Published on: Mar 3, 2016
Transcripts - NabinMulepatiNCUR2010
Automated Segmentation of Liver Tumors from CT Data
with a Two-stage Multi-resolution Approach
Nabin Mulepati (Dr. Toshiro Kubota), Department of Computer Science, Susquehanna University, Selinsgrove, PA
This paper presents an automated system that detects and segments liver tumor
from CT (Computed Tomography) data and compares the results with the
previous system developed by Kubota1. The system consists of two stages. The
first stage sub-samples the volume by 4x4x4, detects candidate locations, and
provides rough segmentation at each candidate location. The second stage
refines each segmentation by sub-sampling at either 2x2x2 or 1x1x1
depending on the size of the segmentation. The Kubota algorithm is then run
on a sub-volume around the first stage segmentation. The system was
experimented with 16 CT data sets of the abdominal section. Four of the CT
data sets with 10 tumors accompany manual segmentation done by
radiologists, and were used in our quantitative evaluation. Visual observations
were made for the rest of the data set. Tanimoto error is used for quantitative
measure. A lower Tanimoto error indicates better segmentation. On average,
the Tanimoto error for our segmentation was (44.2 ± 13.8)% compared to
(63.3 ± 17.9)% by the system reported in . The results show that the
proposed method makes significant improvements over the previous method.
Liver cancer, the fourth leading cancer in the world, is responsible for
approximately one million deaths every year.2 The National Cancer Institute,
estimated 22,620 new cases and 18,160 deaths because of liver and inter-
hepatic bile duct cancer in the United States in 2009.
In order to keep track of the growth of a liver tumor, doctors need to know the
volume of the tumor. Manual segmentation of tumors from CT data is difficult
and time consuming. An automated segmentation will significantly reduce the
time required and potentially reduce both intra and inter-observer variations.
The system presented below is an extension of the system reported in  and
makes use of the same algorithm.
'()$ *+,-.%$& /$01"-(.+0%($
'10130%($ !.+$16?%00%$& *.%3%$& ;1:%$1,1$0
;1 %$01"-(.+0%($ 5<"12<(.>%$&
Fig. 1. A flow diagram of the system reported in .
Large number of candidates are generated which are first filtered by location and then by
intensity. Filtering by location removes all candidate locations near the boundaries of the
volume whereas filtering by intensity removes all candidates whose intensity values at the
location are above the cut-off value of 1100, which is set so that true positives are not filtered.
In this stage the main part of the tumor is extracted.1 At each candidate location a rough
segmentation is computed. False positives are removed and overlapping candidates are grouped
together into clusters.
A more accurate segmentation is computed for each cluster from the step above. Competition-
diffusion algorithm3 is applied to extract candidates whose intensity values lie in the minimum-
maximum range of intensities of the core of the cluster. Then region tracing is implemented and
the segmentations are selected based on their size.
The segmentation obtained is first rescaled to the size of the original CT volume and then a
threshold of 0.5 is applied.
In the second stage of the system, the original CT volume is sub-sampled by either 2x2x2 or
1x1x1 depending on the size of the segmentation obtained from the first stage. The five steps
described in the first stage above is then re-run on the sub-sampled CT volume to get the final
The first stage consists of five steps as shown in Figure 1.
The CT Volume is down-sampled to 4x4x4 voxels to make the segmentation
Gaussian smoothing is applied to make the sub-volume smoother and also to
reduce the amount of aliasing that may have resulted from down sampling.
Next, the sub-volume is analyzed for candidate points that are likely to be
inside a tumor
Experiments and Results:
(a) (b) (c)
Fig. 4. Original CT volumes with manual segmentation super-imposed.1
(a) (b) (c)
Fig. 5. Results of the first stage segmentation.1
The system was experimented with 16 CT volumes of the abdominal section. Four CT volumes
with 10 tumors, which accompany manual segmentation done by radiologists, were used in
Fig. 4 shows some of the CT volumes with manual segmentation overlaid. Fig. 2 shows the
result at each step of the first stage segmentation. Fig. 2 (a) shows an axial slice of the CT
volume named IMG04-L4 in Table 1. Fig. 2(b) shows the same axial slice after 4x4x4 sub-
sampling. Fig. 2(c) shows the result of Gaussian smoothing on the sub-sampled axial slice with
candidate points overlaid by circles. Fig. 2(d) shows the result of the initial segmentation.
Fig. 2(e) is obtained by applying competition diffusion on the smoothed slice in fig. 2(c).
Fig. 2(f) is obtained by applying a threshold of 0.5. Fig. 2(g) shows the result of the first stage
segmentation and fig. 2(h) shows the corresponding manual segmentation superimposed.
The original CT volume is sub-sampled by 2x2x2 or 1x1x1 depending on size of the
segmentation of the first stage and then the five steps from the first stage is implemented. Fig. 3
shows the final result of the second stage segmentation. Figs. 4 (a), 4(b), and 4(c) show the CT
volumes named IMG01-L1, IMG02-L1, and IMG03-L1 in Table 1 with manual segmentation
super-imposed. Fig. 5 shows the results of the first stage segmentation on the same volumes.
Likewise, fig. 6 shows the final results of the second stage segmentation. Visual observations
were made on the remaining 12 CT volumes that do not accompany manual segmentations.
Fig. 6. Results of the second stage segmentation.
Fig. 2. Intermediate steps of the first stage segmentation.1
!,#$/0-"-&,1 !2#$3.4& 5,*61)' !7#$(*..+8)' !'#$95+$()"*)&+,+-.&$:.0)
!)#$;&-+-,1$:.&<-"= !<#$>.0)"0.?&' !8#$@6 5,*61)'
Fig. 3. Result of the second stage
① T. Kubota, Efficient Automated Detection and Segmentation of Medium and Large Liver
Tumors: CAT Approach, 3d Segmentation in Clinic, NYC, NY. Sep 2008.
② “Looking at liver cancer,” EBSCOhost Academic Search Premier, Nursing [0360-4039]
Pellegrino yr:2006 vol:36 iss:10 pg:52 -55.
③ T. Kubota and F. Espinal, Reaction-Diffusion Systems for Hypothesis Propagation, ICPR
Table 1 showing quantitative results.
Using a multi-resolution approach on the method reported in
 has proven to provide better segmentation. On average the
Tanimoto error for our system was was (44.2 ± 13.8)% which
is much lower compared to (63.3 ± 17.9)% of the method in
. These results show that the proposed method makes
significant improvements over the previous method.
Fig. 5 shows the segmentation results of the system reported
in ; the segmentations seem to be under-segmented. The
final segmentations of our system (fig. 6) is closer to the
manual segmentations than the segmentations of the previous
method. Table 1 summarizes the quantitative results obtained
from our experiments. On average the Tanimoto error for our
system was (44.2 ± 13.8)% which is better compared to (63.3
± 17.9)% by the system reported in .