Absolute Quantification in 1H MRSI of the Prostate at

Absolute Quantification in 1 H MRSI of the Prostate at 3 Tesla

Guillaume Lemaˆıtre

Department of Engineering and Physical Sciences Heriot-Watt University

Department of Computer Architecture and Technology Universitat de Girona

LE2I Laboratory Heriot-Watt Université de Bourgogne

A Thesis Submitted for the Degree of MSc Erasmus Mundus in Vision and RoBOTics (ViBOT) · 2011 ·

Abstract Prostate cancer is the second most frequently diagnosed cancer agmonst men in the world. Many techniques are existing and used in order to detect and stage prostate cancers. However, magnetic resonance imaging techniques propose a several non-invasive exams. Magnetic Resonance Spectroscopic Imaging (MRSI) allows to study the metabolite concentrations. 1 H-MRSI was performed on sixteen men with a suspected prostate cancer using a pelvic multi-channel phased-array (8 channels) coil. The aim of this study is to estimate absolute concentrations of choline and citrate using a method the most adapted to our acquisition data. Absolute quantification was done using water signal as reference. Different variations of choline and citrate concentrations were observed compare to the one from literature. Finally, further research directions are suggested.

Keywords: Prostate cancer, MRSI, absolute quantification, water reference, choline, citrate.

Science, like Love, is blind. That’s why she likes to proceed by trial and error . . . Jean O’Neil

Contents Acknowledgments

vii

1 Introduction

1

1.1

Anatomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.2

Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

1.2.1

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

1.2.2

Risk Factors

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

Medical Exams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4

1.3.1

Digital Rectal Examination . . . . . . . . . . . . . . . . . . . . . . . . . .

4

1.3.2

Prostate Specific Antigen (PSA) test . . . . . . . . . . . . . . . . . . . . .

4

1.3.3

Biopsy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4

1.3.4

Magnetic Resonance Imaging . . . . . . . . . . . . . . . . . . . . . . . . .

5

1.4

Project Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.5

Structure of the Document . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.3

2 Literature Review of Absolute Quantification 2.1

2.2

9

Spectroscopy & Prostate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9

2.1.1

Spectroscopy of the Prostate . . . . . . . . . . . . . . . . . . . . . . . . .

9

2.1.2

Metabolite Localization & Interpretation . . . . . . . . . . . . . . . . . .

9

Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 i

2.3

2.2.1

Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2.2

Peak Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2.3

Peak Fitting with Prior Knowledge . . . . . . . . . . . . . . . . . . . . . . 13

Quantification Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.1

Relative Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.3.2

Absolute Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Methodology 3.1

3.2

18

Materials and Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.1.1

MRSI Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.1.2

Study Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.2.1

Baseline Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.2.2

Water Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2.3

Rough Prostate Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.2.4

Choline - Citrate Quantification . . . . . . . . . . . . . . . . . . . . . . . 23

3.2.5

Absolute Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4 Results and Discussion 4.1

4.2

29

”Healthy” Metabolisme Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . 29 4.1.1

Position (Apex-Median-Base) Behaviour . . . . . . . . . . . . . . . . . . . 29

4.1.2

Peripheral Zone - Central Gland Behaviour . . . . . . . . . . . . . . . . . 31

Cancer versus Healthy Tissue Behaviour . . . . . . . . . . . . . . . . . . . . . . . 31 4.2.1

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.2.2

Citrate Concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4.2.3

Choline Concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5 Conclusion 5.1

37

Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 ii

5.2

Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

A Rough Prostate Segmentation Results

39

Bibliography

47

iii

List of Figures 1.1

Sagittal anatomy scheme of the male reproductive system [8]. . . . . . . . . . . .

1.2

Representation of the prostate. 1: Vas deferens, 2: Ampulla, 3: Seminal vesicle,

1

4: Excretory duct of seminal vesicle, 5: Prostate contour, 6: Ejaculatory duct, 7: Prostatic urticle, 8: Glandular tissue, 9: Urethral sphincter, 10: Urethra, 11: Seminal colliculus, 12: Urethral crest [33]. . . . . . . . . . . . . . . . . . . . . . . 1.3

2

Presentation of the different zones of the prostate. AFT: anterior fibromuscular tissue, CZ: central zone, ED: ejaculatory duct, NVB: neurovascular bundle, PUT: periurethral tissue, PZ: peripherical zone, U: urethra, TZ: transitional zone [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2

1.4

Cancer estimations in 2008 by the World Health Organization (WHO) [5]. . . . .

3

1.5

Biopsy of the prostate using transrectal ultrasound . . . . . . . . . . . . . . . . .

5

1.6

Different functional images available using magnetic resonance imaging . . . . . .

6

2.1

Variations and localization of the different disciminative metabolites. Choline (Cho) is localized at 3.21 ppm while citrate (Cit) is localized 2.64 ppm. From Kurhanewicz et al. [12, 13] the citrate should decrease while the choline should increase comparing respectively healthy and cancer prostate sprectrum. . . . . . 10

2.2

Representation of the different methods in order to analyze the data . . . . . . . 12

2.3

Representation of the three different type of well-known function: Voigt, Lorentzian and Gaussian functions. In MRSI, Voigt profile is commonly used to fit peaks. . 13 iv

3.1

Presentation of the method and criteria of Xi et al. method [34]. . . . . . . . . . 20

3.2

Example of the three different water peaks at different echo times (TE). We can observed a decay exponential from echo time (TE) of 30, 80 and 140 milliseconds. 22

3.3

Water integration method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4

Fitting of Voigt function after detecting the choline peak. . . . . . . . . . . . . . 24

3.5

Citrate integration method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.1

Variations of choline and citrate along Apex-Median-Base in the peripheral zone. 30

4.2

Variations of choline and citrate along Apex-Median-Base in the central gland. . 30

4.3

Variations of choline and citrate between Peripheral Zone and Central Gland. . . 31

4.4

Citrate concentration variations at different part of the prostate between healthy and cancer proved cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.5

Citrate concentration variations at different part of the prostate between healthy and cancer proved cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.6

Choline concentration variations at different part of the prostate between healthy and cancer proved cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

A.1 Result of segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 A.2 Result of segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 A.3 Result of segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 A.4 Result of segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 A.5 Result of segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 A.6 Result of segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

v

List of Tables 2.1

Localization and variation of the discriminative metabolites . . . . . . . . . . . . 10

2.2

Strengths and weaknesses of data analysis methods . . . . . . . . . . . . . . . . . 16

2.3

Strengths and weaknesses of quantification mehtods . . . . . . . . . . . . . . . . 17

4.1

Variations of choline and citrate concentrations expressed in mmol.kg −1 depending of the region prostate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

vi

Acknowledgments I wish to express my gratitude to Doctor Paul Walker who supervised this master thesis, to explained to me the many subtleties regarding magnetic resonance and allowed me to work in complete freedom. I would like to manifest my acknowledgments to Professor Fabrice Meriaudeau who advised me during my entire schooling and helped me with my different administrative problems. I would like to present my gratefulness to the teachers who allowed me to improve my knowledge in this entire Master studies. A particular acknowledgment to all persons of the ViBOT-Computer Vision family who will recognize themselves into this acknowledgment, but escpecially Mojdeh, Alex, Amir, Antoine, Carlos, Chengjia, Dio, Emilie, Fran¸cois, Frida, Isa, Jian, Kassem, Miro, Motasem, Oscar, Peter, Rocio, Romuald, Sophie, Tong and Toni. Last but not least, I would like to express more warm regards to my family and my friends without who I would not be where I am.

vii

Chapter 1

Introduction 1.1

Anatomy

The prostate is an exocrine gland of the male reproductive system and possesses an inverted pyramidal form. Its mensurations are usually about 3 centimeters in height by 2.5 centimeters in depth. Its weight is estimated between 15 and 25 grams for an adult. The size of the prostate increases at two moments during development: initially at puberty to reach its normal size then after 60 years of age leading to benign prostatic hyperplasia. The prostate is located below the bladder and in the front of the rectum and the urethra goes through the prostate as shown on figure 1.1. The urethral sphincter is located at the apex of the prostate around the prostatic urethra in order to drain the glands. The prostate is also composed of a muscle which allows the expulsion of the sperm during the ejaculation. The prostate has an inverted pyramidal form. The base is the upper part and closest to the

Figure 1.1: Sagittal anatomy scheme of the male reproductive system [8]. 1

Chapter 1: Introduction

2

Figure 1.2: Representation of the prostate. 1: Vas deferens, 2: Ampulla, 3: Seminal vesicle, 4: Excretory duct of seminal vesicle, 5: Prostate contour, 6: Ejaculatory duct, 7: Prostatic urticle, 8: Glandular tissue, 9: Urethral sphincter, 10: Urethra, 11: Seminal colliculus, 12: Urethral crest [33]. bladder while the apex is lower down and further from the bladder (figure 1.1 and 1.2). The seminal vesicles are located above the base of the prostate localized between the rectum and the bladder (figure 1.1). The prostate can be divided in different zones (figure 1.3) as proposed by McNeal [21] and widely accepted in the literature [4, 9, 24, 32]: central zone, transitional zone and peripheral zone. The peripheral zone which represents about 70% of the prostate is composed of glandular tissue. Roughly 70% of prostate cancers originate in this zone. The central zone which accounts for about 20-25% of the prostate is composed of stromal

(a) Transverse anatomy of the (b) Sagital prostate. prostate.

anatomy

of

the

Figure 1.3: Presentation of the different zones of the prostate. AFT: anterior fibromuscular tissue, CZ: central zone, ED: ejaculatory duct, NVB: neurovascular bundle, PUT: periurethral tissue, PZ: peripherical zone, U: urethra, TZ: transitional zone [3].

3

1.2 Statistics

tissue. The excretory ducts of the seminal vesicles and ampulla go through the base and join to form the ejaculatory duct. The transitional zone is composed of two symetric lobes localized on each sides of the urethra. The transitional zone represents 5% of the prostate. The size of this zone increases with age and with the development of a pathology known as benign prostatic hyperplasia. Approximately 30% of prostate cancers are found in these two zones. On MRI images, the central and transitional zones are usually difficult to distinguish. The peripheral zone accounts for about 70% of glandular tissue. 70% of prostate cancers arise in this zone.

1.2

Statistics

1.2.1

Overview

The World Health Organization (WHO) published in 2008 that prostate cancer was the second most frequently diagnosed cancer of men and the fifth most common cancer overall [5]. No less than 899,000 new cases where detected worldwide in 2008 [5]. As presented on figure 1.4, prostate cancer accounts for approximately 7.1% (figure 1.4(a)) of all cancers diagnosed in 2008 and 3.4% (figure 1.4(b)) of all cancers deaths in 2008 [5].

(a) Estimated number cancers cases for both sexes and (b) Estimated number cancers deaths for both sexes and all ages. all ages.

Figure 1.4: Cancer estimations in 2008 by the World Health Organization (WHO) [5].

1.2.2

Risk Factors

The risk factors can be categorized in three different classes: • Age: age is the most important risk factor for prostate cancer. The diagnosis of prostate cancers for men over 50 years old. Prostate cancer rate increases upto about 70 and declines thereafter [1].


4

• Genetic factors: it has been shown that the probability to have a cancer is higher when a member of the family has been already diagnosed [1]. • Race: in the United States, the Africo Americans have a higher probability of developing a prostate cancer than European American and Hispanic men [1].

1.3

Medical Exams

The presence of prostate cancer may be suggested in several ways: digital rectal examination, Prostate Specific Antigen (PSA) test, biopsy using transrectal ultrasound and magnetic resonance imaging.

1.3.1

Digital Rectal Examination

Both benign prostatic hyperplasia and cancer may lead to an increasing size of the prostate. A rectal examination may allow detection of harder nodules within the softer prostatic tissue. The advantages are that this method is very fast and does not need any special equipment.

1.3.2

Prostate Specific Antigen (PSA) test

The Prostate Specific Antigen (PSA) is a protein secreted by the prostate. A higher-thannormal level of PSA can indicate an abnormality of the prostate: a benign prostatic hyperplasia or a cancer. However, other factors can lead to an increasing level of PSA such as prostate infections, irritations, a recent ejaculation or a recent rectal examination, etc. The PSA can be found in the blood in two different forms: free PSA (about 10%) and linked to another protein (about 90%). A level of PSA higher than 10 ng.mL−1 is considered as pathologic [24]. If the PSA level is between 10 ng.mL−1 and 4 ng.mL−1 , the patient is considered as suspicious [24]. In that case, the ratio free PSA over total PSA is computed. If the ratio is higher than 15%, the case is considered as pathologic.

1.3.3

Biopsy

As described in section 1.1, the prostate is localized in front of the rectum. Hence, its position allows one to carry out a biopsy using transrectal ultrasound in order to localise more precisely an eventual cancer (figure 1.5). At least six different samples are taken from the right and left parts of the three different zones: apex, median and base. The samples are analyzed in order to determine the presence of a cancer.

5

1.3 Medical Exams

Figure 1.5: Biopsy of the prostate using transrectal ultrasound The biopsy is usually prescribed when the PSA level is higher-than-normal or abnormalities were detected during a rectal examination.

1.3.4

Magnetic Resonance Imaging

Magnetic resonance imaging is a relatively recent technique. This exam allows one to obtain a better spatial resolution and a more precise localization of the cancer compared with the previous methods. T2 Weighted Imaging T2 weighted imaging provides a spontaneously high contrast image and a good spatial resolution (figure 1.6(a)). Fat, muscles and prostate can easily be differantiated through the contrast between each tissue. The peripheral zone is mainly glandular, which implies a high-signal intensity enclosed by a thin border of low-signal intensity [2]. The central and transition zone are fibrous zones and give a lower signal than that of the peripheral zone [2]. Tumors are characterized by very low signal intensity [2]. Notice that cancer tissue might be distinguished easily from normal tissue only in the peripheral zone. Diffusion Imaging Diffusion imaging represents the degree of diffusion of the water molecules inside the tissue. On diffusion images, the prostate can be divided in two parts. The glandular nature of the prostate


6

(a) T2 weighted image. The (b) Diffusion image. The prostate (c) Perfusion image. The prostate prostate is highlighted in green. is highlighted in green. The cancer is highlighted in green. The cancer The cancer is highlighted in red. is highlighted in red. is highlighted in red.

(d) Spectroscopy image. The prostate voxels are highlighted in color. The cancer voxels are highlighted in orange.

(e) Spectra of an healthy voxel.

(f) Spectra of a cancer voxel.

Figure 1.6: Different functional images available using magnetic resonance imaging zone means that tissue. Water is able to move freely and gives a hyper intense signal. The central zone is more chaotic and inhibits the motion of the water. Thus, the signal of central zone will be lower and more heterogeneous than that of the peripheral zone. Tumors imply a greater density of membranes. These membranes also inhibit water motion. Hence, cancer tissue will appear to be darker on diffusion imaging as shown on figure 1.6(b). Perfusion Imaging Perfusion imaging provides a cinetic study. This technique consists to acquire a series of images in both space and time during the passage of contrast agent (figure 1.6(c)). The arrival and departure of the contrast agent will be studied in order to determine the type of tissue. The peripheral zone being mainly composed of glandular tissue the amount of intersitial space is limited involving restrected exchanges and a limited abundance of contrast agent. The signal over time neither rise quickly nor fall off quicly. The central zone has a more disorganised

7

1.4 Project Goals

structure which allows easily the arrival of contrast agent. The signal intensity over time will have a rapid increase followed by a plateau. Tumors, due of a high vacularity, will have high exchange in input and in output. Thus, the signal observed during time will have a rapid increase and rapid fall. Spectroscopic Imaging Spectroscopic imaging is a technique that allows the study of metabolite concentrations. The difference in metabolism between healthy and cancer tissues allows one to find a pattern in the concentrations observed. The resolution of the spectroscopy is inferior to the other techniques (figure 1.6(d)) but this technique is particularly sensitive. An example of data acquired in spectroscopy are presented on figures 1.6(f) and 1.6(e). Figure 1.6(f) represents a cancer tissue while 1.6(e) represents a healthy tissue. A decreased citrate concentration suggests a difference between the two different tissues [13].

1.4

Project Goals

The ultimate goal of this project was to implement a method allowing to compute as accurately as possible the absolute concentration of metabolites for prostate imaging using Magnetic Resonance Spectroscopy Imaging (MRSI). In order to identify the best possible approach to solve this problem, a survey was compiled of the state-of-the-art in absolute quantification using MRSI. From this survey, strengths and weaknesses of each method were pointed out. From this survey, a new method was proposed in order to fix the possible weaknesses viewed from the literature. Finally, the method was evaluated on a population of patients. A comparison was performed between the results obtained and those from the literature. Future directions for this work were pointed also out.

1.5

Structure of the Document

This thesis will be organized as follows: • Chapter 2: The state-of-the-art in absolute quantification is presented and discussed. • Chapter 3: The method implemented in order to compute absolute concentration is presented. • Chapter 4: Results of the method implemented are presented. Comparisons with the literature are given.


• Chapter 5: Conclusions are drawn and the future works are presented.

8

Chapter 2

Literature Review of Absolute Quantification In this review, particular attention will be given to three main points. First, the specificities of the MRSI used for prostate studies will be introduced. Then, methods used in order to analyze the data acquired and derive information about metabolite concentrations are presented. Finally, the different types of quantification in order to normalize are given.

2.1

Spectroscopy & Prostate

2.1.1

Spectroscopy of the Prostate

Some researchers have highlighted the behavior of some specific metabolites between healthy and cancer prostate tissue. Kurhanewicz et al. showed the discriminative effect of the citrate [13] and the choline [12]. Later on, Lynch and Nicholson [18] and Garcia-Segura et al. [6] showed that polyamines (mainly spermine), creatine and myo-inositol could be other dioristic biological markers. In our study, we will be interested only in choline and citrate metabolites which are easier to detect with our acquisition method.

2.1.2

Metabolite Localization & Interpretation

Metabolite Localization As previously introduced, citrate, choline, creatine and spermine are the main metabolites which are distinctive of prostate cancer. Each metabolite has a specific frequency at which 9

Chapter 2: Literature Review of Absolute Quantification

(a) Healthy prostate prostate

10

(b) Cancer prostate spectrum

Figure 2.1: Variations and localization of the different disciminative metabolites. Choline (Cho) is localized at 3.21 ppm while citrate (Cit) is localized 2.64 ppm. From Kurhanewicz et al. [12, 13] the citrate should decrease while the choline should increase comparing respectively healthy and cancer prostate sprectrum.

Metabolites Choline (3.21 ppm) Spermine (3.11 ppm) Creatine (3.02 ppm) Citrate (2.64 ppm)

Concentration in cancer tissue Increasing concentration [12] NA NA Decreasing concentration [13]

Concentration in healthy tissue Low concentration [12] NA NA High concentration [13]

Table 2.1: Localization and variation of the discriminative metabolites

it will occur on the spectrum obtained by the MRSI (figure 2.1). Table 2.1 summarizes the different chemical shift. Citrate will occur at 2.64 ppm, choline at 3.21 ppm, creatine at 3.02 ppm and spermine at 3.11 ppm. Figure 2.1 illustrates the possible form which can be obtained in healthy or cancer tissue for the prostate.

Metabolite Interpretation Kurhanewicz et al. studied the metabolism of the prostate and deduced that a decreased level of citrate [13] and an increased level of choline suggested prostate cancer [12]. These behaviours are shown on figure 2.1. Our study agrees with the first aspect regarding the citrate. However, we will show that we did not find a significative increasing level of choline.

11

2.2 Data Analysis

2.2

Data Analysis

In MRSI, the area under a peak at a given resonance frequency will be proportionnal to the concentration of that specific metabolite. Thus, the goal is to find the area under the peak as accurately as possible. Different methods exist in order to achieve this task. A summary of the strengths and weaknesses of each of these approaches is presented in table 2.2.

2.2.1

Integration

Due to the fact that information about metabolite concentrations is given by the peak integral of each metabolite, the most intuitive method is to compute the integral for a given frequency range as shown on figure 2.2(a). Integration required an algorithm or an intervention of the operator in order to define the boundaries. Meyer et al. [22] show that numeric integration can lead to an underestimation. This underestimation is due of the truncation of the wings of each peak [22]. This error is illustrated by comparison of the figures 2.2(a) and 2.2(b). Furthermore, if the boundaries are fixed from one analysis to another one, a problem of accuracy can arise due to a frequency shifting during the acquisition of the data.

2.2.2

Peak Fitting

In order to solve the problem that was pointed out with the integration method, another method is to fit the data to a known profile [10]. Common methods used from the literature fit a Gaussian or Lorentzian function (figure 2.3) to the data [10]. In order to fit a model as accurately as possible, pure Gaussian and Lorentzian models are not ideal. In fact, spectra are obtained from a Fourier transform of a decaying exponential which should give a pure Lorentzian as in equation 2.1. L(x) =

γ 1 π (γ)2 + (x − µ)2

(2.1)

where µ is the abscissa of the maxima and γ is the width at half-height. However, the signal is contaminated by a noise and is filtered. Indeed, it has been shown that the forms of peaks of MRSI spectra are following a Voigt function (figure 2.3) [19]. A Voigt function is a convolution between a Gaussian function and a Lorentzian function with a shared apex (equation 2.2). Z

∞

V (x) = (G ∗ L)(x) =

G(t)L(x − t) dt −∞

where G(x) is a Gaussian function anf L(x) is a Lorentzian function.

(2.2)


(a) Principle of integration. The area under the (b) Principle of peak fitting. The peak is fitted using curve is directly integrated which gives a value rep- a Voigt profile which will be integrated in order to resentative of the metabolite concentration estimate the area under the Voigt profile

(c) Principle of peak fitting with prior knowledge. In this case, peaks are fitted using LCModel. Problem of LCModel is located to the citrate fitting. The algorithm does not take in account the wings of the citrate peak.

Figure 2.2: Representation of the different methods in order to analyze the data

12

13

2.2 Data Analysis

Figure 2.3: Representation of the three different type of well-known function: Voigt, Lorentzian and Gaussian functions. In MRSI, Voigt profile is commonly used to fit peaks. Thus, the quantification is performed integrating the fitted Voigt function. This method has the merit of being fairly robust [10]. However, peaks are not found automatically and need a peak detection algorithm or an intervention of the operator.

2.2.3

Peak Fitting with Prior Knowledge

Several methods have been developed in order to compute an estimation of spectra using prior knowledge or a metabolite basis set. From LCModel, developed by S. Provencher [25], which is one of the most used in the medical field, to AMARES (Vanhamme et al. [31]) and QUEST (Ratiney et al. [27]) which both are implmented in the MRUI package [23], the choice of methods is wide. However, all these methods have a weak point. Indeed, they all have been developed in order to estimate spectra from brain. MRSI applied to prostate cancer is relatively a recent technique and demands are not enough important in order to create a specific analytic tool. We will present succintly the specificities of the three methods previously cited. LCModel LCModel is available as a stand-alone commercial software (http://s-provencher.com/). This method allows to estimate the relative concentrations of metabolite in demand and is performed in the frequency domain. The approach assumes that each spectrum can be seen as a linear combination of each metabolite studied [25]. The model estimated is as in equation 2.3.   NB NM Ns X X X Yˆ (vk ) = exp(−iΦ(vk ))  βj Bj (vk ) + Cl Γn Ml (vk−n ; ξl , δl ) j=1

l=1

n=−Ns

(2.3)


with Cl ≥ 0, ξl ≥ 0,

PNs

n=−Ns

14

Γn = 1

and • The ”baseline” or background is represented by

PNB

j=1

βj Bj (vk ) which is a sum of Nb cubic

splines. • Cl is the concentration of each metabolite. • Ml (vk−n ; ξl , δl ) is the model of the metabolite spectrum l. • Γn is a normalization factor. The algorithm is trying to minimize the error between the model (equation 2.3) and the real spectrum. As said before, LCModel was initially dedicated for brain studies. The spectra model of each metabolite (Ml ) was created from an in vitro solution. However, in order to try to adapt LCModel to the prostate case, the data set was created with simulated spectra which might not be representative of our acquired data. An example of fitting is shown on figure 2.2(c). A recurrent problem takes place to the citrate fitting. In fact, the algorithm is fitting to a single Gaussian. However, in our case, citrate in represented by a trident with small side lobes (figure 2.2(c) - the citrate trident is represented between the range 2.4-2.9 ppm). This error of fitting leads to an important underestimation of the citrate concentration and this feature is highly distinctive of prostate cancers [13]. MRUI Unlike LCModel, the MRUI package [23] is proposed free of charge to nonprofit organizations. In this package, two methods are available in order to perform quantification: AMARES [31] and QUEST [27]. • AMARES used a prior knowledge given by the user about relative frequencies, amplitude ratios, scalar coupling and phases of resonance [10]. The algorithm is performing in the time domain. The approach is to minimize the error between the model presented in equation 2.4 and the real signal in the time domain. This problem is non linear problem [31]. Parameters are found using an adaptive non linear least squares algorithm [31].

yn =

K X

ak exp(iΦk ) exp((−dk + i2πfk )tn ) + en

(2.4)

k=1

where ak is the amplitude, Φk is the phase, dk is the damping factor, fk is the frequency of the k th sinusoid and en is a Gaussian noise.

15

2.3 Quantification Strategies

• QUEST proposes a similar idea as AMARES. Using the algorithm of Levenberg-Marquardt (to solve a non linear problem), the parameters are found in order to minimize the error between equation 2.5. However, instead of having prior knowledge from the operator, a metabolite basis set is used in order to solve the problem fitting. The second main difference is that the algortihm in QUEST will try to estimate the baseline following the model presented in equation 2.6. xn = exp(iΦ0 )

M X

am x ˆm n exp [(∆αm + i∆ωm )tn + i∆Φm ]

(2.5)

m=1

where am is the amplitude, Φm is the phase, ∆αm , Deltaωm , ∆Φm are small changes of the damping factors, angular frequencies and phase shift. x=x ˆM et (p) + b(θ) + e

(2.6)

where w ˆM et is the metabolite part, b is the ”basline” or background and e is an additive Gaussian noise.

2.3

Quantification Strategies

Metabolite concentration estimation is usually divided into two different ways [10]. On the one hand, the concentrations can be expressed as ratios. This strategy is called relative quantification. On the other hand, the concentrations can be normalized by an absolute reference. This method is so called absolute quantification. A summary of relative merits anf drawbacks of each of these approaches is presented in table 2.3.

2.3.1

Relative Quantification

As presented in section 2.1.2, variations of choline and citrate has been considered as discriminative to detect prostate cancer [12, 13]. Thus, these metabolites are widely used in the literature [11, 26, 30] to establish a ratio as shown in equation 2.7. [ratio] =

[Cho] [Cit]

(2.7)

Due to the fact that choline and creatine peaks are sometimes merged, the ratio computed is as in equation 2.8. [ratio] =

[Cho] + [Cre] [Cit]

(2.8)


Data Analysis Methods

Stengths Complexity of the method

Integration

Flexibility of computation moving easily integration boundaries

16

Weaknesses Problem of accuracy due of frequency shifting Underestimation due of negligence of the peak wings

Fixed the problem of negligence of the peak wings Peak Fitting

Method semi-automatic

The computation is time comsuming

Accuracy The computation is time consuming Complexity of the method Peak Fitting with Prior Knowledge

Accuracy Method fully automatic

Time consuming to pare metabolite basis set

pre-

Reduce the degree of freedom of the fitting and are not adaptative to some data Table 2.2: Strengths and weaknesses of data analysis methods

17

2.3 Quantification Strategies

Quantification methods Relative quantification

Absolute quantification:

Stengths Easy implementation More sensitive changes

to

Weaknesses Impossible to precise variations of metabolite concentrations detect

Accurate knowledge about variations of metabolite concentrations Independences between metabolite concentrations

Less sensitive to abnormalities Complexity of the implementation More sensitive malities

to

Might be less sensitive to changes

Table 2.3: Strengths and weaknesses of quantification mehtods The use of ratios has the great advantage that the method is easy to implement. It is sensitive to metabolite concentration changes. However, ratios do not allow one to know precisely which variations of the metabolite concentrations occur. It has been shown that ratios were less sensitive to abnormalities [10].

2.3.2

abnor-

Absolute Quantification

The second way to express the concentrations is to normalize the values obtained by the data analysis by a reference. The main advantage of the absolute quantification is that variations of metabolites concentrations are known. It has been shown that abnormalities are better detected [10]. Different techniques exist in order to have a reference: external reference method, replace-and-match method, water signal reference method and principle of reciprocity. The water signal reference method will be used in our case and will be detailed in section 3.2.5.

Chapter 3

Methodology 3.1 3.1.1 1

Materials and Patients MRSI Protocol

H MR spectroscopic imaging have been collected at the University Hospital of Dijon). The

data were acquired on a clinical 3 Tesla whole body magnet (Siemens Magnetom Trio TIM, Erlangen, Germany) with a pelvic multi-channel phased-array (8 channels) coil. The study was performed using a water and lipid suppressed double-spin-echo point-resolved spectroscopic (PRESS) sequence optimized for quantification detection of choline and citrate metabolites. Water and lipid were suppressed using a dual-band spectral spatial pulse technique. In order to eliminate signals from adjacent tissues, especially periprostatic lipids and the rectal wall up to eight outer voxel saturation pulses were used. Data sets were acquired as 16 × 12 × 16 (interpolated to 16 × 16 × 16) phase-encoded spectral arrays, an echo time (TE) of 140 milliseconds, an repetition time (TR) of 720 milliseconds and 13 minutes acquisition time. A spectral bandwidth of 1250 Hertz was used with 512 data points. A combination of an elliptic weighted averaged k-space acquisition scheme 3D filtering of the signal in k-space were used, the latter in order to reduce intervoxel signal combination. Shimming was carried out using the Siemenbens 3D Mapshim routine on a voxel adapted to the volume of the entire prostate gland. Additional unsuppressed water acquitions (at TE 30, 80 and 140 milliseconds) of 1.5 minutes was also performed in order to allow quantification with respect to prostate water. Systematic verification of the global shim (ie. over the complete 3D PRESS-selected volume) revealed line widths at half-height of the water peak of the order of 20-30 Hertz, routinely. The line widths for individual voxels were of the order of 8-12 Hertz. The total examination time, including the time spent positioning the patient, was approximately 45 minutes. 18

19

3.2 Method

3.1.2

Study Population

”Healthy” Patients The data set was composed of eight patients who had negative biopsies. Mean subject age was 61.8 (range 53.3-71.1). PSA value ranged between 2.7 to 15.0 ng.mL−1 (mean, 8.0 ng.mL−1 ). Patients with Cancers The data set was composed of eight patients with biopsy proven cancer. Mean subject age was 70.9 (range 57.8-82.3). PSA value ranged between 0.4 to 74.0 ng.mL−1 (mean, 15.8 ng.mL−1 ). Gleason score estimated via the biopsies ranged from 6 to 7.

3.2

Method

As presented in section 2.1.2, choline and citrate are considered as distinctive metabolites in the literature [12, 13]. In the section 2.2.3, it was pointed out that LCModel and MRUI underestimate the citrate concentration during the quantification. That is why, our goal will be to implement an algorithm so as to obtain absolute concentration of these two metabolites. The following sections will present the scientific reasoning.

3.2.1

Baseline Detection

In order to obtain more accurate results for the integration of peaks, the baseline of the spectra has to be computed and substracted. Several methods exist to solve this problem: • The baseline might estimated by a polynomial minimizing the least-square error between the polynomial and a data subset [15]. • The baseline can be estimated using a wavelet decomposition. Theoretically, noise should be represented by detail coefficients while the baseline being a low frequency signal should be represented by approximation coefficients. Thus, keeping only the approximation coefficients, it should be possible to estimate the baseline in theory [20]. However, it is difficult to determine the level of decomposition for the wavelet transform [20]. • Some parametric methods allow to estimate the baseline and offer the advantage to customize the smoothness of the estimated baseline [34]. The method of Xi et al. [34] was the method chosen to solve the estimation of the baseline. A brief explanation of this method is given.

Chapter 3: Methodology

20

(a) Example of baseline detection (b) The first criterion is that the (c) The second criterion is that that using the method of Xi et al. [34]. baseline has to be smooth but not the baseline has to run through the flat. middle of the data in the portions where there is no signal.

Figure 3.1: Presentation of the method and criteria of Xi et al. method [34].

Xi et al. propose a statiscally based baseline detection derived from a parametric smoothing model [34]. They assumed a model where a measured spectra can be decomposed as the sum of a pure signal, the ”baseline” or background and an additive Gaussian noise (equation 3.1). This model is a common model and was also used in LCModel (section 2.2.3). yi = bi + µi eni + εi

(3.1)

where yi is the signal, bi is the baseline, µi is the true signal and ni and εi are Gaussian noises. Xi et al. built the cost function presented in equation 3.2 based on the assumption of the model chosen (equation 3.1) and that the baseline estimated should be smooth without being necessarily flat (figure 3.1(b)) and that should be laid on the middle of the data in segments where there is only noise (figure 3.1(c)) [34]. F (b) =

X i

bi − A

X

(bi+1 + bi−1 − 2bi )2 − B

X

i

(bi − γi )2 g(bi − γi )

(3.2)

i

where g(•) is the Heaviside function. The cost function F (b) is composed of three main terms: •

P

i bi

• −A

is the sum of all baseline points.

P

i (bi+1

+ bi−1 − 2bi )2 is the smoothness penalty. This term tends to be small for

linear segments while it will be large for small curvature radii. • −B

P

i (bi

− γi )2 g(bi − γi ) is the negative penalty. This term tends to be nonzero when

the baseline is above the data point.

21

3.2 Method

In order to compute the optimal baseline, the first term has to be maximized subject to both second and third terms (equation 3.3). b0 = arg max F (b)

(3.3)

At this stage, the two parameter A and B have to be chosen. Xi et al. have shown that the theoritical values are as in equation 3.4 and 3.5. 5 × 10−9 n4 σ 1.25 B= σ

A=

(3.4) (3.5)

where σ is an estimation of the standard deviation of the noise and n is the total number of data point. Hence, an estimation of the standard deviation of the noise of the spectrum has to be performed. In order to estimate this parameter, Xi et al. proposed to divide the spectrum into smaller sections and compute the variance and mean intensity for each region. Then, they proposed to use the LOWESS regression to fit a model to the variance versus mean intensity signal. Finally, the estimation of the standard deviation of the noise is the square root of the variance at zero mean intensity in the LOWESS regression (equation 3.6). σ=

p

σε2

(3.6)

where σε2 is the variance at zero mean intensity inferred using the LOWESS regression. Once that all parameters are found,

∂F (b) ∂b

= 0 has to be solved in order to maximize the

function F (b) [34].

3.2.2

Water Quantification

In order to calculate the absolute concentration using the tissue water as reference, a method allowing to integrate the water peak have to be implemented. In this part, we will explain only the method used to integrate the peak. The algorithm is described in algorithm 1. Details about normalization using the water reference is presented in section 3.2.5. The water has to be quantified on three sets of data where the echo time (TE) is the only variable. Indeed, the acquisitions were carried out at TE 30, 80 and 140 milliseconds. The longer is TE, the smaller will be the amplitude of the water peak as shown on figure 3.2. The position of the water peak is ca. 4.65 ppm. It might occur that the water peak could be shifted due to the non homogeneity of the magnetic field. So as to integrate the water peak, the steps of the algorithm 1 have to be performed.


22

Figure 3.2: Example of the three different water peaks at different echo times (TE). We can observed a decay exponential from echo time (TE) of 30, 80 and 140 milliseconds. Algorithm 1 Water integral peak determination for all spectra do Crop the signal between range 3.96-5.94 ppm. Compute a rough approximation of the signal using wavelet decomposition (Haar with 3 decomposition levels). Detect valleys of the water peak using the algorithm of Lavielle [14]. Check if valleys are local minimum. Detect the baseline using the algorithm of Xi and al. [34]. Substract the baseline to the original signal Compute the numeric integral of the previous signal between the detected valleys. end for

(a) Original signal is in blue. Red signal is the rough signal after wavelet analysis. Black crosses are the valleys detected using Lavielle’s algorithm [14] follow by a research of local minimum.

(b) Original signal is in blue. Green signal is the baseline detected using Xi et al. algorithm [34]. Red area will be integrated using Simpson’s rule. Black crosses are the valleys detected after research of local minimum.

Figure 3.3: Water integration method

23

3.2 Method

Approximation signal using Wavelet Decomposition and Synthesis Firstly, the signal is cropped between 3.96 and 5.94 ppm since the water peak cannot occur outside of this range. Then, before detecting the valleys of the water peak, the signal is processed so as to obtain a rough signal where limits will be easier to distinguished. This rough signal is obtained by decomposing the original signal in the wavelet domain with the Haar family at the third level. Then, only approximation coefficients are used to reconstruct the signal. Comparison of these two signals are presented on figure 3.3(a). Limits Detection Lavielle proposed an algorithm to detect multiple changes inside data sequence [14]. This algorithm is used to find the limits in the previous signal computed which are the valleys of the water peak. However, it might occur that the detected limits are not real local minima. Thus, each limit is moved to the local minima analyzing previous and next sample iteratively. Figure 3.3(a) and 3.3(b) illustrate the valley detection. Numeric Integral Computation The water peak will be integrated between the two limits previously detected. However, the baseline will be detected as presented in section 3.2.1 [34] and substracted from the original signal. The numeric integral will be computed using the Simpson’s rule as shown on equation 3.7. Z

a

f (x) dx = b

b−a a+b f (a) + 4f + f (b) 6 2

(3.7)

Red area on figure 3.3(b) is the area which will be integrated.

3.2.3

Rough Prostate Segmentation

To reduce the time of computation and to exclude non-prostate voxels, a rough prostate segmentation was performed. We observed that the prostate was quite nicely delimited when observing the water intensity signal. The prostate position was estimated using a simple Kmeans algorithm. Results for one patient are presented in annex A.

3.2.4

Choline - Citrate Quantification

In this section, we will present our method allowing to extract the relative concentrations of choline and citrate.


24

Figure 3.4: Fitting of Voigt function after detecting the choline peak.

Choline Quantification

Algorithm 2 Choline integral peak determination for all spectra do Crop the signal between range 3.17-3.29 ppm. Detect maxima of the peak. Detect valleys on each side of the maxima. These valleys are define to be the first nearest points of the baseline. Crop the real choline peak. Compute the baseline. Substract the baseline to the signal cropped. Fit a Gaussian G(x) and Lorentzian L(x) functions. Compute the convolution (G ∗ L)(x) in order to obtain a Voigt profile V (x). Compute the numeric integral of V (x) using Simpson’s rule. end for As presented in the table 2.1 in section 2.1.2, theoretically the peak of choline is localized at 3.21 ppm. Using this preknowledge, the signal was reduced to the range 3.17-3.29 ppm. The maxima of this range will be the apex of choline peak. The baseline is computed as described in section 3.2.1 [34] and substracted from the original signal. We will search iteratively each local minima on each side of the maximum. Then, only using the data between the limits detected, a Gaussian (equation 3.8) and a Lorentzian function (equation 3.9) will be fitted. In order to find the different parameter for the Gaussian and Lorentzian function (equation 3.8 and 3.9), the algorithm of Levenberg-Marquardt was used. G(x) = a exp(−

(x − b)2 ) 2c2

(3.8)

25

3.2 Method

L(x) =

1 b π (x − a)2 + b2

(3.9)

As presented in section 2.2.2, peaks in NMR are considered to follow a Voigt function (equation 3.10). Thus, the convolution between the estimated Gaussian and Lorentzian function is performed. Z

∞

V (x) = (G ∗ L)(x) =

G(t)L(x − t) dt

(3.10)

−∞

Having a discrete signal, the integration of the function V (x) was computed using the Simpson’s rule presented in equation 3.7. Figure 3.4 presents the fitting of a Voigt function to the choline peak. Citrate Quantification Algorithm 3 Citrate integral peak determination for all spectra do Smooth signal using cubic spline. Find minimum between 2.75-2.95 ppm and 2.40-2.50 ppm. Compute the baseline. From minimum find points the nearest to the baseline. Crop the real citrate peak. Substract the baseline to the cropped signal. Compute the numeric integral of the absolute previous signal using Simpson’s rule. end for The prior knowledge about the citrate peak is as follows: • The citrate peak is a trident with small side lobes. • The maximum of the citrate peak occurs at 2.64 ppm. • The two minima of the two side lobes occur at 2.47 ppm and 2.81 ppm, respectively. Using these hypothese, the original signal is cropped between 2.40 and 2.95 ppm. The signal is smoothed using a cubic spline function [28] (figure 3.5(a)). Using this smoothed function, the detection of minima is performed in the intervals 2.40-2.50 ppm and 2.75-2.95 ppm. The baseline of the absolute signal is computed (figure 3.5(a)). From the minima of the lower range (2.40-2.50 ppm), we will find the sample which will be the nearest to the baseline with a smaller abscissa. In the same way, from the minima of the higher range (2.75-2.95 ppm), we will find the sample wich will be the nearest to the baseline with a larger abscissa. These


26

(a) The blue signal is the original signal. The red signal is the signal after a cubic spline smoothing. The green signal is the baseline of the absolute value of the original signal while the two black crosses are the limits detected.

(b) The blue signal is the signal obtained taking the absolute value of the substraction between the original signal and the baseline. The red area will be integrated using Simpson’s rule between the two black crosses.

Figure 3.5: Citrate integration method two previous samples will be considered as limits of the citrate peak. The integration of the signal was computed using the Simpson’s rule presented in equation 3.7. The integration will be computed on the absolute substraction between the original signal and the baseline previously computed between the two limits computed as shown on figure 3.5(b).

3.2.5

Absolute Quantification

In this section, normilzation by a tissue water reference will be presented. This normalization allows to obtain absolute quantification. Absolute Quantification Absolute quantification using water reference is based on the fact that the fully relaxed signal from water or metabolites is proportional to the number of moles of the molecules in the voxel [7]. Thus, we can deduce equation 3.11. S0met S0H2 O

=

[met]

=

where: • S0 is the fully relaxed signal.

molmet × nHmet molH2 O × 2 2 × [H2 O] × S0met nHmet × S0H2 O

(3.11)

27

3.2 Method

• [met] is the absolute concentration for each metabolite. • H2 O is the tissue water concentration approximated at 39.4 mol.kg −1 [17]. • nHmet is the number of protons associated with each metabolite, four protons for citrate and nine protons for choline. Parameters which have to be computed are presented in the following sections. Determinatation of S0H2 O The fully relaxed corrected signal of the water S0H2 O can be obtained as expressed in equation 3.12. S0H2 O =

∗ S0H exp(− TTE ) 2O 2

1 − exp(− TTR ) 1

(3.12)

where: • The numerator expresses T2 correction (see next paragraph). • T R is the repetition time set up for the examination which is 720 milliseconds. • From previous studies, we observed that water T1 was approximately 1600 milliseconds. To obtain the numerator value of equation 3.12 corresponding to T2 correction of the water signal, we computed the integral of the water peaks as explained in section 3.2.2 for the three different TE 30, 80 and 140 milliseconds and fitted these data to an exponential decay of the form shown on equation 3.13. y = a exp(−bt)

(3.13)

The parameters a and b are obtained solving a non linear problem using the LevenbergMarquardt’s algorithm. In this case the parameter a corresponds to the numerator of equation 3.12. Determinatation of S0met As in the previous section, the fully relaxed corrected signal of the metabolite considered S0met can be obtained as expressed in equation 3.14. S0met =

∗ ) S0met exp(− TTE 2

1 − exp(− TTR ) 1

(3.14)


28

where: ∗ • S0met is obtained using methods of integration explained in section 3.2.4 regarding relative

concentrations of choline and citrate. • T E during the acquisition is set up at 140 milliseconds. • T2 is metabolite dependent. From the literature [29], citrate has a T2 of 180 milliseconds and choline has a T2 of 220 milliseconds. • From previous studies, we observed that T1 for choline was approximately 1500 milliseconds and for citrate was 600 milliseconds.

Chapter 4

Results and Discussion All results are presented using ”box and whisker” plots. For each box, the central red mark is the median while the blue edges are the twenty-fifth and seventy-fifth percentiles. The whiskers extend to the minima and maxima of the data.

4.1

”Healthy” Metabolisme Behaviour

4.1.1

Position (Apex-Median-Base) Behaviour

Peripheral Zone Both citrate and choline concentrations seem to increase from the apex to the base of the prostate (25 % between apex-median and 15 % between median-base) (figure 4.1). Furthermore, both citrate and choline concentrations tend to be higher (around 25 %) in healthy tissue of patients with negative biopsies than healthy tissue of patients with biopsy proven cancer.

Central Gland Both choline and citrate concentrations are almost constant in the central gland in the three different zones apex, median and base (figure 4.2). However, a very low increase (around 10 % respectively between apex-median and median-base) can be observed in healthy tissue of patients with negative biopsies through the different parts of the prostate. As in the peripheral zone, concentrations of patients with negative biopsies are higher than healthy tissue with biopsy proven cancer (around 20-30 %). 29

Chapter 4: Results and Discussion

30

(a) Variations of choline along Apex-Median-Base (b) Variations of citrate along Apex-Median-Base for ”Healthy” tissue. for ”Healthy” tissue.

Figure 4.1: Variations of choline and citrate along Apex-Median-Base in the peripheral zone.

(a) Variations of choline along Apex-Median-Base (b) Variations of citrate along Apex-Median-Base for ”Healthy” tissue. for ”Healthy” tissue.

Figure 4.2: Variations of choline and citrate along Apex-Median-Base in the central gland.

31

4.2 Cancer versus Healthy Tissue Behaviour

(a) Variations of citrate between peripheral and cen- (b) Variations of choline between peripheral and tral zone. central zone.

Figure 4.3: Variations of choline and citrate between Peripheral Zone and Central Gland. In order to find out the explenations of this phenomenon, an acquistion using a phantom with homogenous concentration of citrate and choline has to be performed and studied.

4.1.2

Peripheral Zone - Central Gland Behaviour

The mean value (± SD) of citrate concentration for peripheral zone was 45.34±14.83 mmol.kg −1 which was significantly higher than that for central gland, at 24.00±8.76 mmol.kg −1 . The same observation can be done regarding the choline concentration where the mean value (± SD) for peripheral zone was 2.25 ± 0.64 mmol.kg −1 which was also significantly higher than that for central gland, at 1.87 ± 0.42 mmol.kg −1 . These pattern are illustrated on figure 4.3.

4.2 4.2.1

Cancer versus Healthy Tissue Behaviour Overview

As presented in section 2.1.2, Kurhanewicz et al. [12, 13], Lowry et al. [17] and Liney et al. [16] present the discriminative role of citrate and choline. That is why, our study was focus in the different variations of these two metabolites. A comparison of our results and the results previouly given in the literature [12, 13, 16, 17] will be given. Results are presented in table 4.1.

4.2.2

Citrate Concentration

Results obtained regarding citrate concentration in healthy and cancer tissues are in coordination with the work presented in the literature [16, 17]. Figure 4.4 are illustrated the different


Tissue Type ”Healthy” Patients

32

Prostate Region Base

Apex

Median

Overall

Choline: 1.70 ± 0.40

Choline: 2.28 ± 0.56

Choline: 2.60 ± 0.60

Choline: 2.25 ± 0.64

Citrate: 33.41 ± 10.10

Citrate: 45.67 ± 14.05

Citrate: 54.28 ± 12.94

Citrate: 45.34 ± 14.83

Choline: 1.87 ± 0.44

Choline: 2.00 ± 0.45

Choline: 1.87 ± 0.42

Citrate: 21.34 ± 6.42

Citrate: 23.87 ± 9.38

Citrate: 26.42 ± 9.52

Citrate: 24.00 ± 8.76

Choline: 1.66 ± 0.32

Choline: 1.80 ± 0.48

Choline: 2.02 ± 0.82

Choline: 1.82 ± 0.57

Citrate: 23.67 ± 10.73

Citrate: 35.01 ± 11.52

Citrate: 39.20 ± 20.82

Citrate: 32.97 ± 15.45

Choline: 1.30 ± 0.29

Choline: 1.45 ± 0.19

Choline: 1.50 ± 0.41

Choline: 1.42 ± 0.30

Citrate: 19.70 ± 7.44

Citrate: 16.77 ± 3.82

Citrate: 16.16 ± 4.52

Citrate: 17.43 ± 5.31

Choline: 1.57 ± 0.64

Choline: 1.39 ± 0.37

Choline: 1.54 ± 0.32

Choline: 1.47 ± 0.40

Citrate: 12.03 ± 8.39

Citrate: 13.24 ± 3.95

Citrate: 16.71 ± 4.90

Citrate: 14.24 ± 5.28

Peripheral Zone

Choline:1.71 0.34

±

Central Gland

Cancer Patients

Peripheral Zone

Central Gland

Cancer Tissue

Table 4.1: Variations of choline and citrate concentrations expressed in mmol.kg −1 depending of the region prostate.

33


(a) Citrate variations through all the prostate. (b) Citrate variations in the apex. Comparison beComparison between negative biopsies and biopsy tween negative biopsies and biopsy proven cancer proven cancer patients. patients.

(c) Citrate variations in the median. Comparison (d) Citrate variations in the base. Comparison bebetween negative biopsies and biopsy proven cancer tween negative biopsies and biopsy proven cancer patients. patients.

Figure 4.4: Citrate concentration variations at different part of the prostate between healthy and cancer proved cases.


34

level of citrate concentration in the different prostate zone. The results are available inside the table 4.1. The main observation which can be done is that the mean value (± SD) of citrate concentration in cancer tissue was 14.24 ± 5.28 mmol.kg −1 which was significantly lower than in peripheral zone, at 39.15 ± 15.14 mmol.kg −1 (figure 4.4(a)). However, it is more difficult to distinguish the difference of concentrations between cancer tissue and tissue of the central gland. The mean value (± SD) in central gland was 20.71 ± 7.04 mmol.kg −1 (mean of healthy tissue of healthy and cancer proven patients) while the concentration inside cancer tissue was 14.24 ± 5.28 mmol.kg −1 . An example for an healthy slice 4.5(b) and a cancer slice 4.5(d) is presented on figure 4.5. On these figures, the level of concentration is graduated where the blue color represent the minima concentration while red is the maxima. On figure 4.5(b), the peripheral zone can be characterized by a high level of citrate concentration compare to the central zone. On figure 4.5(d), the same distinction can be done. However, we can observed that the cancer zone present a relatively low level of citrate concentration and have a high contrast with the peripheral zone citrate concentration. Moreover, it is more difficult to distinguish the difference between cancer tissue and central gland tissue only observing citrate concentration. Therefore, tumor tissue can be characteristic of a dicreasing concentration of citrate when occuring in the peripheral zone.

4.2.3

Choline Concentration

Kurhanewicz et al. has been shown an increasing level of choline concentration between normal and cancer tissue [12]. Obtaining different results, we wanted to rise some points about the method used by Kurhanewicz et al. Indeed, in order to proof the variations of choline between healthy and cancer tissues, they compute ratios of choline concentration of different tissues (cancer, normal tissue, BPH, etc.) over the mean value of choline concentration of the normal peripheral zone of the same patient. Our results obtained in section 4.1 tend to show a variability of choline and citrate concentrations according to the type of tissue or to the prostate position of the voxel. These variations could have an influence on the computation of the ratios. In our results, we observed that the choline concentration is not varying in order to detect a cancer 4.6(a). Furthermore, we observed a small decreseasing level of choline concentration. The mean value (± SD) of choline concentration in peripheral zone was 2.03 ± 0.60 mmol.kg −1 while the choline concentration in cancer tissue was 1.47 ± 0.40 mmol.kg −1 . This change of choline concentration cannot be considered as distinctive.

35


(a) Example of an anatomic view of an healthy pa- (b) Maps of the citrate concentrations. We can obtient. Blue voxels are peripheral zone, red voxels are served that the citrate concentration is higher than merged zone while green zone is the central gland. in the central gland.

(c) Example of an anatomic view of a cancer prooved patient. Blue voxels are peripheral zone, red voxels are merged zone while green zone is the central gland.

(d) Maps of the citrate concentrations. We can observed that the citrate concentration is higher than in the central gland. Furthermore, the cancer zone localized in the cancer zone is characterized by a low level of citrate concentration.

Figure 4.5: Citrate concentration variations at different part of the prostate between healthy and cancer proved cases.


36

(a) Choline variations through all the prostate. (b) Choline variations in the apex. Comparison beComparison between negative biopsies and biopsy tween negative biopsies and biopsy proven cancer proven cancer patients. patients.

(c) Choline variations in the median. Comparison (d) Choline variations in the base. Comparison bebetween negative biopsies and biopsy proven cancer tween negative biopsies and biopsy proven cancer patients. patients.

Figure 4.6: Choline concentration variations at different part of the prostate between healthy and cancer proved cases.

Chapter 5

Conclusion 5.1

Considerations

A novel method to analyse the spectroscopy data of prostate has been done. A customize method using Voigt function fitting and integration allows to compute accurately absolute concentrations of choline and citrate metabolites since these two metabolites have been shown to be discriminative of prostate cancers [12, 13, 16, 17]. As shown in [12, 16, 17], we observed a significant decreasing level of citrate concentration from normal tissue to cancer tissue. However, this difference was only distinctive between the peripheral zone and the cancer zone. It was more difficult to observe a difference between the central gland and the cancer tissue. The contribution of this work lie in the study of the variations of choline concentration. Indeed, Kurhanewicz et al. has been shown an increasing level of choline concentration in cancer tissue compare to normal tissue [12]. Our results shown that no significant increase was observed between cancer tissue and normal tissue. Furthermore, a small decreasing level of choline concentration was observed.

5.2

Future Works

Future works which have to be carried out for future publication or extension of this actual works are: • An acquisition of a phantom of choline and citrate have to be done in order to explain the variations through the different prostate zones (apex-median-base). • This acquisition will allow to proof the real accuracy of the algorithm. 37

Chapter 5: Conclusion

38

• These absolute concentrations could be used as features in order to create a framework using different MRI modalities to detect and classify prostate cancers via usual method of classification (SVM, neural networks, etc.).

Appendix A

Rough Prostate Segmentation Results A K-means algorithm was used in order to segment the prostae using the water signal at time TE of 30 milliseconds. On the anatomic images, the prostate is highlighted in red. The automatic segmentation is delimited by the blue voxels.

(a) Water intensity - slice 1

(b) Prostate segmentation - slice 1

Figure A.1: Result of segmentation

39

Chapter A: Rough Prostate Segmentation Results

40



(c) Water intensity - slice 3

(d) Prostate segmentation - slice 3

(e) Water intensity - slice 4

(f) Prostate segmentation - slice 4


41









42








43









44








Bibliography [1] AmericanCancerSociety. Cancer facts and figures 2010, 2010. [2] P. R. Carroll, F. V. Coakley, and J. Kurhanewicz. Magnetic resonance imaging and spectroscopy of prostate cancer. Rev Urol, 8 Suppl 1:S4–S10, 2006. [3] Y. J. Choi, J. K. Kim, N. Kim, K. W. Kim, E. K. Choi, and K. S. Cho. Functional MR imaging of prostate cancer. Radiographics, 27:63–75, 2007. [4] F. V. Coakley and H. Hricak. Radiologic anatomy of the prostate gland: a clinical approach. Radiol. Clin. North Am., 38:15–30, Jan 2000. [5] J. Ferlay, H.R. Shin, F. Bray, D. Forman, C. Mathers, and D.M. Parkin. Globocan 2008, cancer incidence and mortality worldwide: Iarc cancerbase no. 10 [internet]., 2010. [6] J. M. Garcia-Segura, M. Sanchez-Chapado, C. Ibarburen, J. Viano, J. C. Angulo, J. Gonzalez, and J. M. Rodriguez-Vallejo. In vivo proton magnetic resonance spectroscopy of diseased prostate: spectroscopic features of malignant versus benign pathology. Magn Reson Imaging, 17:755–765, Jun 1999. [7] C. Gasparovic, T. Song, D. Devier, H. J. Bockholt, A. Caprihan, P. G. Mullins, S. Posse, R. E. Jung, and L. A. Morrison. Use of tissue water as a concentration reference for proton spectroscopic imaging. Magn Reson Med, 55:1219–1226, Jun 2006. [8] Geckomedia. Natom anatomy, 2011. [9] H. Hricak, G. C. Dooms, J. E. McNeal, A. S. Mark, M. Marotti, A. Avallone, M. Pelzer, E. C. Proctor, and E. A. Tanagho. MR imaging of the prostate gland: normal anatomy. AJR Am J Roentgenol, 148:51–58, Jan 1987. [10] J. F. Jansen, W. H. Backes, K. Nicolay, and M. E. Kooi. 1H MR spectroscopy of the brain: absolute quantification of metabolites. Radiology, 240:318–332, Aug 2006. 45

BIBLIOGRAPHY

46

[11] B. M. Kelm, B. H. Menze, C. M. Zechmann, K. T. Baudendistel, and F. A. Hamprecht. Automated estimation of tumor probability in prostate magnetic resonance spectroscopic imaging: pattern recognition vs quantification. Magn Reson Med, 57:150–159, Jan 2007. [12] J. Kurhanewicz, D. B. Vigneron, H. Hricak, P. Narayan, P. Carroll, and S. J. Nelson. Three-dimensional H-1 MR spectroscopic imaging of the in situ human prostate with high (0.24-0.7-cm3) spatial resolution. Radiology, 198:795–805, Mar 1996. [13] J. Kurhanewicz, D. B. Vigneron, S. J. Nelson, H. Hricak, J. M. MacDonald, B. Konety, and P. Narayan. Citrate as an in vivo marker to discriminate prostate cancer from benign prostatic hyperplasia and normal prostate peripheral zone: detection via localized proton spectroscopy. Urology, 45:459–466, Mar 1995. [14] M. Lavielle. Detection of multiple changes in a sequence of dependent variables. Stochastic Processes and their Applications, 83(1):79–102, September 1999. [15] C. A. Lieber and A. Mahadevan-Jansen. Automated method for subtraction of fluorescence from biological Raman spectra. Appl Spectrosc, 57:1363–1367, Nov 2003. [16] G.P. Liney, L. W. Turnbull, M. Lowry, L. S. Turnbull, A.J. Knowles, and A. Horsman. In vivo quantification of citrate concentration and water t2 relaxation time of the pathologic prostate gland using 1H MRS and MRI. Magn Reson Imaging, 15:1177–1186, 1997. [17] M. Lowry, G. P. Liney, L. W. Turnbull, D. J. Manton, S. J. Blackband, and A. Horsman. Quantification of citrate concentration in the prostate by proton magnetic resonance spectroscopy: zonal and age-related differences. Magn Reson Med, 36:352–358, Sep 1996. [18] M. J. Lynch and J. K. Nicholson. Proton MRS of human prostatic fluid: correlations between citrate, spermine, and myo-inositol levels and changes with disease. Prostate, 30:248–255, Mar 1997. [19] I. Marshall, J. Higinbotham, S. Bruce, and A. Freise. Use of Voigt lineshape for quantification of in vivo 1H spectra. Magn Reson Med, 37:651–657, May 1997. [20] Vincent Mazet. Développement de méthodes de traitement de signaux spectroscopiques : estimation de la ligne de base et du spectre de raies. PhD thesis, Université Henri Poincaré - Nancy 1, 2005. [21] J. E. McNeal. The zonal anatomy of the prostate. Prostate, 2:35–49, 1981. [22] R. A. Meyer, M. J. Fisher, S. J. Nelson, and T. R. Brown. Evaluation of manual methods for integration of in vivo phosphorus NMR spectra. NMR Biomed, 1:131–135, Jun 1988.

47

BIBLIOGRAPHY

[23] A. Naressi, C. Couturier, I. Castang, R. de Beer, and D. Graveron-Demilly. Java-based graphical user interface for MRUI, a software package for quantitation of in vivo/medical magnetic resonance spectroscopy signals. Comput. Biol. Med., 31:269–286, Jul 2001. [24] S. Parfait, J. Miteran, and P.M. Walker. Classification de spectres et recherche de biomarqueurs en spectroscopie par résonance magnétique nucléaire du proton dans les tumeurs prostatiques. PhD thesis, Université de Bourgogne - LE2I, 2010. [25] S. W. Provencher. Estimation of metabolite concentrations from localized in vivo proton NMR spectra. Magn Reson Med, 30:672–679, Dec 1993. [26] A. Rajesh, F. V. Coakley, and J. Kurhanewicz. 3D MR spectroscopic imaging in the evaluation of prostate cancer. Clin Radiol, 62:921–929, Oct 2007. [27] H. Ratiney, M. Sdika, Y. Coenradie, S. Cavassila, D. van Ormondt, and D. GraveronDemilly. Time-domain semi-parametric estimation based on a metabolite basis set. NMR Biomed, 18:1–13, Feb 2005. [28] Christian H. Reinsch. Smoothing by spline functions. Numerische Mathematik, 10:177–183, 1967. [29] T. W. Scheenen, G. Gambarota, E. Weiland, D. W. Klomp, J. J. Futterer, J. O. Barentsz, and A. Heerschap. Optimal timing for in vivo 1H-MR spectroscopic imaging of the human prostate at 3T. Magn Reson Med, 53:1268–1274, Jun 2005. [30] M. Seitz, A. Shukla-Dave, A. Bjartell, K. Touijer, A. Sciarra, P. J. Bastian, C. Stief, H. Hricak, and A. Graser. Functional magnetic resonance imaging in prostate cancer. Eur. Urol., 55:801–814, Apr 2009. [31] Leentje Vanhamme, Aad van den Boogaart, and Sabine Van Huffel. Improved method for accurate and efficient quantification of mrs data with use of prior knowledge, 1997. [32] A. Villers, A. Steg, and L. Boccon-Gibod. Anatomy of the prostate: review of the different models. Eur. Urol., 20:261–268, 1991. [33] Wikipdia. Prostate — wikipdia, l’encyclopdie libre, 2011. [En ligne; Page disponible le 17-mai-2011]. [34] Y. Xi and D. M. Rocke. Baseline correction for NMR spectroscopic metabolomics data analysis. BMC Bioinformatics, 9:324, 2008.