nearly automatic liver vessels segmentation of computed ... · the blood vessels starting from a...

49
Nearly Automatic Liver Vessels Segmentation of Computed Tomography Angiography Patient Scans: Method and Experimental Results A thesis submitted in fulfillment of the requirements for the degree of Master of Science by Miriam Natanzon supervised by Prof. Leo Joskowicz The Selim and Rachel Benin School of Computer Science and Engineering The Hebrew University of Jerusalem Jerusalem, Israel August 16, 2009

Upload: others

Post on 18-Aug-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Nearly Automatic Liver Vessels Segmentationof Computed Tomography Angiography

Patient Scans: Method and ExperimentalResults

A thesis submitted in fulfillmentof the requirements for the degree of

Master of Science

by

Miriam Natanzon

supervised byProf. Leo Joskowicz

The Selim and Rachel BeninSchool of Computer Science and Engineering

The Hebrew University of JerusalemJerusalem, Israel

August 16, 2009

Page 2: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Abstract

Advanced medical image processing is intensively studied to develop a highspeed and high accuracy imaging technology to be integrated throughoutthe clinical procedure starting by the diagnostic stage, onwards to preoper-ative planning and even real-time procedures. Automatic segmentation isconsidered a key task in developing any patient-specific application.This thesis presents a nearly automatic graph-based segmentation methodof the liver vascular system given a CTA image that can be integrated intovarious treatments of numerous liver diseases.

The method begins with liver contour extraction and partial segmentation ofthe blood vessels starting from a single user-defined pixel seed inside the liverarea. Then, the seed area is used for construction of a prior intensity proba-bility distribution function for liver blood vessels. The entire vascular systemof the liver is eventually segmented using a graph min-cut method based ona new edge weights function that adaptively couples the voxel intensity, theintensity prior, and geometric vesselness shape prior.Our method accurately segments the three main liver vessels with theirbranches: the Hepatic Artery, the Portal Vein and the Hepatic Vein. Itrequires only single seed point as user initialization and is completely free ofparameters adjustment. The suggested method is quite fast, it reaches goodcomputation-time on a standard PC (71-153 secs).

Examination of the method was done by two experiments, segmenting overall42 CTA images acquired from two sources taken by different machines. Eachexperiment was evaluated differently: expert radiologist’s evaluation provedour results to be high qualified (PV = 3.273 , MHV = 3.227 in a scale of1 to 4 ), comparison of the results to those generated by a semi-automaticsegmentation tool showed clearly that our method is significantly better thanexisting tools (added value for PV visualization is 3.5 for coloring and 1.75for spreading , in a scale from 1 to 4).

The results show that the method is highly satisfying medical requirementsfrom one hand, and is by far better than the offered implementations fromthe other hand. Other parameters including minimal user intervention re-quirement, good computation time and applicable memory usage simplifiesintegration of this method in various preoperative planning relevant in dis-eased liver treatment procedures.

Page 3: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Acknowledgments

I would like to express my gratefulness to many people that helped me, each

one in his way , along the process of writing this thesis.

First, I thank my supervisor, Prof. Leo Joskowicz, for his helpful and devoted

guiding during all the steps of my work from the implementation stage to

the design of the stucture of this thesis and for being always ready to give a

good advise.

I thank Moti Freiman, a Phd. candidate at Prof. Joskowitz lab that famil-

iarized me with the ITK software and gave great ideas of how to progress.

Also for his help in the evaluation stage.

Dr. Sosna from the Hadassah Medical center and the people from his labora-

tory provided us CTA images, performed accurate work of producing semi-

automatic segmentation and gave us essential information from a medical

point of view.

I thank Naama Lev-Cohen from Dr. Sosna group for her significant contri-

bution in the assessment of the results and for introducing the liver CTA

characteristics and common clinical procedures.

I thank my dear friend, Noah Broide for her friendship along all the way of

the M.Sc. studies and especially for the collaboration on working on this

project. It was all easier when you work with a closest friend..

I thank my parents that have directed me since my childhood to expand my

knowledge in all fields and gave me warmth and love. My whole family: my

grandparents, parents, sisters and brothers that helped me with taking care

of my children during the long hours I have worked in the laboratory.

My children, Yehonatan that got with understanding that ”Mom is in the

ii

Page 4: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

university now”, and Adi that was inseparable part of the process of writing

even before she was born.

Last, but not the least, I would like to thank my husband, Yosef for his love

and understanding. Without your help and support during this period ,this

lines may not be written...

This research was supported in part by MAGNETON grant 38652 from the

Israeli Ministry of Trade and Industry.

iii

Page 5: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Contents

1 Introduction 1

1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Computed Tomography . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Clinical Background . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Method Overview . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.5 Novel Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.6 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Literature Review 10

2.1 Intensity-Based Segmentation Approaches . . . . . . . . . . . 11

2.1.1 Thresholding Methods . . . . . . . . . . . . . . . . . . 11

2.1.2 Region Growing Methods . . . . . . . . . . . . . . . . 12

2.2 Geometric Shape-Based Segmentation Approaches . . . . . . . 13

2.2.1 Circular Object Detection . . . . . . . . . . . . . . . . 13

2.2.2 Matching Filters Methods . . . . . . . . . . . . . . . . 13

2.3 Active Contours . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.4 Statistical Active Shape Models (ASM) . . . . . . . . . . . . . 14

2.5 The Graph-Cut Approach . . . . . . . . . . . . . . . . . . . . 15

2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Method Overview 17

3.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.2 Graph Structure . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.3 Graph Weights . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.4 Min-Cut Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 23

3.5 Shape Criteria: Vesselness . . . . . . . . . . . . . . . . . . . . 23

iv

Page 6: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

4 Experimental Results 27

4.1 First Experiment:

Hadassah Hospital Dataset . . . . . . . . . . . . . . . . . . . . 27

4.2 Second Experiment:

MICCAI 2008 Workshop Dataset . . . . . . . . . . . . . . . . 31

4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5 Conclusions 35

5.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

v

Page 7: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

List of Figures

1.1 Computed Tomography (CT) . . . . . . . . . . . . . . . . . . 2

1.2 Liver anatomy model. . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Liver dieases’s image-guided treatments . . . . . . . . . . . . . 6

3.1 two-mode algorithm results . . . . . . . . . . . . . . . . . . . 20

3.2 Seed area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.3 Graph-cut example . . . . . . . . . . . . . . . . . . . . . . . . 24

3.4 Vesselness map . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4.1 Artifacts examples . . . . . . . . . . . . . . . . . . . . . . . . 28

4.2 Comparison of Hadassah experiment results . . . . . . . . . . 30

4.3 Experiment final results . . . . . . . . . . . . . . . . . . . . . 34

vi

Page 8: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

List of Tables

1.1 HU values range . . . . . . . . . . . . . . . . . . . . . . . . . . 3

3.1 Algorithm and workflow . . . . . . . . . . . . . . . . . . . . . 18

4.1 Hadassah experiment. . . . . . . . . . . . . . . . . . . . . . . 29

4.2 Miccai experiment. . . . . . . . . . . . . . . . . . . . . . . . . 32

vii

Page 9: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1

Introduction

This thesis presents a nearly automatic graph-based segmentation method of

the liver vascular system given a CTA image, in purpose of patient-specific

medical procedures.

Section 1.1 explains the motivation for this work. Section 1.2 provides a

brief description of Computed Tomography (CT) and in particular Computed

Tomography Angiography (CTA). Section 1.3 provides some background for

the anatomy of the liver, possible diseases and their common image-guided

treatments. Section 1.4 presents and overviews the algorithm. Section 1.5

discusses the novel aspects of this thesis. Finally, Section 1.6 presents the

organization of this thesis.

1.1 Background

One of the great challenges of modern medicine is to find alternatives for fully

invasive surgeries in order to reduce the complexity and risks of such proce-

dures and minimize the recovery period as well. Medical image processing

among other computer-based technologies is a key component in achieving

this goal. By presenting an accurate picture of the intra-patient anatomy,

medical image processing has grown to be a useful tool in wide range of

clinical procedures including diagnostics, preoperative planning and image-

guided surgeries.

1

Page 10: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 2

(a) (b)

Figure 1.1: CT scanning. (a) CT scanner. (b) CT slices.

Developing patient-specific techniques is a primary task in the world of

computer-aided surgery and medical image processing research, which leads

to intensive efforts in improving the accuracy, efficiency and automation

level of image-related hardware and software including facing with the inter-

patient high-variance anatomies, while preserving acceptable memory and

run-time performance in order to make them practical for physicians to use.

This thesis presents a nearly automatic method for blood vessels segmenta-

tion. The liver has a complex vascular system composed of three branched

vessels ended with thin capillaries: the Hepatic Artery, the Portal Vein and

the Hepatic Vein. Many medical treatments performed in response to various

liver diseases would be improved by trustworthy knowledge of location and

size of liver vessels from a precise segmentation. For example, ablation of tu-

mors and metastasis, infusion of chemotherapeutic dose and shunt (artificial

channel in the liver) insertion.

1.2 Computed Tomography

CT scanning was introduced by a British inventor Sir Godfrey Hounsfield

who was awarded a Nobel Prize for this great invention. This commonly used

2

Page 11: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 3

Organ HU values rangeBone 1000Liver 40-60White Matter 46Grey Matter 43Blood 40Muscle 10-40Kidney 30Cerebrospinal Fluid 15Water 0Fat -50 - -100Air -1000

Table 1.1: Hounsfield Units (HU) values range for CT. Angio scans (CTA) pro-duce significantly higher HU values for blood 100-300 which highlights blood vesselsover other organs.

3D imaging technique was initially integrated in medical procedures only in

1974. From then until today, CT has grown to be an inseparable part of

various medical treatments including both preoperative and intra-operative

planning and diagnosis in general.

CT scanner (see figure 3.2(a)) is an advanced version of X-ray machine.

X-ray device consists of X-ray source and detector. The source sends out ra-

diation aimed to a specific area of the body that eventually hits the detector.

Different tissue types such as bones, muscles and blood, produce different

rate of radiodensity which is the relative transparency of a substance to X-

ray radiation. Radiodensity is measured using Hounsfield Units (HU), table

1.1 shows typical HU values for CT. The CT scanner involves, in addition,

a translating bed and both source and detector are positioned in opposite

sides on a ring-shaped structure that surrounds the patient. The patient is

lying on the bed that moves gradually through the ring while the ring is

revolving around, therefore, the radiodensity values are recorded for differ-

ent directions. This leads to more accurate and detailed information in each

horizontal slice joined together to 3D image.

Computed Tomography Angiography (CTA) image is obtained by a standard

CT scanner as described above with an additional preliminary step of con-

trast substance insertion in order to provide better identification of blood

vessels. CT images with contrast such as CTA can give a better distinction

3

Page 12: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 4

and highlight specific organs. The contrast substance can be taken by mouth

or injected into an intravenous line.

Complex organs consisting of various tissue types are often required to follow

different CT scanning protocols in order to enhance visibility of the desired

element. The protocols differ in scan timing and infusion rate. Liver scanning

has several objectives. For each, radiologists follow the appropriate protocol

for example veins and arteries in the liver are acquired in different phases

known as ”hepatic phase”, ”portal phase” etc.

1.3 Clinical Background

The liver is a vital organ in the human body. It is located below the di-

aphragm, in the right upper quadrant of the abdominal cavity and its normal

weight for adults is around 1.5 kilogram. The liver has two different sources

for blood supply, the Hepatic Arteries and the Portal Veins which supplies 75

percents of the blood. The Hepatic Arteries supply the liver with oxygen-rich

blood from the hepatic system while the Portal Vein carries venous blood into

it. Unlike most veins, the Portal Vein does not carry blood directly to the

heart but through the liver. The liver processes the nutrient-rich blood which

comes from the gastrointestinal tract via the Portal Vein before it reaches

the systemic circulation and returns back to the heart .In addition, the liver

utilizes the metabolic pool contained in that blood to its functionalities. The

Hepatic veins function as regular veins, the de-oxygenated blood and the

blood supplied by the Portal Vein (to be cleaned in the liver), are merged

into those veins. Then, the merged blood is drained into the inferior vena

cava (IVC) which infuses it into the right atrium of the heart.

Both the arteries and veins split after getting into the liver to left and right

vessels that terminate in very small capillaries after several levels of bifurca-

tions. This creates separate branched trees for the liver blood vessels, as can

be seen in the model of liver illustrated in figure 1.2. The vessels branch-

ing separate the liver into several lobes. Some systems divide the liver into

four lobes while the Couinaud system [2] further divide into a total of eight

sub-segments.

4

Page 13: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 5

Figure 1.2: Liver anatomy model. Image source: [1]

The liver plays a main role in various functions in the body including vital

processes of synthesis and breakdown. The liver converts glucose into glyco-

gen and store it, responsible for cholesterol and protein synthesis, produces

hormones and clean the blood from toxic substances. The liver functionality

is essential, it is involved in numerous biological procedures in the human

body and currently it is impossible to survive in the absence of it.

If the liver becomes inflamed or infected, its ability to perform these func-

tions may be impaired. Liver disease and infections are caused by a variety

of conditions including viral infections, bacterial invasion, and chemical or

physical changes within the body. The most common cause of liver damage

is malnutrition, especially that which occurs with alcoholism.

Cirrhosis and liver cancer are long- term effects of liver disease. Cirrhosis

is the term used to describe a diseased liver that has been severely scarred,

usually due to many years of continuous injury. Cirrhosis causes portal hy-

pertension, high blood pressure in the Portal Vein and its tributaries, which

frequently leads to intestinal bleeding or the buildup of fluid within the ab-

domen.

A TIPS (Transjugular intrahepatic portosystemic shunting) procedure is an

endovascular insertion and placement of an artificial channel (shunt) in the

5

Page 14: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 6

(a) (b)

(c) (d)

Figure 1.3: Liver dieases’s image-guided treatments. (a)-(b) TIPS: Theshunt is used a bypass, connecting the Portal Vein directly to the Hepatic Vein.(c) RFA probe. (d) TACE. Illustration of the chemoembolization proccess. Imagesources: [3]-[6]

liver connecting the portal and the systemic venous system. A TIPS de-

creases the effective vascular resistance of the liver. The result is a reduced

pressure drop over the liver and a decreased portal venous pressure which

prevent from future bleeding to occur within high rate and might make less

fluid develop as well.

Modern liver cancer treatment allows non-surgical ablation of tumors and

metastasis known as Radio Frequency Ablation (RFA). In this procedure a

needle-like RFA probe is placed inside the tumor and the radiofrequency

waves passing through the probe increase the temperature within tumor tis-

sue that results in destruction of the tumor. In cases of unresectable tumors

6

Page 15: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 7

or marginal liver functions, the only solution for complete healing is liver

transplant, but until such is performed (waiting period for liver donation is

usually long) the progression of the cancer must be delayed. Transcatheter

arterial chemoembolization (TACE ) is an interventional radiology catheteri-

zation procedure where a catheter is passed into the proper Hepatic Artery

and onwards to the branches of the Hepatic Artery supplying the tumor(s).

When a blood vessel supplying tumor has been selected, alternating aliquots

of the chemotherapy dose and of embolic particles, or particles containing

the chemotherapy agent, are injected through the catheter.

The method described in this thesis can significantly increase and improve

the use of non-surgical image-guided procedures as described above. Seg-

mentation of liver blood vessels can contribute to patient-specific modeling

for interventional radiology simulation.

1.4 Method Overview

In this thesis we present a nearly automatic algorithm for segmentation of

the liver blood vessels. The algorithm takes an abdominal CTA image as

input and outputs the segmentation result that can be used for accurate

diagnostics or for generating a patient-specific anatomy model.

We use the graph theory’s well-known ’Min-Cut Max-Flow’ theorem as the

main basis of our method. As a first step, the CT image is represented by a

weighted graph where its nodes identify voxels in the original image and its

edges, neighboring relations. The weighting function calculation is based on

both intensity and shape considerations. A seed area that is known to be a

part of the blood vessels is given to the ’graph-cut’ algorithm as an input and

used for initialization purposes by establishing a prior intensity probability

distribution function for the liver vascular system.

Since the graph-cut output is two disjoined groups representing the deviation

of the image to background and object, the relation of a voxel to the desired

segmentation object is defined accordingly.

We begin with finding the liver contour and a minor portion of the blood

7

Page 16: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 8

vessels using Bayesian classification as the major method starting from a

given single user-defined point that can be marked even by not-professional

user [31]. This partial segmentation is functioned as initial seed area for the

’graph-cut’ performing.

Additional functionality we provide is what we call the ’two-mode’ option.

This allows the algorithm to choose the better parameters for the ’graph-cut’

stage considering the high-variance of input’s quality. The proper mode can

be determined before running the algorithm based on prior knowledge about

the image quality or as an accurate enhancing step, re-run the graph-cut

using the other mode.

1.5 Novel Aspects

Most existing software tools used for blood vessels segmentation require user

interaction in various levels (choosing seed points, tuning parameters dur-

ing initialization or throughout the process progressively), we offer a nearly

automated segmentation algorithm where user intervention level is negligi-

ble. The only requirement from the user is to supply a single point located

in the liver area, but even in cases of wrong initialization, the algorithm

would convergences into the right direction. This applies that our method is

easy-to-use even to non-expert user.

Our method deals successfully with the high variance of liver CTA images.

Low contrast is a typical situation in the input images that may be caused due

to pathological phenomena such as Cirrhosis or due to the scanning situation,

and makes the task of blood vessels detection very challenging. We suggest a

multiple-mode algorithm, allowing parameters adjustment according to the

image quality, done in reasonable computation time.

The method collects the information about the intensity values range from

the image itself and does not rely on earlier assumptions acquired from train-

ing set or based on medical literature. In result, we achieve stability and ro-

bustness i.e. dealing with the high-variance of image qualities differing from

one machine to another and applying this method to segmentation of blood

8

Page 17: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 1. Introduction 9

vessels in other organs.

In order to asses the proposed method we deliberately use two different eval-

uation techniques, the accuracy and contribution of the segmentation is ex-

amined in comparison to both the ideal and reality, medicine’s expectations

and existing industrial tools. The medical qualitative evaluation relies on

physicians expertise while the industrial comparison evaluation checks the

method’s contribution beyond other existing segmentation tools available in

the market. This two-scaled, wide range evaluation provides a better per-

spective of judgment.

The results of this thesis have submitted and accepted to the 2009 RSNA

conference [7].

1.6 Thesis Organization

This thesis consists of five chapters. Chapter 2 is a survey of previous work

in the field. Chapter 3 presents the proposed segmentation method. Chapter

4 describes our experiment, the evaluation technique and its results. Chap-

ter 5 concludes with a summary of our contributions and examines further

directions for future work.

9

Page 18: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2

Literature Review

There are many automatic and semi-automatic vascular structure segmenta-

tion methods [8]. The main approaches rely on intensity values, geometric

shape, edge-based active contours and statistical active shape models. The

segmentation algorithms often yield inaccurate vessel diameters, miss ves-

sels segments and entire small vessels, and include non-vessel anatomical

structures and incorrectly model vessel bifurcations and pathologies. These

failures are often due to the local vessels appearance and characteristics vari-

ability in various regions of the body. In addition, most methods require

extensive user interaction and the adjustment of non-intuitive parameters,

which difficult their routine clinical use.

This chapter reviews previous works on vessels extraction. Section 2.1 and

Section 2.2 describe in detail the two main approaches: intensity-based and

geometric shape methods. These are the groundwork of our method de-

scribed in this work. Section 2.3 and 2.4 review other significant segmentation

methods. Section 2.5 presents previous applications usage of the graph-cut

method. Finally, section 2.6 summarizes the previous methods and applies

our new method’s benefit over them.

10

Page 19: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2. Literature Review 11

2.1 Intensity-Based Segmentation Approaches

Intensity-based segmentation methods rely on the assumption that different

anatomies can be characterized by different Hounsfield unit (HU ) values.

There are many vessel segmentation methods of this kind; they utilize the

advantage of the CTA that sharpens the contrast between blood vessels and

other anatomies in the image.

2.1.1 Thresholding Methods

Thresholding is a simple but powerful technique. It yields segmentation over

an image by simply determining intensity values as threshold. The parti-

tion is done according to this threshold; all pixels within the same range are

grouped together. It does not take in consideration any spatial characteristic

and requires the choice of the threshold value. Its simplicity is a great ad-

vantage, which tempts many to use this, although its limitations which are

high sensitivity to noise, intensity in-homogeneities and to high variance of

areas of classes to be segmented. Therefore, it is usually not used as ’stand-

alone’ method but integrated with other methods. Most common usage of

this method is as pre or post processing such as noise removal and initial

filtering.

Higgins et al [9] and Niki et al [10] both perform a pure thresholding method

followed by connected component analysis. The main advantage of these

algorithms is their excellent run-time performance. Though, they are not

suitable enough to all anatomies, inaccurate results are expected when per-

formed on areas with high variance intensity values. Other works integrated

the basic thresholding method within the general process of their more com-

plex algorithms. Kim et al [11] use thresholding as a core action in its

algorithm. First, a histogram of the intensity values is used to initialize the

threshold value. Second, thresholding is performed locally on each slice of

the image, and determined adaptively. The Region of Interest (ROI ) in the

next slice is adjusted according to the segmentation result of the previous

slice. This algorithm utilize the benefits of thresholding as described above

(’fast and simple’), it overcomes thresholding’s main limitation by taking

11

Page 20: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2. Literature Review 12

into consideration spatial proximity of pixels and avoids the need to choose

the threshold value by adaptive determination of this value, defined by a

histogram.

2.1.2 Region Growing Methods

Segmentation is performed based on two important criteria, value similarity

and spatial proximity. Two pixels can be grouped together if they have the

similar intensity characteristics or if they are physically close to each other.

It is assumed that pixels that are closed to each other and have similar

intensity values are likely to belong to the same object. The process starts

from a single or many seed points known to be part of the desired region, this

requires prior knowledge. The main disadvantage of region growing methods

in the context of vessel segmentation is that it can result in holes and over

segmentation due to the variations in image intensities (such as variations

caused by stents, made of metal or plastic, that are inserted into the vessel

in order to keep it open) and noise (such as nearby anatomical structures

with similar intensity range). Thus, it sometimes requires post-processing of

the segmentation result.

Revol-Muller et al [12] describe an advanced method of region-growing. They

performed an iterative process of region growing, increasing the threshold

value in each step. Eventually an assessment function is executed to de-

termine the most suitable result from the previous stage (region growing).

In their method they overcome some drawbacks of region growing. For in-

stance, not using a fixed threshold value helps avoiding the problems caused

by variations in image intensities, but this does require a longer computation

time.

12

Page 21: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2. Literature Review 13

2.2 Geometric Shape-Based Segmentation Ap-

proaches

Geometric features of the desired anatomy drive shape-based segmentation

methods. Vessels segmentation algorithms will try to recognize tubular struc-

ture characteristics.

2.2.1 Circular Object Detection

Tubular structure can be identified by a sequence of circles in 2D planes. A

common method for this purpose is the ’Hough Transform’. The standard

Hough transform to locate round objects employs an edge detector. The

resulting edge information is used to find candidate center locations. Finally,

candidate center locations are averaged to obtain an estimate of the position

of the object’s center. In the 3D image result, the tubular structure will

be seen due to the sequential circles, found by Hough transform. The main

obstacle is that there might be nearby, non-vessel, tube-like anatomies and

the distinction between them is not an easy task. [13]

2.2.2 Matching Filters Methods

Matching filters convolves the image with multiple matched filters for the

extraction of objects of interest. In extracting vessel contours, designing

different filters to detect the vessels with different orientation and size plays

a crucial role. An orientation detection filter can use the eigenvalues of the

Hessian matrix i.e. second derivative eigenvalues to determine the likelihood

of a voxel to be a part of a vessel [14, 15]. Matching filters are usually followed

with some other image processing operations like thresholding and connected

component analysis to get the final vessel contours, to avoid segmentation of

any other tubular structured anatomies that may be extracted as well within

pure matching filter. In previous works [14, 15], a significant advantage is the

ability to cope with varying width of vessels, a common scenario e.g. in case

of segmentation of different vessels or different pathologies such as stenosis

13

Page 22: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2. Literature Review 14

and aneurysms. But, this requires re-computation of the second derivative

values with multiple scales allow for different radius. There is a tradeoff

between accuracy and run-time performance; the exact extraction may be

problematic since in general the number of scales needs to be small to reduce

the computational cost.

2.3 Active Contours

Active contours (’Snakes’) is a model-based technique to find object contours

using parametric curves that deform under the influence of internal and ex-

ternal forces [16]. Physically, a snake is a set of control points in an image

that are connected to each other. Each one has an associated energy that

either rises or falls depending upon the forces that act on it. These forces are

known as snake’s internal and external forces, respectively. Internal forces

serve to impose smoothness constraints on the contour while external forces

pull the snake towards the desired image features like lines and edges. Vari-

ous vessels segmentation methods use the active contour technique and differ

in the energy function they define. Lorigo at al. [17] present an edge-based

active contour method. Its energy criterion is based both on intensity val-

ues and on local smoothness properties of the object boundary (vessel wall).

Nain at al. [18] combine image statistics and shape prior in their energy

function.

Active contour’s main disadvantage is that it usually requires user interac-

tion to initialize the snake. [19, 20] deal with this lack by automating the

initialization process. Additionally, Active contour is problematic in its ex-

cessive computing time and memory requirements. In general, active contour

is better for segmentation of local objects since snakes cannot move toward

objects that are too far away from the initial point.

2.4 Statistical Active Shape Models (ASM)

Statistical active shape models (ASM) [21] are statistical models of the shape

of objects which iteratively deform to fit to an instance of the object in

14

Page 23: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2. Literature Review 15

a new image. The model is defined by a training set of segmented data,

its segmentation is represented by landmarks of the anatomy. The ASM

algorithm aims to match the model to a new image. The main drawback

is the training-set production, it requires involvement of an expert in the

field. Furthermore, strong noises in many medical images usually introduce

false edges that may trap the shape instance to incorrect locations. Some

techniques try to address this limitation. Lekadir et al [22] uses statistical

shape metrics based on inter-model landmark-based distances to revaluate

the error between the segmentation result of the instance and the model.

Another method that tries to avoid cases of wrong segmentation is the AAM

(Active Appearance Models) [23]. This method use the training set images

to estimate the relationship between model parameter displacements and the

residual errors induced between an image and an assignment in the model.

In order to segment a new image, AAM measurements are used to predict

changes to the current parameters that achieve better results. However,

AAM requires high-quality training images which can be very expensive,

while the former [22] does not consider the wealth of information provided

by the parameterized shape model.

2.5 The Graph-Cut Approach

A promising approach is the graph min-cut segmentation method [24, 25].

It classifies the voxel nodes that separate the objects of interest and the

background based on weighted voxel adjacencies. The advantages of graph

min-cut segmentation are that it is generic and that it is nearly parameter-

free. However, it relies on significant user interaction and requires fine-tuning

of the vessels intensity priors to capture the small vessels variability. It does

not incorporate vessels geometric information, is computationally intensive,

and has extensive memory requirements.

Recent improvements to the graph min-cut interactive segmentation method

address some of these drawbacks. Slabaugh and Unal [26] add an elliptical

shape prior term to the edges cost function that improves the segmentation

results. Though, in their work, this is done only on 2D images. Sinop and

15

Page 24: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 2. Literature Review 16

Grady [27] use a Laplacian pyramid to accelerate the segmentation and to

reduce the memory requirements. Ning et al. [28] improve the segmenta-

tion with graph-cut active contours based on prior object surface estimation.

Rother et al [29] reduce user interaction with user-defined enclosing rectan-

gular regions around the objects of interest. While this method is useful for

extracting simple objects in 2D images of natural scenes, it is laborious for

3D images of complex vascular structures. A common key drawback of these

methods is that their generic segmentation framework is often ill-suited for

volumetric vessels segmentation.

2.6 Summary

The reviewed methods offer different solutions to patient-specific segmen-

tation task. The challenge is to satisfy the many competing goals such as

accuracy, robustness, automation and efficiency. Each solution answers the

requirements in various levels and tries to overcome the contradiction be-

tween them.

The task presented in this thesis is a patient-specific application that can be

used as a pre-operative tool for treatment planning purposes. Therefore, it

should be fast and easy to use with the least user intervention as can be,

cope with high variability of input as in real-world data and most important,

produce an accurate and reliable result.

This thesis presents a graph-cut based method for vessel segmentation. The

core of the graph-cut process is the cost function. We define it as an inte-

grated function of both local intensity estimation and vesselness value (ves-

selness calculation is based on [15]). This overcomes weakness points of each

criterion when used as stand-alone. It is responsible for the method’s robust-

ness as well.

In contrast to previous works, our method finds the seed area almost au-

tomatically, therefore, it only requires negligible user interaction for initial-

ization. This trait makes it practical and compatible with patient specific

applications.

16

Page 25: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3

Method Overview

This chapter describes an automatic graph based segmentation method of the

liver vascular system. The method uses min-cut algorithm to vary between

the object (liver blood vessels) and the background. It utilizes the prior

characterization of blood vessels in CTA image, i.e. intensity range and

shape.

The contents of this chapter are as follows. Section 3.1 describes the al-

gorithm in general. Section 3.2 describes the graph structure. Section 3.3

explains how the graph weights are determined. Section 3.4 provides an

overview of the min-cut algorithm. Finally, Section 3.5 describes into details

the shape criteria calculation.

3.1 Algorithm

Given an abdominal CT image with injected contrast liquid (CTA), we pro-

pose a nearly automatic segmentation algorithm to extract the liver’s blood

vessels. The main idea of this algorithm is to distinguish between the object

voxels and the background voxels, where the object is the desired segmenta-

tion area. To do so, we use the graph-cut algorithm.

Initially, we need to build the graph which represents the input image. The

nodes in the graph represent voxels in the image, in addition there are two

17

Page 26: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 18

Input: a CTA image of the abdominal area.Output: Binary image which segments the vascular structure in the liver.

Main method: ’graph-cut’

1. Build a graph that represents the image.

(a) Each pixel is a node in the graph.

(b) Two dummy nodes - object and background.

(c) Edges from each pixel to it’s six immediate neighbors.

(d) Edges from each pixel to the dummy nodes.

2. Calculate edge weights. Based on intensity and vesselness values.

3. Perform ’min-cut’ algorithm.

Workflow:

1. Extract the Region of Interest - The liver surface.

2. Identify partial area of blood vessels.

3. Create the vesselness map for the Region of Interest (ROI ).

4. Run ’graph-cut’ algorithm on the Region of Interest, use the preliminaryvessel segmentation result (from step 2) as seed.

5. Noise removal - choose the largest connected components.

Table 3.1: Segmentation of liver blood vessels: Algorithm and workflow -schematic view.

18

Page 27: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 19

dummy nodes representing the object and background accordingly. Edges

represent connections in the image. Every two neighbor voxels in the image

have a connecting edge between them. This applies all three axes; therefore,

each node will have six neighbors. Two additional edges are drawn from each

node in the graph: one to the object node and the other to the background

node. These edges represent the probability of a voxel to belong to the

object or background correspondingly. Weights are assigned to the edges as

a combined function of intensity values and shape characterization. Now, we

perform the graph-cut algorithm which outputs the desired partition of the

image nodes to object nodes vs. background nodes.

Liver CTA images often differ in their quality. The reasons are various,

including pathological cases such as sick patient (high level of cholesterol)

and different quality of CT machines. To overcome this common obstacle

we offer a two-mode procedure of the graph-cut algorithm using different

radius values for the vesselness map calculation. In general, this allows a

more precise segmentation as can be seen from figure 3.1.

As a preliminary step, we find simultaneously the liver contour and a small

portion of the liver vessels in the CTA (see figure 3.2 for examples). For this

purpose we use a nearly automatic algorithm for liver analysis, described

in details in [31]. This method was examined on various datasets [32] and

was graded within best segmentation results. Method seemed to be robust

and accurate and as consequence is well-suited for the purpose of finding

initial good seed area. The process is based on multi-resolution iterative

approach. It requires single user-defined pixel seed inside the liver that can be

marked even by non-expert user since the method is robust enough and copes

with not-accurate initial point. The algorithm starts with generation of an

intensity model describes the distributions of the Hounsfield units values for

liver and other organs in the abdominal CTA dataset. Based on this model,

each voxel is classified as Liver parenchyma, tumor, vessel, or background

using the Bayesian classification method. Adaptive morphological operators

are applied to identify the liver, fill holes inside the liver and disconnect

it from surrounding organs. Finally, active contours method is applied to

refine the liver surface. Then, the algorithm repeats those steps- generate

19

Page 28: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 20

(a)

(b) (c)

Figure 3.1: Illustration of the two-mode algorithm. (a) The grayscaleinput. In this case the segmentation task is very challenging due to the poorquality of the CTA image. (b) The first mode result. Segmentation is noisy. (c)The second mode result. Segmentation is more accurate.

again the intensity model and the histogram and so on, according to the

knowledge acquired from the process. The process continues to iterate until

convergence, when no changes to the liver surface are detected.

The vesselness map is generated for the Region of Interest - the liver area

extracted in the previous stage to retrieve the shape information per voxel.

Voxels within the seed area found in the first step are assuredly object vox-

els. We use this preliminary knowledge within the graph-cut algorithm in

order to characterize the object’s intensity features (i.e. mean and standard

deviation).

20

Page 29: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 21

(a) (b)

Figure 3.2: Seed area. Partial segmentation of the blood vessels serves as aseed area to the ’graph-cut’ algorithm

3.2 Graph Structure

We use the graph-cut method to distinguish between the ’object’ (vascular

tree) and the ’background’. All image voxels are nodes in the graph. The

graph contains two additional dummy nodes representing the ’object’ and

the ’background’. An edge is ”drawn” between each voxel to each of its six

immediate neighbors: two in the x axis, two in the y axis and two in the z

axis and to each of the dummy nodes: one to the ’object’ node, and another

one to the ’background’ node.

Formally, let graph G = (V,E) where V = {vi, s, t} are the voxels nodes (vi)

and object (s) and background (t) nodes, and E = {(vi, vj), (vi, s), (vi, t)} for

each voxel vi and its 6-neighbor voxels (vj).

3.3 Graph Weights

Edge weights are divided into two kinds in accordance to the edge types:

(a) The edge weight between a node to its neighbor is an intensity based

normalized ’distance’ function. In this manner, high distance value will re-

21

Page 30: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 22

sult with low weight value and vice versa. (b) The vascular structure can

be characterized by two attributes: typical intensity range of blood vessels

(˜100–300 HU) and tubular structure (vesselness, based on Frangi’s formula

[15]).

Since the role of an edge weight between a node to a dummy node is to

evaluate the relevance of a voxel to the object (vascular tree) or background,

the weight calculation is a combined function of both voxel’s intensity and

vesselness.

We use the mean and standard deviation values computed from the seed area

which is known to be an object.

Let:

µobj = mean value of the object intensity values

σobj = standard deviation of the object intensity values

Formally, for each (vi, vj) ∈ E where vi and vj are adjacent, the weight

function W is:

W (vi, vj) = exp

(−(

(I(vi)− I(vj))

σ

)2)

(3.1)

where σ calculation is based on σobj. (Tuned according to the voxel vesselness

value, as far as the vesselness value is high , we let a wider variance.)

For each (vi, s) ∈ E, the weight function W is:

W (vi, s) = exp

(−(

(I(vi)− µobj)

σobj

)2)M(vi) (3.2)

where M(vi) is the vesselness value.

For each (vi, t) ∈ E, the weight function W is the complement of the above

function:

W (vi, t) = exp

(−(

(I(vi)− µobj)

σobj

)2)

(1−M(vi)) (3.3)

22

Page 31: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 23

3.4 Min-Cut Algorithm

Cut C is defined as a subset of edges such that the removal of C from the

graph G separates the source from the target nodes, in our case, the two

dummy nodes s and t. A cut defines two disjoint subsets, the object set O

and the background set B, such that s ∈ O and t ∈ B. The cost of a cut is

defined as the sum of weights of the cut edges e ∈ C:

|C| =∑e∈C

We (3.4)

The optimal cut is the minimum-cost cut. It can be efficiently computed

using standard graph theory algorithms. The theorem of Ford and Fulkerson

[30] states that the minimum-cost cut can equivalently be calculated with the

Max-flow algorithm. The idea behind Ford and Fulkerson algorithm is very

simple. As long as there is a path from the source (s ∈ O in our graph) to the

sink (t ∈ B), with available capacity on all edges in the path, we send flow

along one of these paths, which means we decrease the edge weights along

the path that produces the residual graph. Then we find another path, and

so on. A path with available capacity is called an augmenting path. When

no more augmenting paths can be found, s will not be able to reach t in the

residual graph. If S is the set of nodes reachable by s in the residual graph,

then the total capacity of edges from S to the remainder of all the nodes in

the (original) graph (V − S) is on the one hand equal to the total flow we

found from s to t, and on the other hand serves as an upper bound for all

such flows. This explains why the flow we have found is maximal.

3.5 Shape Criteria: Vesselness

One of the characteristics of a vascular structure is its tubular shape. We use

a Hessian-based measure in order to determine voxel’s relevancy to a vessel

like structure at a given scale. The measure describes the geometric structure

of vessels by analyzing the eigensystem of the 3×3 Hessian matrix H of each

voxel from the volumetric image. For bright vessels on a dark background,

23

Page 32: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 24

Figure 3.3: Graph-cut example. Illustration of a 2D image graph and its cut.

the following shell exists:

λ1 ≈ 0 (3.5)

λ1 � λ2

λ2 ≈ λ3

The eigenvector associated with λ1 is the direction along the vessel, therefore

we expect its eigenvalue to tend to zero. The eigenvectors associated with λ2

and λ3 are perpendicular to λ1. Therefore, we expect λ2 and λ3 to be greater

then λ1, with a low difference.

24

Page 33: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 25

Figure 3.4: Vesselness map. Shape-based probability map of each voxel to belongto the object according to Frangi’s vesselness measure.

We use Frangi’s vesselness measure [15]:

(3.6)

M(σ) =

0 λ2 > 0 or

λ3 > 0,1−exp

− R2A

(2a)2

exp

− R2B

(2b)2

1−exp

− S2

(2c)2

otherwise

where:

RA =|λ2||λ3|

, RB =|λ1|√|λ2λ3|

, S =√λ2

1 + λ22 + λ2

3

and σ is the scale at which the measure is computed. RA differentiates be-

tween plate and line-like structures, RB measures the deviation from blob-like

structures, and S differentiates between foreground (vessel) and background

(noise). Constants a, b and c are predefined weights determining the influ-

ence of RA, RB and S. The vesselness measure value is close to 1 for voxels

with tube-like structures, and close to 0 otherwise. The vesselness measure

is computed for each voxel and for several scales. The maximal value among

of the different scales is chosen as the vesselness of the current voxel. This

25

Page 34: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 3. Method Overview 26

measure behaves like a probability map which maps the probability of each

voxel to be a vessel. Figure 3.4 illustrates the vesselness response overlaid on

the original datasets. The colors are for visualization purpose only.

26

Page 35: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4

Experimental Results

We implemented our segmentation method using a powerful open-source

toolkit known as ITK (Insight Toolkit) [33] which implements state-of-the-

art algorithms in medical image processing and analysis. To visualize the

CTA images and their segmentation results we used ITK-Snap [34], a very

user-friendly software for presenting 3D images. We have tested our seg-

mentation method on a series of 42 abdominal CTA images taken from two

different datasets. The first dataset was taken from the radiology depart-

ment at Hadassah hospital from Dr. Sosna’s lab. The second dataset was

provided by the MICCAI 2008 3D Liver tumors segmentation workshop [32].

The results evaluation was done differently for each dataset.

The contents of this chapter are as follows. Section 4.1 describes the exper-

iment and the results of the Hadassah CTA images. Section 4.2 describes

the experiment and the results of the MICCAI CTA images. Section 4.3

discusses and summarizes the experimental results.

4.1 First Experiment:

Hadassah Hospital Dataset

The dataset contains twenty scans taken from ten patients in age range of

32 to 79 years and average age of 54 years, consist of both hepatic phase

27

Page 36: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 28

(a)

(b) (c)

Figure 4.1: CTA images with artifacts. (a) Cholesterol accumulation yieldspoor contrast. Whiter pixels can be seen on the top right edge. (b) A stent isdetected as high intensity area. (c) Tumors and metastasis are detected as blackholes.

and portal phase images for each patient. The scans were acquired on a 64-

channel detector scanner (Brilliance 16, Philips Medical Systems, Groningen,

The Netherlands) with 2 mm slice thickness,1 mm increment, 200-250 mAs

and 120 kVp. These CTA images had 512 × 512 × ˜400 (The Z axis vary

between the images, the average size is around 400) voxels. The clinical

findings of these scans included healthy livers and livers with various diseases

such as cirrhosis, metastases and primary liver tumors. Artifacts in the CTA

image can result from unhealthy livers such as these i.e. tumors, stents

and fatty liver, or from circumstantiated factors revolving around the scan’s

timing. See figure 4.1 for examples.

28

Page 37: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 29

Portal Vein Hepatic Veinmain vessel coloringquality spreading main vessel coloring

quality spreading artifacts

3.5 1.75 2.25 2.25 0.75

Table 4.1: Hadassah experiment. Average results of segmentation comparisongrading. The first two columns present scores of both coloring and spreadingmeasures for the Portal Vein (PV). The third and fourth columns are the samefor the Hepatic Vein (HV). The fifth column presents artifact rate for the entireimage.

We compared the results of our method to those obtained from an extrac-

tion of the liver vessels performed by a dedicated technologist using a semi

automatic paintbrush tool in a Philips EBW3.0 workstation, (Philips Medi-

cal Systems, Groningen, The Netherlands), which serves as a representative

example of the segmentation abilities currently present. This comparison

was done by an expert radiologist. The evaluation consists of the following

measures:

Coloring quality: on a scale from 0 to 4, where 0 indicates similar coloring

quality between the representative example and the segmentation result

and 4 indicates significant improvement of the segmentation over the

representative example. Coloring quality measure takes into account

both accuracy and consecutiveness.

Spreading level: on a scale from 0 to 4, where 0 indicates similar spreading

level between the representative example and the segmentation result

and 4 indicates significant improvement of the segmentation over the

representative example. Spreading level indicates till which peripheral

location the segmentation has reached.

Artifacts rate: on a scale from 0 to 4, where 0 indicates the segmenta-

tion has no additional artifacts over the representative example and

4 indicates the high rate presence of artifacts in the segmented image

compared to the representative example.

Table 4.1 summarizes the evaluation of our results. In order to give a more

specific and useful evaluation, the first two measures, coloring quality and

29

Page 38: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 30

(a)

(b) (c)

Figure 4.2: Comparison of Hadassah experiment results. (a) The input -2D slice grayscale image. (b) and (c) are the semi-automatic tool and our methodoutputs correspondingly. Note the high coloring quality and spreading level in(c).However, the segmentation in (c) suffered from the presence of artifacts.

spreading level referred to the Hepatic and Portal Vein separately. Artifacts

rate assessment referred to the entire segmented region. Since our automatic

segmentation generated more detailed results, the lowest grading was equiva-

lent to similarity of the two, any higher scoring proves our results to be better

and more accurate. Note that we accomplished significant improvement in

both coloring quality and spreading (e.g. coloring quality average for PV is

3.5 and for HV is 2.25) while the artifacts rate is negligible (average of 0.75).

30

Page 39: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 31

4.2 Second Experiment:

MICCAI 2008 Workshop Dataset

The dataset contains twenty two portal phase scans gathered for the purpose

of liver tumor detection competition organized by MICCAI [32]. As such,

they are representative sample of real world cases. The scans were acquired

on one 64-channel detector and two 40-channel detector CT scanners using

a standard four-phase contrast enhanced imaging protocol with 1-1.5 mm

slice thickness and in-plane resolution of 0.6-0.9mm. These CTA images had

512 × 512 × ˜350 (The Z axis vary between the images, the average size is

around 350) voxels.

The evaluation method used in this experiment is the more common one.

An expert radiologist determines the grading according to his experience in

the field. The radiologist assessed the visibility of individual vessels in the

3D segmentation images in comparison to the original CT images using the

following measures: zero indicates poor visualization and four indicates excel-

lent visualization. Eight main liver venous vessels were examined including

Portal Vein branches as well as the three Hepatic Veins.

Table 4.2 summarizes the results for those cases and the overall results. Note

that our method successfully segmented the LPV and LHV (average grading:

3 and 3.045 correspondingly) despite their relative low contrast in the original

CT images.

4.3 Summary

The two experiments described above differ in both their dataset source

(Hadassah hospital, MICCAI workshop) and the used evaluation method.

Each evaluation method examines the contribution of our algorithm in a dif-

ferent aspect. In the first experiment, the evaluation procedure was based on

a comparison between our segmentation results to such generated by an up-

to-date industrial tool. From this experiment we conduct that our method

gives more accurate and detailed results which is presumably a great advan-

31

Page 40: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 32

RHV MHV LHV PV RPV LPV RAPV RPPVMICC-01 4 4 4 4 4 4 4 4MICC-02 1 1 1 1 1 1 1 1MICC-03 4 4 4 4 4 3 4 4MICC-04 1 1 1 4 4 4 4 1MICC-05 2 2 2 4 4 4 4 4MICC-06 1 1 1 1 1 1 1 1MICC-07 4 4 4 4 4 4 4 4MICC-08 4 4 4 4 4 4 4 4MICC-09 3 3 3 4 4 4 2 2MICC-10 4 4 4 4 4 1 4 4MICC-11 4 4 4 4 4 3 4 4MICC-12 4 4 4 4 3 4 2 2MICC-13 4 2 1 4 4 4 4 4MICC-14 4 3 1 1 4 1 2 2MICC-15 4 4 4 4 4 4 4 4MICC-16 4 4 4 4 4 3 4 4MICC-17 3 3 2 3 3 3 3 3MICC-18 4 4 4 4 4 4 4 4MICC-19 1 3 2 1 3 2 2 2MICC-20 4 4 4 4 4 4 4 4MICC-21 4 4 4 1 1 1 1 1MICC-22 4 4 4 4 3 4 4 4Average 3.273 3.227 3 3.273 3.409 3.045 3.182 3.045

Table 4.2: Miccai experiment. Evaluation grading for 22 CTA scans. Dividedinto the eight main liver venous vessels: (1) RHV - Right Hepatic Vein (2) MHV- Middle Hepatic Vein (3) LHV - Left Hepatic Vein (4) PV - Portal Vein (5) RPV- Right Portal Vein (6) LPV - Left Portal Vein (7) RAPV - Right Anterior PortalVein (8) RPPV - Right Posterior Portal Vein. The scoring range is 0 to 4.

tage over the existing tools presented up until now. In the second experiment,

the evaluation was done by a physician grading our results according to his

knowledge and experience concerning medicine’s requirements and needs.

The high scores we achieved in this experiment indicate that our method

fulfils the medical expectations above and beyond.

The sufficient results described in Table 4.1 and Table 4.2, were generated

with good performance. The mean computation time on a standard PC

dual-core 2.4GHz with 3GB of memory machine was 94.15 secs (std=33.1

secs).

The experiment’s input contains various types of datasets which covers high

32

Page 41: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 33

range of different phenomena, as can be met in real-world clinical proce-

dures. In all cases, we have accomplished high level results in reasonable

computation time and acceptable memory usage.

This qualifies our method to be easily integrated in a preoperative planning as

it meets the physician’s requirements: nearly automatic, fast enough response

time and applicable with standard resources.

33

Page 42: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 4. Experimental Results 34

(a) (b)

(c) (d)

(e) (f)

(g) (h)

Figure 4.3: Experiment final results. In the left column, the 2D results aredisplayed. The right column presents the corresponding 3D segmentation.

34

Page 43: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 5

Conclusions

This chapter concludes the thesis. Section 5.1 summarizes the background,

goals, and contributions of this thesis. Section 5.2 suggests directions for

improvements and future work.

5.1 Summary

This thesis presents a nearly automatic method for liver blood vessels segmen-

tation from CTA images that can highly improve the preoperative planning

process in clinical treatment for various liver diseases.

Many previous works suggests wide variety of vascular structure segmenta-

tion techniques. The existing methods don’t offer a sufficient solution for the

purpose of automatic or nearly automatic segmentation of the liver blood

vessels that can be merged into a clinical routine. Some of the methods

don’t cope successfully with the significant inter-patient liver geometry and

intensity variances or the artifacts of many kinds. Other methods are not

applicable to clinical situation because they require extensive user interven-

tion in selecting many seed points or tuning parameters or due to long run

time and expensive resources.

The segmentation method suggested in this thesis is based on the ’graph-

cut’ algorithm which is known to be cheap in the term of run time. It

35

Page 44: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 5. Conclusions 36

takes into account both intensity and shape considerations and looks for

vessels with different radius. This yields that our method produces impressive

segmentation results of the entire liver vascular system including all three

main vessels, the Hepatic Artery, the Hepatic Vein and the Portal Vein, and

their branches, all done in a reasonable time.

We examined the method on overall 42 ”real-world” CTA images acquired

from two sources, using different machines with different qualities. The im-

ages create a representative model reflecting the variance in liver CTA images:

different anatomies, presence of pathologies and artifacts.

We used two different evaluation methods in order to examine the contri-

bution of our method in two different aspects: the ideal objective from a

medical point of view and the added value to existing tools in the industry.

We achieved high quality results that answer the strict requirements of ac-

curacy and reliability as medicine states. In addition, our method proved

to be significantly better in comparison to a widely used segmentation tool

assumed to be one of the best in the field.

In conclusion, we present an accurate, robust, easy-to-use segmentation method

that can be integrated into image guided procedures such as liver tumors

treatments (e.g. chemoembolization, radiofrequency).

5.2 Future Work

Since automatic blood vessels segmentation is relevant in various patient-

specific applications for different anatomies, the demand for such is con-

stantly increasing by the medical community. These segmentations can be

used as a preliminary step for producing an input to training simulators

or as an accurate estimate of situation report detailing critical information

such as size and location of blood vessels, required before taking an invasive

procedure e.g. in brain surgeries.

Therefore, the main direction for possible future work is to apply this segmen-

tation method on blood vessels in other anatomies besides the liver’s vascular

system. First attempt in this direction has already been taken; our method

36

Page 45: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Chapter 5. Conclusions 37

segments successfully the carotid arteries and its full vascular tree [35]. As

we expected, we can use our core method with few minor adjustments for

other blood vessels segmentations, this is possible thanks to the robustness

and generalization of our method and its nearly complete independence of

parameters. Any implementation of our method for other organs would re-

quire some adjustments and modifications i.e. anatomy-related features like

seed area and range of radius sizes must be customized accordingly. Further-

more, it is foreseen that CTA images of separate anatomies will challenge our

method with their own unique features, different common pathologies and

artifacts. Other vessels that we plan to segment with this method are the

Abdominal Aorta for detection of Aneurysm (known as Triple A or AAA)

and brain vessels.

Another direction for future work is to achieve accurate deviation of the liver

according to Couinaud ’s segment model [2] based on this liver blood vessels

segmentation.

We hope that this method will assist the influential goal of improvement of

the clinical pre-operation step and increase treatment’s accuracy in surgeries

from many kinds.

37

Page 46: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Bibliography

[1] Collin County Community College District (CCCD), Dr. Dr. Mary Weis

site. [Online]. Available: http://iws.ccccd.edu/mweis

[2] C. Couinaud , A. Delmas, and J. Patel ”Le foie; Etudes Anatomiques

et Chirurgicales”. Masson. Paris 1957.

[3] AngioDynamics company site. [Online].

Available: http://www.angiodynamics.com

[4] A.D.A.M. Medical Illustration Team. [Online].

Available: http://www.nlm.nih.gov/medlineplus/ency/

imagepages/19859.htm

[5] The University of Texas, M. D. Anderson Cancer Center. [Online]. Avail-

able: http://www.wentek.com/jean/ablation.htm

[6] Fuda Cancer Hospital ,Guangzhou , Jappan. [Online].

Available: http://www.orienttumor.com/pages/products/

transjugularaccess_products.asp

[7] N. Lev-Cohen, M. Natanzon, N. Broide, M. Freiman, L. Joskowitz, L.

Appelbaum, P. Berman, J. Sosna, ”Minimal Cut Method for Liver Vessel

Segmentation from MDCT: Algorithm Assessment and Verification,”

the 95th Scientific Assembly and Annual Meeting, RSNA Radiological

Society of North America ,2009.

[8] C. Kirbas and F. Quek ,”A review of vessel extraction techniques and

algorithms,” ACM Computing Surveys 36(2) 81-121, 2004.

38

Page 47: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Bibliography 39

[9] W.E Higgins, W.J.T Spyra, and E.L Ritman, ”Automatic extraction of

the arterial tree from 3-d angiograms,” in IEEE Conference Engineering

in Medicine and Biology Society vol. 2. 563-564, 1989.

[10] N. Niki, Y. Kawata, H. Satoh, and T. Kumazaki, ”3d imaging of blood

vessels using x-ray rotational angiographic system,” in IEEE Conference

Nuclear Science Symposium and Medical Imaging vol. 3 1873-1877, 1993.

[11] D. Kim and J. Park, ”Connectivity-based local adaptive thresholding

for carotid artery segmentation using MRA images,” Image and Vision

Computing. 23(14) 1277-1287, 2005.

[12] C. Revol-Muller, F. Peyrin, Y. Carrillon, and C. Odet , ”Automated

3D region growing algorithm based on an assessment function,” Pattern

Recognition Letters. vol. 23 137-150, 2002.

[13] E. Davies, ”A high speed algorithm for circular object detection,” Pat-

tern Recognition Letters. vol. 6, 323-333, 1987.

[14] Y. Sato, S. Nakajima, H. Atsumi, T. Koller, G. Gerig, S. Yoshida, and

R. Kikinis, ”3D multi-scale line filter for segmentation and visualization

of curvilinear structures in medical images,” in J. Troccaz, E. Grimson,

and R. Mosges, eds., Proc. CVRMed- MRCAS97, LNCS, 213-222, 1997.

[15] A. Frangi, W. Niessen, K. Vincken, and M. Viergever, ”Multiscale ves-

sel enhancement filtering,” in Wells, W.M. III, Colchester, A., Delp, S.,

eds.: Proc. of the 1st International Conference on Medical Image Com-

puting and Computer-Assisted Intervention, MICCAI98. Volume 1496

of LNCS. 130-137, 1998.

[16] M. Kass, A. Witkin, and D. Terzoopoulos, ”Snakes: Active contour

models,” International Journal of Computer Vision. 1(4) 321-331, 1998.

[17] L. Lorigo, O. Faugeras, W. Grimson, R. Keriven, R. Kikinis, A. Nabavi,

and C. Westin, ”Curves: Curve evolution for vessel segmentation,” Med-

ical Image Analysis. vol. 5 195-206, 2001.

39

Page 48: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Bibliography 40

[18] D. Nain, A. Yezzi, and G. Turk, ”Vessel segmentation using a shape

driven flow,” in Proc. of the 7th International Conference on Medi-

cal Image Computing and Computer-Assisted Intervention, MICCAI04.

Volume 3216 of LNCS. 51-59, 2004.

[19] V. Casseles, F. Catte, T. Coll, and F. Dibos, ”A geometric model for

active contours in image processing,” Numerische Mathematik 66(1) 1-

32,1993.

[20] C. Xu and J. Prince, ”Snakes, shapes, and gradient vector flow,” IEEE

Transactions on image Processing. vol. 7, 359-369, 1998.

[21] T. Cootes, C. Taylor, D. Cooper, and J. Graham, ”Active shape models

- their training and application,” Computer Vision and Image Under-

standing 61(1) 38-59, 1995.

[22] K. Lekadir, R. Merrifield, and Y. Guang-Zhong, ”Outlier detection and

handling for robust 3-D active shape models search,” IEEE Transactions

on Medical Imaging 26(2) 212-222, 2007.

[23] T. Cootes, G. Edwards, and C. Taylor, ”Active appearance models,”

IEEE Transactions on Pattern Analysis and Machine Intelligence 23

681-685, 2001.

[24] Y. Boykov, and G. Funka-Lea, ”Graph cuts and efficient n-d image

segmentation,” International Journal of Computer Vision 70(2) 109-

131, 2006.

[25] L. Kang, W. Xiaodong, D. Chen, and M. Sonka, ”Optimal surface seg-

mentation in volumetric images - a graph-theoretic approach,” IEEE

Transactions on Pattern Analysis and Machine Intelligence 28(1) 119-

134, 2006.

[26] G. Slabaugh and G. Unal, ”Graph cuts segmentation using an elliptical

shape prior,” in Proc. of the 2005 IEEE International Conference on

Image Processing, ICIP05. vol. 2 1222-1225, 2005.

40

Page 49: Nearly Automatic Liver Vessels Segmentation of Computed ... · the blood vessels starting from a single user-de ned pixel seed inside the liver area. Then, the seed area is used for

Bibliography 41

[27] A. Sinop and L. Grady, ”Accurate banded graph cut segmentation of

thin structures using laplacian pyramids,” in Proc. of the 9th Interna-

tional Conference on Medical Image Computing and Computer-Assisted

Intervention, MICCAI06. Volume 4191 of LNCS. 896-903, 2006.

[28] X. Ning, A. Narendra, and B. Ravi, ”Object segmentation using graph

cuts based active contours,” Computer Vision and Image Understanding

107(3) 210-224, 2007.

[29] C. Rother, V. Kolmogorov, and A. Blake, ”Grabcut: interactive fore-

ground extraction using iterated graph cuts,” ACM Transactions on

Graphics 23(3) 309-314, 2004.

[30] L.R. Ford and D.R. Fulkerson, ”Maximal flow through a network,”

Canadian Journal of Mathematics vol. 8. 399-404, 1956.

[31] M. Freiman, O. Eliassaf, Y. Taieb, L. Joskowicz, and J. Sosna, ”A

bayesian approach for liver analysis: Algorithm and validation study,”

in Proc. of the 11th International Conference on Medical Image Com-

puting and Computer-Assisted Intervention, MICCAI08. Volume 5241

of LNCS. 8592, 2008.

[32] http://lts08.bigr.nl/

[33] L. Ibanez, W. Schroeder, L. Ng, and J. Cates, ”The ITK Software

Guide”. Kitware, Inc. ISBN 1-930934-15-7, 2005. http://www.itk.org/

ItkSoftwareGuide.pdf

[34] P. Yushkevich, J. Piven, C. Heather, R. Gimpel Smith, H. Sean, J.

Gee, and G. Guido, ”User-guided 3D active contour segmentation of

anatomical structures: Significantly improved efficiency and reliability,”

Neuroimage 31(3) 1116-1128, 2006. http://www.itksnap.org/.

[35] M. Freiman, M. Natanzon, N. Broide, L. Joskowicz, L. Weizman, J.

Sosna, E. Nammer, and O. Shilon, ”Patient-specific modeling of the

carotid arteries for surgical simulation.” In proc. of the 23rd Interna-

tional Conference on Computer Aided Radiology and Surgery, CARS’09,

2009.

41