popular semi-automatic segmentation
method. The user places labels for each class to be segmented in the volume (or
image) in a fashion similar to a painting application. The class of each unlabeled
voxel is then determined as the class which has the highest probability of being
reached by a random walker starting at the respective position. The probability
for the random walker moving from one voxel to another is based on the intensity
difference between both voxels.
The lack of knowledge about the location and orientation of the structure, for
example due to diseases or different acquisition procedures, poses another major
challenge for any segmentation algorithm. Starting from a deficient position,
most algorithms will either fail completely, stop at a poor local minimum of
their target function, or at best require a large computational effort.
In order to address these problems of varying shape, position, and orien-
tation, we propose a novel two-stage semi-automatic algorithm. Our algorithm
starts with the manual delineation of cross-sections of the anatomical structure
in 2-D multi-planar reconstruction (MPR) views. This is advantageous compared
to the delineation in parallel slices as in [4], where problems arise when the or-
gan surface is almost coplanar to the slice. From this set of contours, an initial

180
3-D surface is reconstructed using radial basis functions (RBFs). According to
our experience, seven manually drawn contours already are sufficient in order to
obtain a good approximation of the complex shape of a liver. In a second stage,
this surface is evolved using a level set [8] algorithm incorporating a new combi-
nation of both image information and shape information. The shape information
is derived from the initial contours and thus tailored to the specific dataset. The
inclusion of image information frees the user from having to perform numerous
manual delineations. Our paradigm is to specify shape information where im-
age information is ambiguous. In other regions, image information prevails. The
algorithm is summarized as follows:
1. Manually delineate cross-sections of the anatomical structure in 2-D MPR
views
2. Reconstruct the 3-D surface defined by the manual delineations through
radial basis functions
3. Evolve the reconstructed surface under both image and shape information
using the level set framework
In Sect. 2, the manual delineation of the anatomical structure and the recon-
struction of the initial 3-D surface are presented. Section 3 details the subsequent
surface evolution using both image and shape information. Results for the seg-
mentation of the liver shape in CT scans are presented in Sect. 4, followed by a
discussion in Sect. 5.

2 Manual Delineation and Surface Reconstruction

The first step in our semi-automatic segmentation algorithm is the manual delin-
eation of cross-sections of the anatomical structure in 2-D MPR views. We have
chosen to fit a 2-D cubic interpolating spline [9] through control points placed
by the user. Smooth organ shapes usually require the placement of only a couple
of control points to capture the cross-section, as can be seen in Fig. 2(a). The
user can insert, move, and delete control points. At any time, the current spline
curve is shown to the user, providing immediate visual feedback about changes.
The contours are defined in three orthogonal views. Complex structures like the
liver shape typically require the definition of six to eight contours in total. In
Fig. 2(b), six contours are defined, two for each view direction.
Other approaches like the live wire method [5] could easily be integrated
into our algorithm, too. Compared to the Random Walker algorithm [6], the
definition of accurate spline curves may seem inconvenient and costly. However,
as noted above, smooth organ shapes usually require the placement of only a
few spline control points, and for low contrast boundaries, the Random Walker
may require an exact placement of labels, too.
After the contours have been defined, radial basis functions are used to in-
terpolate an initial surface. It has been demonstrated in [10, 11] that radial basis
functions deliver high quality smooth surface reconstructions even for sparse
datasets. The surface to be reconstructed is implicitly represented as the zero

181
level set of a signed distance function d. Thereby, no parameterization of the
surface is required and complex topologies can be handled easily. In order to
define d and thus the surface, all contours are sampled at N equidistant points
in total and the points are transformed to world coordinates. These points are
assigned a distance of 0. In addition, the planar normal for each point is calcu-
lated, too. These normals can be analytically determined from the spline curves.
At the intersection 1 of two contours, both planar normals can be extended to
3-D using the non-planar component of the respective other normal. For the
N points sampled from the contours, the missing component of the normal is
then found by linear interpolation between the respective component of the two
closest intersections, as illustrated in Fig. 2(c). The linear interpolation is per-
formed according to the arc length distances along the respective contour . Given
the N points in world coordinates and their 3-D normals, 2N additional con-
straints can be defined for the signed distance function d by moving each point
one length unit in its positive and negative normal direction and assigning that
point a signed distance value of +1 or −1, respectively. This sparse signed dis-
tance function containing all 3N points denoted with pi is interpolated using
radial basis functions according to

3N
X
d (x) = q (x) + ϕ (|x − pi |) λi , (1)
i=0

where q is a polynomial of degree 1 and ”|.|” denotes the Euclidean norm. We use
the biharmonic kernel ϕ (t) = t, which ensures a smooth surface reconstruction
[11]. The fitting of polynomial coefficients and the estimation of interpolation
coefficients is described in [11]. In order to keep the fitting and evaluation time
small, N is chosen in the range 300 < N < 500. In our experiments, we found
this usually sufficient to capture most if not all of the detail the user has specified
when drawing the contours. Furthermore, it is not necessary to capture all detail
with the initial surface due to the subsequent surface evolution. The interpolation
is performed on a grid with spacing h, where h is the minimum spacing of the
dataset. Evaluating (1) for the whole grid containing the liver is not necessary
and would be too time-consuming. Instead, the surface is traced and the signed
distance function is only evaluated at grid nodes belonging to grid cells which
are intersected by the surface. In order to generate a rendering of the surface,
as shown in Fig. 2(d), each intersected grid cell is polygonized according to the
Marching Cubes algorithm [12].

1
To aid the definition of intersecting contours, the intersection points of all other
contours with the current slice are shown. When performing a new delineation, spline
control points are then placed such that the curve passes through these intersection
points. As this will usually not yield true intersections, two contours are considered
to intersect if the distance between their closest points is below a threshold.

182
 (a) (b)

(c) (d)

Fig. 2. (a) Manual delineation of a liver cross-section; spline control points are de-
noted with ”×”. (b) Six cross-sectional contours are defined, two for each orthogonal
view orientation. (c) Visualization of calculated 3-D normals; at the points A – H,
the planar normal of the contour defined in the x/y plane is extended to 3-D using
the z-components of normals of intersecting contours defined in x/z and y/z planes,
respectively; the z-component of the normal at P is determined by linear interpolation
between the z-components at D and E according to the arc length distances from D to
P and from P to E. (d) Rendering of the interpolated surface.

3 Surface Evolution
The initial surface calculated only with contour information passes through the
manual delineations. In distant regions, we found it to be a good approximation
of the actual organ shape. In order to improve accuracy, the surface is evolved in
a second step using both image and contour information. Evolving the surface
solely using image information would forfeit the valuable information of the
manual delineations and fail for low-contrast or missing boundaries, as illustrated
in Fig. 1(b). As mentioned before, our paradigm is to specify shape information
where image information is ambiguous. In other regions, image information is
used, freeing the user from having to perform numerous manual segmentations.
The level set evolution takes place using the same grid as the initial surface
reconstruction. We decided for edge-based Geodesic Active Contours [7], because
we want pathologic structures to be included in the segmentation and the sta-

183
tistical information required for region-based approaches is often destroyed by
such structures. The level set function φ that represents the deforming surface
through its zero level set is evolved according to
 
∂φ (x) ∇φ (x)
= |∇φ (x)| div g (|∇I (x)|) , (2)
∂t |∇φ (x)|
where I is the input image or volume, g is a stopping function [7], and ”div”
denotes the divergence operator. Most level set applications use a constant speed
component to drive the deforming surface to object boundaries. With the initial
surface calculated in the previous step being a good estimate of the actual organ
shape, no such constant speed, which might drive the deforming surface across
weak boundaries (see Fig. 1(b)), is necessary here.
In order to incorporate the shape information, a method is required which
fits a surface to a set of contours or contour points. As the surface reconstruction
will be combined with the level set segmentation, the RBF approach described
before cannot be used again. We instead employ the level set surface reconstruc-
tion approach proposed in [13], which can be combined straightforwardly with
the level set segmentation. While delivering locally smooth surfaces, the level
set reconstruction method lacks the higher order global smoothness of the RBF
reconstruction in regions with a sparse set of input points, which prohibits using
it for determining the initial surface. Now the shape information shall only in-
fluence the deforming surface in regions close to predefined contours. The level
set surface reconstruction is well suited for that task.
Given a set of points, the level set surface reconstruction finds a surface which
is smooth and which minimizes an energy taking the distance of the evolving
surface to the points into account [13]. The according level set equation is
 
∂φ (x) ∇φ (x)
= |∇φ (x)| div δ (x) , (3)
∂t |∇φ (x)|
where δ is the distance function to the input points. In contrast to the surface
reconstruction using radial basis functions presented in the previous section, this
approach relies only on the unsigned distance function, no normals are required.
We calculate δ for (3) using the fast sweeping algorithm [14]. Therefore M input
points are sampled in total from all contours. With the fast sweeping algorithm
having linear complexity in both the number of input points M and the number
of grid points, M can be chosen large, e.g. M = 10N , ensuring that the shape
constraint captures all detail the user has specified when defining the initial
contours.
Combining (2) and (3), the joint image and shape information evolution
equation takes the form
 
∂φ ∇φ
= |∇φ| div (µ (δ) g (|∇I|) + (1 − µ (δ)) δ) . (4)
∂t |∇φ|
The spatial position x is omitted in (4) for the sake of a simpler notation. µ
is a function that weights the influence of each component with respect to the

184
 Dataset Overlap Error Volume Diff. Avg. Dist. RMS Dist. Max. Dist. Total
[%] Score [%] Score [mm] Score [mm] Score [mm] Score Score
1 10.1 61 10.0 47 1.7 58 2.6 64 20.0 74 61
2 9.7 62 8.7 54 1.5 63 2.6 63 27.3 64 61
3 8.7 66 5.3 72 1.8 55 3.1 57 26.2 65 63
4 13.1 49 13.5 28 2.5 37 4.1 43 25.3 67 45
5 10.9 57 10.1 46 2.1 47 3.6 50 35.1 54 51
6 7.6 70 3.1 84 1.2 70 2.4 67 23.8 69 72
7 9.1 65 8.7 54 1.4 64 2.3 68 16.6 78 66
8 11.1 57 11.4 39 2.0 51 2.7 63 13.7 82 58
9 10.1 61 8.6 54 1.4 65 2.3 69 19.7 74 65
10 10.9 57 8.6 54 1.9 53 3.1 57 18.8 75 59
Average 10.1 60 8.8 53 1.7 56 2.9 60 22.6 70 60

Table 1. Results of the comparison metrics and scores for all ten test cases.

distance from a manually delineated contour. We have chosen a sigmoid function

for µ, whose shift and slope control the width and offset of the transition band
from pure shape based evolution (µ = 0) to pure image based evolution (µ = 1).

4 Results

The proposed algorithm has been evaluated for ten abdominal CT scans of the
liver. The segmentation results were compared to manual expert segmentations
and given a score. The scale was set such that a score of 100 points was awarded
to a perfect liver segmentation with 100% volume overlap and a score of 75 points
was awarded to a segmentation with a volumetric overlap error of 6.4%, a relative
absolute volume difference of 4.7%, an average symmetric absolute surface dis-
tance of 1.0mm, a symmetric RMS surface distance of 1.8mm, and a maximum
symmetric absolute surface distance of 19.0mm. These values reflect deviations
of a typical manual segmentation from the perfect reference segmentation. The
results are presented in Table 1. Figure 3 shows the segmented liver boundary
and the reference boundary for a selection of slices. The initial and final position
of the surface is illustrated in figure 4(a), figure 4(b) shows a rendering of the
segmented liver shape of dataset No. 6.
Defining the initial contours involves selecting MPR slices and placing spline
control points in them. This typically took around 5 min. in total for six to
eight manual delineations. Our non-optimized implementation performed the
reconstruction of the initial surface in usually less than 2 min., evolving the
surface took 2.5 to 5 min. on standard computational hardware.

185
Fig. 3. From left to right, a sagittal, coronal and transversal slice from a relatively easy
case (1, top), an average case (4, middle), and a relatively difficult case (3, bottom). The
outline of the reference standard segmentation is in red, the outline of the segmentation
of the method described in this paper is in blue. Slices are displayed with a window of
400 and a level of 70.

5 Discussion and Conclusion

We have presented a novel two-stage semi-automatic algorithm for the 3-D seg-
mentation of complex organ structures and demonstrated its value for the seg-
mentation of the liver. In the first stage, the user manually delineates cross-
sections of the anatomical structure. Given these contours, an initial surface is
interpolated using radial basis functions. In the second stage, this surface is au-
tomatically evolved under the level set framework using a new combination of
both image and shape information.
Combining manual delineation with automatic surface reconstruction and
image segmentation has several advantages. With the reconstructed surface pro-
viding a good initialization, only a few level set iterations are required until the
equilibrium between image- and shape-derived speed components is reached. In
addition, a constant speed component, which might drive the deforming sur-

186
 (a) (b)

Fig. 4. (a) Starting at the position marked in blue, the initial surface deforms using
both image and shape information to the final position marked in red. (b) Rendering
of the segmented liver surface of dataset No. 6.

face across weak boundaries, is superfluous. During the surface evolution, shape
information is used in areas close to manually delineated contours. Thereby, am-
biguous image information can be resolved by the user. In other areas, image
information prevails, thus reducing the required number of manual delineations.
Another benefit of our approach is that no underlying model of the shape char-
acteristics and therefore no training phase is required.
The results listed in Table 1 show that the semi-automatic algorithm is able
to accurately segment the liver. While it does not reach the score of a manual
segmentation, the required amount of user-interaction is significantly smaller,
only 7 slices on average were manually delineated.
The main cause for segmentation errors is an initial surface including high-
contrast boundaries not belonging to the liver. Such boundaries are for example
encountered for ribs, the right kidney, or in areas where organ or muscle tissue
is adjacent to fat. During the deformation stage, the surface will usually stick to
those spurious boundaries. An accurate initialization is also of high importance
for the central part of the liver, which has an extremely complex anatomy due to
separating lobes and branching vessels. Otherwise, the deforming surface may
fail to recover concave regions or to exclude vessels. This usually leads to an
oversegmentation and can be observed especially for the datasets No. 4 and 5.
In both cases, the initial surface was reconstructed from six delineations.
The afore-mentioned problems can be solved by specifying additional con-
tours in regions where the initial surface is inaccurate. Even a single additional
contour often leads to a drastic improvement. With our non-optimized imple-
mentation taking around two minutes for the surface reconstruction, future work
will focus on increasing the performance. Having a prompt reconstruction from
the specified set of contours would conveniently allow the user to add exactly
the required amount of information. Instead of specifying additional contours,
the problem of the surface sticking to spurious high-contrast boundaries between

187
muscle or organ tissue and fat might also be addressed by incorporating intensity
information in the deformation stage. With fat having characteristic absorption
values in CT images, the surface could be forced to retract from those regions
erroneously included in the initialization and to move back to the liver.

References
1. Leventon, M.E., Grimson, W.E.L., Faugeras, O.D.: Statistical shape influence in
geodesic active contours. In: CVPR. (2000) 1316–1323
2. Soler, L., Delingette, H., Malandain, G., Montagnat, J., Ayache, N., Koehl, C.,
Dourthe, O., Malassagne, B., Smith, M., Mutter, D., Marescaux, J.: Fully au-
tomatic anatomical, pathological, and functional segmentation from ct scans for
hepatic surgery. In Hanson, K.M., ed.: Proc. SPIE Medical Imaging 2000: Image
Processing. Volume 3979 of Presented at the Society of Photo-Optical Instrumen-
tation Engineers (SPIE) Conference. (2000) 246–255
3. Heimann, T., Wolf, I., Meinzer, H.P.: Active shape models for a fully automated 3d
segmentation of the liver – an evaluation on clinical data. In Larsen, R., Nielsen, M.,
Sporring, J., eds.: Medical Image Computing and Computer-Assisted Intervention
– MICCAI 2006. Volume 4191 of LNCS., Springer (2006) 41–48
4. Schenk, A., Prause, G.P.M., Peitgen, H.O.: Efficient semiautomatic segmentation
of 3d objects in medical images. In: MICCAI 2000: Proceedings of the Third
International Conference on Medical Image Computing and Computer-Assisted
Intervention, Springer-Verlag (2000) 186–195
5. Falcao, A.X., Udupa, J.K., Samarasekera, S., Sharma, S., Hirsch, B.E.,
de A. Lotufo, R.: User-steered image segmentation paradigms: live wire and live
lane. Graph. Models Image Process. 60(4) (1998) 233–260
6. Grady, L.: Random walks for image segmentation. IEEE Transactions on Pattern
Analysis and Machine Intelligence 28(11) (2006) 1768–1783
7. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. In: ICCV. (1995)
694–699
8. Sethian, J.A.: Level Set Methods and Fast Marching Methods. second edn. Cam-
bridge University Press (1999)
9. Späth, H.: One Dimensional Spline Interpolation Algorithms. AK Peters (1995)
10. Turk, G., O’Brien, J.F.: Shape transformation using variational implicit functions.
In: SIGGRAPH ’99: Proceedings of the 26th annual conference on Computer graph-
ics and interactive techniques, New York, NY, USA, ACM Press/Addison-Wesley
Publishing Co. (1999) 335–342
11. Carr, J.C., Beatson, R.K., Cherrie, J.B., Mitchell, T.J., Fright, W.R., McCallum,
B.C., Evans, T.R.: Reconstruction and representation of 3d objects with radial
basis functions. In: SIGGRAPH ’01: Proceedings of the 28th annual conference on
Computer graphics and interactive techniques. (2001) 67–76
12. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface con-
struction algorithm. In: SIGGRAPH ’87: Proceedings of the 14th annual conference
on Computer graphics and interactive techniques. (1987) 163–169
13. Zhao, H.K., Osher, S., Fedkiw, R.: Fast surface reconstruction using the level
set method. In: Proceedings of the IEEE Workshop on Variational and Level Set
Methods (VLSM’01), Washington, DC, USA, IEEE Computer Society (2001) 194
14. Zhao, H.K.: A fast sweeping method for eikonal equations. Mathematics of Com-
putation 74 (2005) 603–627

188