AUTOMATIC COUINAUD LIVER AND VEINS SEGMENTATION

FROM CT IMAGES

Dário A. B. Oliveira, Raul Q. Feitosa and Mauro M. Correia

Department of Electric Engineering, Catholic University of Rio de Janeiro, Rio de Janeiro, Brasil

Department of Computer Engineering, Rio de Janeiro State Universityl Brasil

Unigranrio and National Cancer Institute-INCA, Rio de Janeiro, Brasil

Keywords: Medical Imaging, Liver Segmentation, Vessel Segmentation, Computed Tomography.

Abstract: This paper presents an algorithm to segment the liver structures on computed tomography (CT) images

according to the Couinaud orientation. Our method firstly separates the liver from the rest of the image.

Then it segments the vessels inside the liver area using a region growing technique combined with

hysteresis thresholding. It separates the vessels in segments without any bifurcation, and using heuristics

based on anatomy, it classifies all vessel segments as hepatic or portal vein. Finally, the method estimates

the planes that best fit each of the three branches of the segmented hepatic veins and the plane that best fits

the portal vein. These planes define the subdivision of the liver in the Couinaud segments. An experimental

evaluation based on real CT images demonstrated that the outcome of the proposed method is generally

consistent with a visual segmentation.

1 INTRODUCTION

By and large the CT data analysis is performed

visually by a radiologist. This is a time consuming

task, whose accuracy depends essentially on the ex-

perience of the analyst. Digital Image Processing

techniques can be used to develop methods that

automatically perform many of the tasks involved in

the CT analysis, improving productivity and the

overall accuracy.

The segmentation process is particularly arduous

in abdominal CT images because different organs lie

within overlaping intensity value ranges and are

often near to each other anatomically. Many

techniques have been proposed in the literature for

the analysis of abdominal CT scans. They can be

roughly divided in two main groups: model driven

and data driven approaches (Masutani et al, 2005).

The blood vessel definition is an essential step in

several medical imaging applications. They can be

used as reference to segment different organs and

structures in the human body. Kirbas et al (2004)

presented a review of vessel extraction, in which

many of the available techniques are described in

details.

This paper presents a data driven method to

segment the liver into the eight different regions

proposed by Couinaud (1957), using the hepatic and

portal veins position in the liver. It deals with the

case of low contrast and erroneous connection

between the hepatic and portal veins, as a

improvement proposed in previous work (Oliveira et

al., 2007).

The subsequent text is organised in the following

way. Section 2 presents segmentation method in

details, section 3 reports some results, and the main

conclusions are presented in section 4.

2 THE 3D SEGMENTATION

METHOD

The segmentation method consists of five main

steps:

a) segmentation of organs and muscle tissues,

b) segmentation of the liver,

c) segmentation of the vessels within the liver ,

d) classification of segmented vessels as hepatic

and portal veins, and

e) determination of Couinaud sectors.

249

A. B. Oliveira D., Q. Feitosa R. and M. Correia M. (2008).

AUTOMATIC COUINAUD LIVER AND VEINS SEGMENTATION FROM CT IMAGES.

In Proceedings of the First International Conference on Bio-inspired Systems and Signal Processing, pages 249-252

DOI: 10.5220/0001063202490252

 SciTePress

Details of each step are presented in the next

subsections.

2.1 Segmentation of Organs and

Muscle Tissue

Organs and muscles tissue are the main presence in

abdominal images. Typical grey values of these

tissues occur around the maximum (CM) of the grey

value histogram for the whole CT sequence.

Figure 1: histogram for gray level range definition.

Figure 1 shows the histogram of a sample CT

exam, the range of intensities corresponding to

organs and muscles and the lower and upper limits

TL and TH defining this range.

Let CM be the maximum CT histogram count,

TM the corresponding intensity value, and CL and

CH the counts corresponding respectively to TL and

TH. It has been observed in our experiments that the

ratios RL=CL/CM and RH=CH/CM do not

significantly change from CT exam to CT exam. In

fact these ratios lied around RL=0.6 and RH=0.2

through all our experiments.

This regularity suggests the following procedure

to select the lower and higher threshold values:

a) Compute and smooth the histogram of the

whole CT exam;

b) Detect the maximum histogram count CM;

c) Multiply CM by factors RL and RH, and

obtain the count values CL and CH.

d) Search the smoothed histogram for the

intensity values TL and TH closest to TM

corresponding to CL and CH, such that

TL<TM and TH>TM.

2.2 Liver Segmentation

The next step consists in segmenting the liver.

Generally the liver appears as nearly homogeneous

areas on CT slices, i.e. its intensities are restricted to

a narrow grey value interval. This can be observed

in Figure 1, where the histogram of pixels belonging

to the liver is drawn in red over the histogram of the

whole CT sequence shown in blue.

The extreme values of this interval are

determined in the following way.

One image of the CT set where the liver is

present is selected as the main sample and passed as

an input parameter to the algorithm. Then, the

largest connected component of this slice located on

the upper-left side of the image (right side of the

human body), is identified and its mean value on the

original image is computed.

Using the pixels of organs and muscle tissue

previously segmented, a new grey level range is

determined following a procedure similar to the one

described in subsection 2.1. The histogram count

value corresponding to the liver mean value is used

as the maximum count value and the range limits are

calculated using as limiting ratios the value 0.8 for

both cases. The threshold values obtained this way

are applied to the regions selected in the previous

step.

A simple procedure extracts the liver from the

remaining objects. Starting on the main sample it is

executed on the next adjacent slice upward and

downward in the CT image set till all slices have

been processed. It consists of three main steps:

a) Select the biggest object in the collection;

b) If its centroid is in the upper left quadrant of

the CT image, go to step c, otherwise discard

this object from the collection and go back to

step a;

c) If the selected object is connected to another

object of an adjacent slice previously

classified as liver, classify it as liver,

otherwise discard the object from the

collection and go back to step a;

Clearly the first iteration does not pass through

step c and the object selected in step b is set as liver

directly.

2.3 Vessel Segmentation

Having segmented the liver, and considering only

the region delimited by this organ, we select a

threshold VH, such that the intensities above it

identify unambiguously the vessels inside the liver.

A second threshold VL (VL<VH) is further selected

such that intensities below it clearly indicate liver

parenchyma.

These two threshold values define three ranges

of pixel intensities, namely:

- the strong vessel range, defined by

intensities above VH,

- the weak vessel range, comprising

intensities between VL and VH, and

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

250

- the liver tissue range, covering intensities

below VL.

The construction of the vessel tree is performed

by a region growing approach consisting of the

following basic steps:

a) Build the weak vessel object set defined by

the pixels with values above VL.

b) Build the strong vessel object set defined by

the pixels with values above VH.

c) Take the strong vessel set computed in the

preceding step as the initial vessel tree

estimate, and add to it all objects of the weak

vessel set connected to it.

d) Repeat the previous step using the current

vessel tree estimated until it stops growing.

We searched appropriate values for VL and VH

manually through many experiments using different

CT sequences. We observed that the histogram

counts for the manually selected values stayed at a

roughly constant ratio to the intensity corresponding

to the maximum count.

Considering NM the maximum liver histogram

count, and NL and NH the counts corresponding

respectively to VL and VH, the ratios rl=NL/NM and

rh=NH/NM do not significantly change from CT

exam to CT exam. These ratios were determined

experimentally as rl=0.5 and rh=0.2.

Based on this regularity the following procedure

is proposed to select the lower and higher threshold

values:

a) Compute and smooth the histogram of the

image region inside the liver;

b) Detect the maximum histogram count NM

and the corresponding intensity VM.

c) Multiply NM by the ratios rl and rh, and

obtain the count values NL and NH.

d) Search the smoothed histogram for the

intensity values VL and VH corresponding to

NL and NH, whereby both VL and VH are

greater than VM.

2.4 Classification of Portal and Hepatic

Veins

The hepatic and portal veins appear as separate three

dimensional objects in most CT exams. However,

sometimes these veins touch to each other on some

CT slice, what may lead to identifying them as a

single object. In such case the Couinaud

segmentation becomes not possible.

This subsection describes a method to correctly

segment the veins even when they touch in some CT

slice.

Firstly, the method separates the vessel objects

segmented previously in connected components,

hereafter called objects, performing the following

steps:

a) The first slice S1 containing any object is

labelled.

b) The area projected by each object in S1 on

the next adjacent slice S2 is verified. If it

intersects only one object, the same label is

set to the object in S2. If it intersects more

than one object, new labels are created for

each intersected object in S2.

c) Step b is repeated until all objects in the CT

sequence are labelled.

As result vessels segments are obtained whose

extremes are determined by bifurcations, as shown

in Figure 2-b.

A second procedure is performed to classify

these vessel segments as hepatic or portal vein.

Based on knowledge of the anatomy, the following

simple algorithm is proposed. It consists of six steps:

a) The first object identified on the top slice is

selected.

b) If it is divided in three other objects in the

next adjacent slice, it is classified as hepatic

vessel, otherwise it is discarded and other

object is selected on the top slice until this

condition is reached.

c) For each of the three objects identified as

hepatic branches, the next adjacent slice is

analysed. The object with the largest

intersection area is selected as continuation of

the respective hepatic branch.

d) Step c is repeated recursively for each hepatic

branch until no other segment can be merged

to the hepatic vessel tree. At the end of this

step, the major hepatic vessels have been

identified.

e) The vessel segments not assigned to the

hepatic vessel tree up to step d are examined

and the largest 3D connected component is

labelled as the portal vein.

f) Non classified segments which are connected

to the hepatic vessel tree are merged to it.

In figure 2-a the hepatic and portal veins are

shown as a single object because they touch on some

CT slice. Figure 2-b shows in different colours

several independent segments delimited by each

bifurcation identified during the classification

process. Figure 2-c shows the final result, where the

hepatic and portal veins appear as separated vessel

trees.

AUTOMATIC COUINAUD LIVER AND VEINS SEGMENTATION FROM CT IMAGES

251

Figure 2: Vessel segmentation results – (a) the portal and

hepatic veins (b) the independent vessel segments

determined by bifurcations (c) the portal and hepatic veins

as separated vessel trees.

2.5 Segmentation of Couinaud regions

The Couinaud paradigm divides the liver into eight

independent segments each one having its own

vascular inflow, outflow, and biliary drainage.

Because of this division into self-contained units,

each can be removed without damaging those

remaining.

Our method estimates the subdivision of the liver

in the eight Couinaud segments, by fitting planes to

the portal vein, and to each of the hepatic vein

branches. To separate the three main branches of the

hepatic vein we apply the k-means algorithm on the

3 dimensional coordinates of the pixels identified in

the preceding step as belonging to the hepatic vein.

It is assumed that there are three clusters. A

restriction for singleton value is imposed so as to

guarantee that no cluster will be empty. This leads to

three different objects corresponding to each branch

of the hepatic vein.

Then, a least squares based procedure determines

the four planes that best fit the points of each branch

of the hepatic vein and the portal vein segmented

before. These planes divide the liver in the Couinaud

segments.

3 RESULTS

A software prototype implementing the proposed

method has been built for validation purpose. It also

implements both the surface and volumetric

visualization of the internal liver structures. It

receives as input the segmented structures of each

image slice and the thickness of the CT slices

available in the DICOM image file header.

Figure 3 shows an example of segmentation

result produced by the proposed procedure as a 3D

surface which can be visualized within our

prototype. It is possible to observe the hepatic vein

and the portal vein respectively in blue and red, and

the Couinaud segments in different colours are also

present. It can be observed that the Couinaud

segments are divided according to the veins

orientation.

Figure 3: 3D models of segmented liver structures.

Experiments performed on seven different CT

sequences have shown that the results produced by

the proposed method are consistent with the visual

perception of a specialist.

4 CONCLUSIONS

This work proposes an algorithm to segment the

liver in computer tomography (CT) images

according to the Couinaud classification.

Experiments conducted on a software prototype

of the proposed algorithm upon 7 CT produced

results consistent with the visual perception. The

method has the potential of becoming a useful tool

in various applications. It can be used to generate 3D

liver representations to aid visual diagnostic and

surgery planning. Shape attributes other than volume

may also be measured from the 3D model and

explored in Computer Aided Diagnostic

environments.

The assessment of segmentation accuracy is a

major concern in the continuation of this work.

REFERENCES

Couinaud, C. “Le Foie: Etudes Anatomiques et

Chirurgicales. Masson”. Paris. 1957.

Kirbas, C. and Quek, F. “A review of vessel extraction

techniques and algorithms.” ACM Comput. Surv. 36,

2 (Jun. 2004), 81-121. 2004.

Masutani Y., Uozumi K., Akahane M. and Ohtomo K.,

“Liver CT image processing: A short introduction of

the technical elements”, E. Journal of Radiology, Vol

58, Liver Lesions, Pages 246-251. 2006.

Oliveira, D.A.B., Mota, G.L.A., Feitosa, R.Q. and Nunes,

R.A., "A region growing approach to pulmonary

vessel tree segmentation using adaptive threshold".

CompIMAGE proceedings, pp. 319-324, 2006, Porto.

BIOSIGNALS 2008 - International Conference on Bio-inspired Systems and Signal Processing

252