A semi-automated method for liver tumor segmentation
based on 2D region growing with
Damon Wong1, Jiang Liu1, Yin Fengshou1, Qi Tian1, Wei Xiong1,
Jiayin Zhou2, Yingyi Qi2, Thazin Han2, Sudhakar K Venkatesh2,
and Shih-chang Wang2
1. Institute for Infocomm Research, Agency for Science, Technology and Research, Singapore
2. Department of Diganostic Radiology, National University of Singapore, Singapore
3. School of Computing, National University of Singapore, Singapore
Abstract. Liver tumour segmentation from computed tomography (CT) scans is
a challenging task. A semi-automatic method based on 2D region growing with
knowledge-based constraints is proposed to segment lesions from constituent
2D slices obtained from 3D CT images. Minimal user involvement is required
to define an approximate region of interest around the suspected legion area.
The seed point and feature vectors are then calculated and voxels are labeled
using a region-growing approach. Knowledge-based constraints are
incorporated into the method to ensure the size and shape of the segmented
region is within acceptable parameters. The individual segmented lesions can
then be stacked together to generate a 3D volume. The proposed method was
tested on a training set of 10 tumours and a testing set of 10 tumours. To
evaluate the results quantitatively, various measures were used to generate
scores. Based on the results obtained from the 10 testing tumours, the method
was resulted in an average score of 64.
According to a recent report on cancer statistics, liver tumor is the third highest
cause of death due to cancer. Although the incidence rate is the sixth highest, at 5.7%
of new cancer cases, or 626,000 cases, the mortality rate due to liver cancer is almost
as high at 598,000. The high mortality rate has been attributed to poor prognosis of
the disease, which underscores the need for an accurate assessment of the cancer.
Computed Tomography (CT) scans are widely used mode of non-destructive and
non-invasive imaging for observation of internal physiological structures in the body.
The near-isotropic features of modern CT machines allows for volumetric
reconstruction obtained via individual slices by stacking the CT slices accordingly.
Particularly for diagnosis of liver cancer, it is important to obtain an accurate
portrayal of the size of the tumour to determine its severity. Furthermore, tumour
localization and volume determination is important for radiotherapeutic treatment
management in techniques such as 3D Conformal Radiotherapy (3DCRT) and
Intensity Moduldated Radiotherapy (IMRT), where tumor information is vital for
correct dosimetry calculations.
To identify tumours from CT slice images, there is a need for identification, or
segmentation, of tumourous lesions. Typically, this has been manually done by
trained clinicians. The task is time-consuming, requiring much effort and can be
subjective depending on the skill, expertise and experience of the clinician. Objective,
computer-aided segmentation of CT images would thus be a great boon for lesion
identification. However, this remains a challenging task, due to a number of factors,
mainly the low density contrast between lesions and surrounding normal liver tissues,
but also, and not limited to, the irregularity of the shape of the lesions and the
similarities in image characteristics of the liver with surrounding organs.
A number of methods have been proposed for the computer-aided segmentation of
tumours from medical images. In , Mahr et al reviewed and compared the various
techniques which included region-growing, isocontour, snakes, hierarchical and
histogram-based methods, and found that region-growing and snakes, also sometimes
known as active contour models, were the most promising for future investigation on
liver volumetry determination. In , Liu et al reported on the use of a snake
algorithm based on gradient vector flow for liver segmentation, incorporating the use
of a liver template and edge detection for better initialization of the method. A
possible limitation on the use of active contour models is their reliance on a model for
initialization to produce an accurate segmentation. In tumour segmentation, the large
variability of the shape of lesions for different patients and possibly even for different
slices of the same tumour can be a challenge to model.
Seeded region-growing  techniques are also another popular method for medical
image segmentation. Typically, features are generated from a select number of
seed points, and the initial region is gradually increased by incorporating
neighbouring pixels with similar feature vectors. The method takes advantage of the
potential similarities of pixels belonging to the same structure and can be effective if
proper features are chosen. However, the method requires the selection of proper seed
points, as incorrect seeds would lead to wrong segmentations. Furthermore, region-
growing can be highly computationally-intensive particularly for large images.
Recently, there has also been a trend to utilize a priori information into
segmentation and classification methods to improve their effectiveness by
incorporating heuristics specific to the task. Also known as knowledge-based
techniques, these methods make use of domain-knowledge to filter out extraneous
regions by employing heuristic-based contraints and filters, or to aid classification
through use of templates and landmarks. Knowledge-based techniques have been used
to identify abdominal organs  and identify brain tumours  from CT slices.
In this paper, we propose a semi-automated method to identify tumours from 3D
CT scans. After decomposing the 3D scan into its component slices, we apply 2D
region growing with knowledge-based constraints on each slice. User interaction is
employed to establish an approximate region of interest (ROI) around the lesion in
each slice image. This improves the performance of region growing, as well as
reduces computational requirements. During the region-growing process, knowledge-
based constraints are implemented to constrain the emergent segmented region to
within acceptable parameters. In Section 2, an overall framework of the method
employed will be presented, followed by a detailed description of the processes used
in the method. Subsequently, in Section 3, we describe the data provided for
benchmarking and report on the results obtained on applying our method to the testing
data. Section 4 ends the paper with a discussion of the results and conclusion of the
findings from the method presented here.
For the method presented, we adopted a semi-supervised method based on 2D
region growing with knowledge-based constraints to segment the lesions from the CT
images. Figure 1 shows the overall framework for the lesion segmentation. The
selection of the start and end slices for the segmentation would need to be manually
defined, after which the method is applied on each of the selected slices. The
following subsections describe the steps employed in greater detail.
Fig. 1. Liver tumour segmentation process framework.
2.1 Preprocessing, ROI selection and seed point generation
First, to reduce the granular noise in the CT slice image, median filtering via a 3x3
voxel square kernel is convolved across the entire image. Subsequently, the contrast
of the CT slice is enhanced to improve visual perception of the structures in the
image. Next, a region of interest (ROI) was manually determined by selecting two
points on the CT slice image. The two points are used to indicate the diagonal limits
of the ROI which contain the tumour lesion. This helps to localize and constrain the
region-growing to within the ROI. Furthermore, it helps to avoid inaccurately
identifying erroneous regions, particularly outside the liver. ROI selection has to be
manually performed only on the first slice. For subsequent slices, the ROI can be
reselected, or it can make use of the ROI from the prior slice.
An additional benefit is the reduced processing time by limiting the ROI. For a
typical CT slice image with a resolution of 512x512 voxels, generally the size of the
lesion is less than 10% of the total number of voxels. Applying the region-growing on
only the ROI instead of the entire CT slice would result in savings in the
2.2 Region Growing
Next, a seed point for the region growing algorithm is automatically defined at the
centre of the ROI. Concurrently, the average voxel intensity in the neighbouring
region is calculated and is defined as the initial feature metric. Subsequently, a search
is conducted in the 4-connected neighbourhood to determine the voxel which has the
least intensity difference from the initial average intensity. The determined voxel is
then added to the region.
After initialization, the subsequent iterative region-growing process is similar to
the method described in . Let L be the set of voxels labeled as part of the lesion,
and I be the set of voxels t in the ROI which are not part of L but are neighbours to L.
N(t) represents the voxels in the immediate neighbourhood of t.
Next, the difference in the grayscale intensity levels between the voxels in I and
the voxels in L are calculated as ( ) t
The region-growing method then determines the minimum ( ) t
( ) t
The voxel t1 corresponding to the minimum ( ) t
the set L. The segmentation labels are updated and the algorithm reiterates until the
values of the grayscale intensity in the neighbouring voxels are higher than a pre-
defined threshold value, which is empirically obtained. Morphological closing using a
flat, disc-shaped structuring element with a voxel radius of 15 is then performed on
the segmented region to smoothen out the region boundaries.
is determined and t1 is added to
Fig. 2. Liver tumour segmentation process. (a) shows the manual ROI corner selection. The red
crosses indicate the selected points and subsequent ROI is the rectangle with the white dotted
outline. The contrast of the CT slice has been enhanced for better differentiation between the
various structures. (b) shows the segmented lesion after region-growing and (c) presents the
overlay of the boundary from (b) on the original CT slice image, as denoted by the red outline.
(d) shows the ground truth segmentation in red of the lesion. The image is from Slice 139 from
Lesion 2 of Patient Data 1.
2.3 Knowledge-based constraints
During the segmentation process, it was found that some attempts resulted in an
unusually high regional average intensity, due to the initial seed point being located
on bright spots. This caused premature termination of the region-growing algorithm
and over-segmentation of the tumour lesion. To avoid these effects, a constraint was
imposed on the initial segmented region to occupy a minimum fraction of the total
ROI area. The fraction was empirically set but a good approximation was found to be
half the size of the ROI. In circumstances where the initially segmented region does
not meet this constraint, the region growing is performed again using the same ROI
but with a larger area to calculate the initial average intensity. The subsequent
segmentation is then fused with the initial segmentation using a logical addition
operator, after which morphological closing is applied again to the combined region
to obtain the final segmentation result for the CT slice.
3 Testing Data & Results
To evaluate the results using the proposed method, liver tumour CT data sets were
provided by the organizers of the workshop. The slices were obtained via a 64-slice
and two 40-slice CT scanner in a standard four-in-plane contrast enhanced imaging
protocol with a slice thickness of 1mm of 1.5mm, and an in-plane resolution of 0.6-
0.9mm. 10 liver tumours from four patients were used for training, and another 10
tumours from five patients were used in testing. A final ten images will be used for
the onsite during the segmentation workshop. The datasets represent a range of
patients, pathology and CT scanning phases. All the ground truths were manually
segmented by an experienced radiologist and confirmed by another radiologist as
reference for evaluation purposes.
Figure 3 and 4 show the visual examples of the segmentation results for the
training set and the testing set respectively. For Figure 3, the corresponding ground
truth results as determined by the radiologists are also included for comparison. The
results were also evaluated quantitatively using the following five measures, (1)
relatively absolute volume differences, (2) average symmetric absolute surface
distance, (3) symmetric RMS surface distance, (4) maximum symmetric absolute
surface distance, and (5) volumetric overlap error, and are tabulated in Tables 1 and 2
for the training set and the testing set respectively. For the testing set, an automatic
scoring system  was used to assign scores to the results obtained under the five
measures, with the score ranging from 0 to 100. A score of 100 represents the perfect
segmentation while 0 is the minimum score one segmentation will get. Table 2
includes the corresponding scores for the measures.
Fig 3. Segmentation results (top row) and ground truth (bottow rom) for Slices 86-
93 of Lesion 2 from Dataset 4.
Table 1. Quantitative results obtained from the training data set
Ave. Surf. Dist.
RMS Surf. Dist.
34.45 11.2544 2.0702 2.832 14.2655
28.7 8.86 1.0226 1.6305 9.8152
25.07 3.8759 0.8661 1.3366 9.5636
28.61 25.568 1.1983 1.8093 9.3226
27.97 7.4968 1.0106 1.5866 8.8689
41.99 21.0929 1.0384 1.6694 9.4826
18.87 0.891 1.505 2.319 18.3036
9.29 2.8418 0.3457 0.6532 5.133
Ground truth Segmentation
Fig 4. Segmentation results for Slices 163 to 170 of Lesion 3 from Patient Dataset 5.
Fig 5. Selected poor segmentation results from (a) Dataset 5 Slice 143, (b) Dataset 5
Slice 149, (c) Dataset 7 Slice 88 and (d) Dataset 7 Slice 119. The poor segmentation
can be attributed to low contrast visibility in (a) and (b), and non-uniform lesion
texture in (c) and (d).
(a) (b) (c) (d)
Table 2. Quantitative results obtained from the testing data set
Overlap Error Volume Difference Ave. Surf. Dist. RMS Surf. Dist. Max. Surf. Dist.
Tumor (%) Score (%) Score (mm) Score (mm) Score (mm) Score
IMG05_L1 36.05 72 6.78 93 2.99 24 4.17 42 18.59 53 57
IMG05_L2 41.85 68 23.52 76 1.73 56 2.53 65 12.48 69 67
IMG05_L3 36.93 71 3.61 96 1.50 62 2.18 70 10.97 73 74
IMG06_L1 50.06 61 41.81 57 1.20 70 1.61 78 5.23 87 70
IMG06_L2 48.25 63 19.74 80 1.22 69 1.78 75 9.35 77 73
IMG07_L1 40.76 69 35.73 63 5.61 0 7.59 0 29.87 25 31
IMG07_L2 31.46 76 18.87 80 1.60 60 2.25 69 10.18 75 72
IMG08_L1 18.24 86 12.18 87 2.16 45 3.11 57 13.62 66 68
IMG09_L1 46.85 64 38.49 60 1.53 61 2.08 71 7.55 81 67
IMG10_L1 43.50 66 41.22 57 2.45 38 2.95 59 9.07 77 60
Average 39.40 70 24.20 75 2.20 49 3.02 59 12.69 68 64
4 Discussion & Conclusions
From an analysis of the results, the proposed method performs well for lesions that
are well-defined and have uniform grayscale intensity throughout the lesion. Although
some user interaction is required, the effort needed is minimal as the user need only to
define an approximate ROI around the suspected lesion, rather than manually
delineating the lesion boundary. Furthermore, the method, combined together with
ROI constraints, is relatively fast, with each slice taking about one to two seconds to
process in a MATLAB operating environment running on a 3GHz Dual Core Pentium
PC with 4 GB RAM. Providing some allowance for adjusting of the parameters,
segmenting an entire tumour typically consisting of approximately 25 slices should
take about ten minutes.
However, it was observed that for lesions with certain characteristics, segmentation
performance was sub-optimal. In particular, for lesions with low contrast compared t0
normal liver tissue, such as in Figs 5(a) and 5(b), the feature space separation between
lesion voxels and normal tissue voxels can be minimal, leading to difficulty in
accurately labeling voxels. Furthermore, when the lesion has a patchy, non-uniform
appearance, as can be seen in Figs 5(c) and 5(d), the visual characteristics of the
lesion can be observed to vary even on the same CT slice, since the resultant labeling
of voxels would depend on the initial seed selection. Where parts of the lesion have
characteristics which are too different from the seed points, these voxels would be
mistakenly mis-labelled as non-tumour voxels. Possible strategies to overcome these
challenges could be the selection of different features, or aggregating feature vectors
from multiple seed regions to increase the robustness of the system.
In this paper, we have presented a semi-automated method for 3D segmentation of
tumours from CT slices. First, the 3D scans are decomposed into its constituent parts,
after which 2D region-growing with knowledge-based constraints is applied to
segment the lesion voxels from the normal tissue voxels. From the results obtained
using the training data set and the testing data set, the method was found to perform
adequately. Further improvements can be implemented to improve the performance of
the proposed method.
This work is supported by a research grant (SBIC RP C-008/2006) from the
Singapore BioImaging Consortium, Agency for Science, Technology and Research.
1. Parkin, D.M., Bray, F, Ferlay, J, Pisani, P.: Global Cancer Statistics 2002.CA Cancer J
Clin, 55:74-108 (2005).
2. Mahr, A., Levegrün, S, Bahner, M.L., Kress, J., Zuna, J., Schlegel W.: Usability of
semiautomatic segmentation algorithm for tumor volume determination. Invest. Radiol. 34,
3. Liu, F., Zhao, B., Kijewski, P.K., Wang, L., Schwartz, L.H.: Liver segmentation for CT
images using GVF snake. Med. Phys. vol. 32(12), 3699-3706, 2005.
4. Adams, R. Bischof, L.: Seeded Region Growing. IEEE Trans. Pattern Anal. Mach. Intell. vol
16(6), pp. 641-647 (1994).
5. Pohle,R., Toennies, K.D.: Segmentation of Medical Images Using Adaptive Region
Growing. Proc. SPIE Medical Imaging, vol. 4322, pp. 1337–1346. (2001)
6. Kobashi,M. and Shapiro,L.G.: Knowledge-based organ identification from CT images.
Pattern Recogn., vol. 28, pp. 475–491 (1995).
7. Clark, M.C. Hall, L.O. Goldgof, D.B. Velthuizen, R. Murtagh, F.R. Silbiger, M.S.:
Automatic tumor segmentation using knowledge-based techniques. IEEE Trans. Med.
Imaging. vol. 17(2), pp. 187-201 (1998).