Molecular magnetic resonance imaging of gene therapy-induced apoptosis and gene transfer: A role for 1H spectroscopic imaging and iron oxide labelled viral particles (Spektroskooppisen magneettikuvantamisen ja rautaoksidi-leimattujen virusten rooli geenit

(1)

A.I. VIRTANEN INSTITUTE FOR MOLECULAR SCIENCES 50

TIMO LIIMATAINEN

Molecular magnetic resonance imaging of gene therapy-induced apoptosis and gene transfer

A role for ¹H spectroscopic imaging and iron oxide labelled viral particles

Doctoral dissertation To be presented by permission of the Natural and Environmental Sciences of the University of Kuopio for public examination in Medistudia Auditorium ML3, University of Kuopio, on

January 12^th 2007, at 12 noon Department of Biotechnology and Molecular Medicine

A.I. Virtanen Institute for Molecular Sciences University of Kuopio

(2)

Distributor: Kuopio University Library

P.O. Box 1627

FI-70211 KUOPIO FINLAND

Tel. +358 17 163 430

Fax +358 17 163 410

http://www.uku.fi/kirjasto/julkaisutoiminta/julkmyyn.html Series Editors: Research Director Olli Gröhn, Ph.D.

Department of Neurobiology

A.I. Virtanen Institute for Molecular Sciences Research Director Michael Courtney, Ph.D.

A.I. Virtanen Institute for Molecular Sciences Author’s Address: Department of Biotechnology and Molecular Medicine

A.I. Virtanen Institute for Molecular Sciences University of Kuopio

P.O. Box 1627 FI-70211 KUOPIO FINLAND

Tel. +358 17 162 088

Fax +358 17 163 030

Timo.Liimatainen@uku.fi

Supervisors: Docent Juhana Hakumäki, M.D., Ph.D.

Department of Radiology

Kuopio University Hospital

Research Director Olli Gröhn, Ph.D.

A.I. Virtanen Institute for Molecular Sciences

Reviewers: Professor Markus von Kienlin

Hoffmann-La Roche Pharmaceuticals

Basel, Switzerland

Research Assistant Professor Harish Poptani

University of Pennsylvania School of Medicine

Philadelphia, USA

Opponent: Associate Professor Dmitri Artemov, Ph.D.

Johns Hopkins University School of Medicine

Baltimore, USA

ISBN 978-951-27-0609-9 ISBN 978-951-27-0432-3 (PDF) ISSN 1458-7335

Kopijyvä Kuopio 2006 Finland

(3)

ISBN 978-951-27-0609-9 ISBN 978-951-27-0432-3 (PDF) ISSN 1458-7335

Abstract

Cancer is a heterogeneous collection of diseases that affect many people each year in the world. It can originate from nearly every cell type from every organ. For example, gliomas (derived from glial cells) comprise the biggest group of all primary brain tumours and aggressive gliomas related deaths are common amongst patients, despite them having a low incidence in the general population. Non-surgical cancer treatment methods, such as gene therapies, are already a part of modern cancer treatment landscape. Using these methods, apoptosis is the desired way of inducing cancer cell death because of its precision in the death of the target cell and the cleanliness of the removal of the dead cell material afterwards. In order to improve the effectiveness of these therapies, it has been acknowledged that the detection of treatment responses in the early stages of the therapy has become more important. At the tissue level, apoptosis generally appears as a subtle process, but with profound changes in morphology, structure and biochemistry brought forth by the activation of distinct proteases and lipases underlie this. These carefully orchestrated molecular changes are available for visualizing apoptosis in vivo as well. Now a days, widely used clinical visualization methods such as magnetic resonance imaging, X-ray (computed tomography), nuclear imaging and ultrasound, together with optical imaging are providing new possibilities for biophysical, biochemical and biomedical research. Gene delivery can be achieved by using several methods, amongst which viruses, such as adenoviruses and retroviruses are the most potent ones. To ensure that gene transfection is efficient enough to delivery a sufficient gene therapy dose to a patient, techniques for the non- invasive, real time in vivo detection of viral vectors will be required.

In the current study, an ultra-short echo time magnetic resonance spectroscopic imaging (MRSI) sequence with water suppression was developed to measure accurately metabolites with short T2-relaxation times, such as lipids. ¹H nuclear magnetic resonance (MR) visible lipids were found to accumulate mostly in the peripheral zones of rat BT4C herpes simplex thymidine kinase- transfected gliomas undergoing apoptosis using this technique, findings which were confirmed by post mortem invasive neutral lipid histology staining of the same tumours. The majority of apoptotic cells were also found in the same regions. A quantitative relationship between membrane lipid catabolism and ¹H MR-visible lipid accumulation in vivo was also found; the biological analyses revealed that the expression of PLA2 together with catabolism of intracellular membrane phosphatidylethanolamine in BT4C gliomas was significantly increased during apoptosis. The biochemical analyses of the tumour cell membranes revealed that membrane fatty acyl moiety losses appear to be reflected in the storage lipid concentrations as observed by ¹H MRSI. Additionally, magnetic resonance spectroscopy (MRS) was used to detect and follow up the cholesterol backbone containing metabolites in BT4C non-invasively during treatment in vivo. The group of molecules containing a cholesterol backbone was found to constitute a remarkable proportion of the ¹H MRS visible lipids and thus cannot be bypassed in accurate tumour lipid analysis of MRS data. The accumulation of the cholesterol backbone containing molecules component was found to take place during the apoptotic cell death of BT4C tumours. Finally, by using unique, biotinylated ultra-small superparamagnetic iron oxide particle (USPIO) -labelled Baavi viruses, it was demonstrated that the biodistribution of these viruses in rat brain could be imaged after intraventricular injection in vivo.

With these studies several technical, biochemical, biophysical issues have been resolved in the use of molecular MRI and MRS in the fields of gene transfection and tumour gene therapy. The findings are likely to lead to better diagnostic techniques and a deeper understanding of tumour metabolism in the clinical setting.

National Library of Medicine Classification: WN 185, QZ 52, QZ 200, QU 470, QW 160

Medical Subject Headings: magnetic resonance spectroscopy; magnetic resonance imaging; diagnostic imaging;

gene therapy; genetic vectors; transfection; viruses; neoplasms; glioma; apoptosis; lipids; cholesterol; contrast media; iron; oxides; biotinylation; baculoviridae

(4)

(5)

These studies were carried out at the A.I. Virtanen Institute for Molecular Sciences, University of Kuopio, during the years 2003-2006.

I would like to express my deepest gratitude to my principal supervisor, docent Juhana Hakumäki M.D. Ph.D., who gave me an opportunity to work with a group of widely experienced specialists in the field of biomedical imaging. This heterogeneous working environment widened a lot of my perspectives in physics, biology and medicine, giving me great opportunities to continue in the field of medical imaging.

I thank also my second supervisor, docent Olli Gröhn Ph.D., who provided me a wide technical MRI/MRS experience, as well as full support in all the fields of scientific work from writing to the smallest details including the particular choice of pulse sequence.

I wish to acknowledge the official reviewers of this thesis, Professor Markus von Kienlin and Research Assistant Professor Harish Poptani for their constructive comments and evaluations of the thesis. I also owe thanks to Thomas W. Dunlop Ph.D., for revising the language of this thesis.

I wish also express my gratitude to my co-authors of the articles that are integral to this thesis, especially docent Mika Ala-Korpela for his friendship and professional collaboration that has meant a great deal to me, Ivan Tkac for sharing his wide expertise in spectroscopy with me, Jani Räty Ph.D. for co-authoring the virus imaging work, Piia Valonen Ph.D for her help in histology and tumours and Mr Kimmo Lehtimäki for fruitful discussions concerning the physical view of biological issues and cholesterol in particular. I would also like to thank people in the Department of Biomedical NMR and the current Biomedical imaging unit, especially Ms Maarit Pulkkinen for her excellent technical assistance, Teemu Laitinen MSc. and the former head of department Professor Risto Kauppinen for conversations which clarified my thinking on technical and biological issues related to magnetic resonance research.

Finally, I want to express my loving thanks to Maarit for her patience and support during this project and deepest thanks to my parents Tuula and Juhani for their continuous support, not only during my Ph.D. project but also throughout my childhood and education. Thanks are also in order to all of my friends for giving other than scientific thoughts.

This study was financially supported by the Instrumentarium Science Foundation, the Orion Corporation Research Foundation, the Finnish Cultural Foundation and the Finnish Cultural Foundation of Northern Savo.

Kuopio, November 27^th, 2006

Timo Liimatainen

(6)

(7)

AFM atomic force microscope

AMARES advanced method for accurate, robust and efficient spectral fitting ANOVA analysis of variance

ATP adenosine triphosphate

B0 external magnetic field

B1 magnetic field created by a radio frequency field Baavi avidin-displaying baculovirus

bUSPIO biotinylated ultra small superparamagnetic iron oxide particle

BW bandwidth

Cb cholesterol backbone

CCT choline cytidyltransferase CDP-Cho cytidine diphosphocholine

CE cholesteryl ester

CHESS chemical shift selective radio frequency pulse Cho choline containing compounds

CK choline kinase

CPT choline phosphotransferase Cr creatine and phosphocreatine

CT computed tomography

δ chemical shift

d damping

∆E energy difference

∆ω offset frequency

∆z pixel size

DFT discrete Fourier transformation

DMF dimethylformamide

DNA deoxyribonucleic acid

DTPA diethylene-triamine penta-acetic acid DWI diffusion weighted imaging

dUTP deoxyuridine triphosphate e noise

EPI echo planar imaging

ER endoplasmic reticulum

f frequency

FASTMAP fast automatic shimming technique by mapping along projections

FFA free fatty acid

FIR finite impulse response

FID free induction decay

FOV field of view

FWHM full-width half-maximum

γ gyromagnetic ratio

g Gaussian decay factor

GBM glioblastoma multiforme

ħ Planck’s quantum constant

hw point spread function

HDL high-density lipoprotein

HLSVD Hankel Lanczos singular value decomposition

(8)

HSVD Hankel singular value decomposition HSV-tk herpes simplex virus thymidine kinase HTLS Hankel total least squares

I quantum number

ISIS image selected in vivo spectroscopy

j molecular environment

J coupling constant

k Boltzmann’s constant

LASER localization by adiabatic selective refocusing

LD lethal dose

LDL low density lipoprotein

M magnetization

MAS magic angle spinning

MR magnetic resonance

MRI magnetic resonance imaging MRS magnetic resonance spectroscopy MRSI magnetic resonance spectroscopic imaging MUFA mono unsaturated fatty acids

n number of coupled equivalent nuclei

n+ and n- spin population in upper and lower energy stage

N number of animals

NAA N-acetyl aspartate

NMR nuclear magnetic resonance NLLS non-linear least-squares

NP number of points

NV number of phase encoding steps OVS outer volume suppression φ absolute phase of the receiver

PC phosphocholine

PCT phosphocholine cytidyltransferase PET positron emission tomography

PLA1 phospholipase A1

PLA2 phospholipase A2

PLC phospholipase C

PLD phospholipase D

PRESS point resolved spectroscopy

PtdSer phosphatidylserine

PtdCho phosphatidylcholine PtdEth phosphatidylethanolamine PUFA poly unsaturated fatty acid

ρ0 spin density

R correlation

rf radio frequency

RNA ribonucleic acid

σ shielding constant

Σ diagonal matrix with singular values in diagonal S signal intensity

S0 initial signal intensity SEM standard error of mean STEAM stimulated echo acquisition mode

(9)

SPIO superparamagnetic iron oxide particle SVD singular value decomposition

θ flip angle

τ inter pulse delay

t time (in seconds)

T temperature (in Kelvin’s)

T1 spin-lattice relaxation time T2 spin-spin relaxation time

T2* apparent spin-spin relaxation time

T1ρ spin-lattice relaxation time in a rotating frame of reference

tCho total choline

TE time-to-echo

TG triglycerides

TLC thin-layer chromatography

TLS total least square

TM mixing time

TMS tetramethyl silane

TR time-to-repetition

TSP trimethylsilylpropionate

TUNEL terminal deoxyribonucleotidyl transferase-mediated dUTP nick end- labelling

U matrix in SVD

US ultra sound

USPIO ultra small superparamagnetic iron oxide particle

V matrix in SVD

VAPOR variable power rf-pulses with optimized relaxation delays VARPRO variable projection method

VOI volume of interest

ω0 Larmor frequency

WATERGATE water suppression by gradient tailored excitation WET water suppression enhanced through T1 effects WHO world health organization

w.w. wet weight

x number of lines in a multiplet

y acquired data

y model signal

(10)

(11)

This thesis is based on the following publications which will be referred to their corresponding numerals in the thesis.

I

Liimatainen T, Hakumäki J, Tkac I, Gröhn O. (2006). Ultra-short echo time spectroscopic imaging in rats: implications for monitoring lipids in glioma gene therapy. NMR in Biomedicine 19, 554-559.

II

Liimatainen T, Erkkilä A, Valonen P, Vidgren H, Lakso M, Wong G, Gröhn O, Ylä-Herttuala S and Hakumäki JM. ¹H Magnetic Resonance Spectroscopic Imaging of Phospholipase-mediated Membrane Lipid Release in Apoptotic Rat Glioma In Vivo, submitted.

III

Liimatainen T, Lehtimäki K, Ala-Korpela M, and Hakumäki J. (2006). Identification of mobile cholesterol compounds in experimental gliomas by ¹H MRS in vivo: Effects of ganciclovir induced apoptosis on lipids. FEBS Letters 580, 4746-4750.

IV

Räty J, Liimatainen T, Wirth T, Airenne K, Huhtala T, Ihalainen T, Hamerlynck E, Vihinen- Ranta M, Närvänen A, Ylä-Herttuala S, Hakumäki JM. (2006). Magnetic resonance imaging of viral particle biodistribution in vivo. Gene Therapy 13, 1440-1446.

(12)

(13)

1 Introduction

Cancer is a widely spread disease in western society. The World Health Organization (WHO) estimates that over 11 million people are diagnosed with cancer and 7 million people die of cancer-related diseases during one year, which is 12.5 % of the total deaths world wide (www.who.int/cancer). Gliomas comprise almost half of all primary brain tumours with glioblastoma multiforme (GBM) is the most aggressive (Levin et al., 2001). Despite the low appearance (2 %) of all brain tumours in respect to the total number of cancer cases, GBM is the leading and fifth leading type of tumour related death in men and women, respectively, in the 20- 39 year old age group in the UK (Jemal et al., 2003; Levin et al., 2001). Amongst patients bearing a malignant brain tumour (GBM), the forecast of remaining lifetime is expected to be approximately 1 year after diagnosis.

Amongst the non-invasive non-surgical cancer treatment methods, such as gene therapy, one type of programmed cell death, apoptosis, is the desired way in which to induce cell death. The detection of treatment responses at an early stage has become more important, especially with new treatment strategies. Although at the tissue level, apoptosis generally appears as a subtle process, the individual cells experience profound changes in their morphology, structure and biochemistry, brought forth by the activation of distinct proteases and lipases. Biophysical changes and the temporal timeline of molecular cell events provides a wide array of markers and targets that could be used for visualization of apoptosis using modern imaging modalities.

Traditionally, the assessment of apoptosis in tissue is based on microscopic methods often assisted with specific histochemical staining procedures. However, microscopic sections must be taken from tissue by biopsy or post mortem. Clearly, these are invasive and set limitations to many follow-up studies. Present in vivo imaging technologies rely mostly on non-specific macroscopic physical, physiological, or metabolic changes that differentiate pathological from normal tissue rather than identifying specific molecular events responsible for disease or treatment.

Molecular imaging methods have been applied widely in the clinical world for several decades.

More recently technical developments have widened the usefulness of these methods to the imaging of small experimental animals. Molecular imaging as such, can be defined as the in vivo characterization and measurement of biologic processes at the cellular and molecular level (Weissleder and Mahmood, 2001), which usually includes imaging of, gene transfer and delivery, molecular events in diseases, as well as, the detection of labelled cells or other particles. New gene therapy methods (Pulkkanen and Ylä-Herttuala, 2005) are developing at a rapid pace. The aim of gene therapy is to deliver genetic material with therapeutic potential into a target tissue. For delivery, several methods can be used, amongst which, viruses, such as adenoviruses and retroviruses are the most potent (Ram et al., 1993; Vincent et al., 1996).

Therefore, ensuring efficient, therapeutically effective gene transfection in vivo will require the non-invasive, real time detection of viral vectors.

The aims this current study can be divided in four main objectives:

1) to develop a magnetic resonance spectroscopic imaging method to optimally detect lipids as apoptotic markers in malignant tissue with high spatial resolution

2) combine magnetic resonance spectroscopic imaging with biochemical data to obtain more information on the physiological basis of apoptosis-related magnetic resonance visible lipid accumulation

(16)

Kuopio Univ. Publ. G. –A.I. Virtanen Institute for Molecular Sciences 50: 1-81 (2006)

3) to explain the true constitution of the ¹H magnetic resonance spectroscopy detectable lipid spectrum during gene therapy induced apoptosis and finally,

4) to develop a magnetic resonance imaging technique for the imaging of virus biodistribution in the normal rat brain to study with the aim of using it as a surrogate marker for the imaging of gene expression as well as measuring the efficiency of gene transfection in any particular gene therapy.

(17)

2 Literature overview

2.1 The theoretical basis of NMR

When nuclei with angular momentum (also called nuclear spin) that are non-zero are placed in an external magnetic field, a net magnetization is created by the equilibrium difference between magnetic dipole-dipole moments existing in a large spin population. The number of nuclei in two different energy states n+ (lower) and n- (upper) obey the Bolztman’s distribution

kT E

n e

n ⁻^∆

+

− = , [1]

where ∆E=γħB0 (γ being the gyromagnetic ratio, ħ is Planck’s constant and B0, the external magnetic field) is the quantized energy difference between the lower and the upper energy states, k is Boltzman’s constant and T is the temperature (in Kelvin’s). The magnitude of the net magnetization dependency on the spin density ρ0, the gyromagnetic ratio γ, temperature T and the main magnetic field B0 is illustrated by the following equation:

0 2 2 0

0 4 B

M ρ γkTh

= . [2]

The resonance frequency (Larmor frequency) ω0, which allows energy absorbance between energy states, is proportional to B0, and γ

0

0 γB

ω = . [3]

The Larmor frequency for hydrogen is γ(¹H) = 42.58 MHz/T. When an energy burst is inserted into the sample with the Larmor frequency, spin equilibrium can be disturbed. The Larmor frequency in the magnetic fields currently used in magnetic resonance (MR) (0.5 – 21 T) is in the radiofrequency range, and a given energy burst is called a radiofrequency pulse (rf-pulse).

Immediately after the disturbance, the spin system begins to relax towards the initial equilibrium state. Two separate relaxation mechanisms can be observed, namely spin-lattice and spin-spin types. Spin-lattice relaxation includes interaction of the spins of the nuclei with the surrounding material (i.e. lattice and spin energy is absorbed by the lattice). When a spin-spin relaxation occurs, energy is exchanged due to interactions between nuclei only. The nuclear magnetic resonance (NMR) signal is damped due to a lose in phase coherence, which is caused by small differences in their precessional frequencies. Bloch equations describe the spin lattice, with a relaxation time T1 and the spin-spin relaxation with T2 relaxation times are obtained from the use of the differential equation group:

1 0 1

2 0

1

2 0

) / (

T M B M

dt M dM

T B M

M B dt M

dM

T B M

dt M dM

z y

z

y x

z y

x y

x

− −

−

=

−

=

−

=

γ

γ ω γ

γ

γ ω γ

, [4]

(18)

where the subscripts x, y and z refers to the spatial directions (where z is direction of B0) and B1

is a magnetic field, produced by the rf-pulse. In general, signals would suffer additional suppression due to dephasing derived from external static magnetic field inhomogeneities. In the case of the spin-spin relaxation, the T2 relaxation time is replaced by a shorter relaxation time, T2*<T2. However using “rephasing” or “echoing”, this source of signal dispersion can be avoided. Without disturbing the spin system, a raw NMR signal is not measurable, because the net magnetization is weak and is parallel to the stronger external field B0. However, if the net magnetization is disturbed by an additional energy burst (a rf-pulse) with the Larmor frequency, a free induction decay (FID) can be measured after switching off the rf-pulse. An ideal FID from one particular substance, at time t, is formed by an exponentially decaying spin-spin relaxation induced terms and sine-form with Larmor frequency,

2*

/

) 0

(t S eⁱ eⁱ ^te ^t ^T

S = ^ϕ ^∆^ω ⁻ , [5]

where S0 is the initial signal intensity after excitation, φ is the absolute phase of the receiver, ∆ω is the offset frequency from Larmor frequency and t is time.

A NMR spectrum can be defined as a plot of the radiofrequencies that nuclei emit, once they are excited. The spectrum is the amplitude and the phase content of certain frequencies of measured FID signal. Using Fourier transformation to treat the data, one can convert a complex number FID signal into a complex number spectrum, where ideally exponential decaying FIDs have a Lorentzian-line shape with full-width half-maximum can be represented as Eq. [6].

FWHM _*

2

1 πT

= . [6]

The real part of a spectrum is called the absorption spectrum and the imaginary part, the dispersion spectrum. Instead of a single Lorentzian resonance, real NMR spectra include multiple resonances. These resonances arise due to the interaction of the nuclei with other nuclei in the surrounding environment. Therefore, resonating nuclei such as ¹H can be located in different molecular environments jx, where the subscript x denotes environments 1, 2, 3, etc.

shifts in the spectra occur because dissimilar electron densities exert small magnetic fields with different amplitudes around environments j1 and j2. Thus, a nucleus in environment j1

experiences a magnetic field Bj1 and nucleus in j2 a magnetic field Bj2 instead of main field B0. Greater and lesser electron density can also be described as shielded, and deshielded because of their relative ability to present resonances in NMR spectra. A quantitative variable between reference species and shielded or deshielded compound j is called chemical shield constant σ and it is defined by the following equation

) 0

1

( B

B_j = −σ_j , [7]

where Bj is a magnetic field modulated by chemical environment j. A new resonance frequency for nucleus in an environment j can now be defined by Eq. [8].

) 2 (B⁰ _jB⁰

j σ

π

ω = γ − . [8]

(19)

Therefore, the resonance frequency (ωj) of nucleus in molecule j depends on the chemical environment surrounding it, as seen in Larmor equation [3]. Generally, a “part per million”

(ppm) scale is used when chemical shifts are discussed. The chemical shift δ is defined using reference resonance with frequency ωref

6 0

×10

= − ω

ω

δ ω ^ref . [9]

The reference frequency is usually defined using standard chemical compounds, such as tetramethylsilane (TMS) or trimethylsilylpropionate (TSP). The ¹H spectrum resonance frequencies of these compounds are of particular use because of their very low electronegativity values (Cowan, 1997). These are not the only compounds used and wider list of 0 ppm reference compounds is given elsewhere in the literature (Certaines and Bovee, 1993). The ppm scale is, for historical reasons usually represented according to the traditional rule of spectroscopy, where frequency decreases to the right.

Close observation of NMR spectra can reveal fine structure in certain resonances as well as multiplicity of resonances, which are only couple of Hz apart from each other. This is not in contradiction with chemical shift, but indicates the existence of a magnetic interaction amongst the protons of two chemical groups. This interaction is generally defined as nuclear coupling.

The intensity of coupling is defined by a coupling constant J. The unit for the coupling constant is the Hz and this quantity is not affected by the intensity of the external magnetic field B0. In low viscosity solutions, values for J are normally not more than 100-300 Hz. In particular, the homonuclear proton (¹H) coupling (that only occurs between protons) is smaller than 20 Hz.

With the assumption J<<∆δ, the number of lines in which a signal is split is related to the number of coupled equivalent nuclei by the equation,

1

2 +

= nI

x , [10]

where x is the number of lines of the multiplet, n is the number of coupled equivalent nuclei and I is their quantum number. Similarly, the relative intensities of signals in multiplet can be obtained using Pascal’s triangle. However, spectrum simulations are needed where the condition J<<∆δ is not fulfilled and when the coupled spin system is more complicated.

The natural occurrence of ¹H is nearly 100% in terms of hydrogen atoms and nuclei. This very high abundance with I = ½ property and the wide spectrum of molecules and molecular groups that contain protons in living material, makes NMR a practical tool for use in both biology and medicine. However, the spread of the chemical shifts in ¹H NMR spectrum is only about 10 ppm, this together with the enormous amount of chemical environments where ¹H can exist, makes ¹H spectra relatively complex to analyse, due to overlapping signals. Information can be obtained however by using both modern analytical and spectral fitting methods. A further inconvenience is also imposed on the analysis of proton spectra by the considerable occurrence of homonuclear coupling events. The natural abundance of other NMR visible nuclei such as magnetically active

13C nuclei is 1.1 % of total carbon population, and makes heteronuclear coupling negligible event when considering ¹H spectra, as it is not visible under normal conditions.

(20)

2.2 Spectral processing

After the measurement of the FID, processing is needed to reveal the chemical shifts (frequencies), phases, and amplitudes of the signals present that enable analysis based on these parameters. The spectral analysis methods can be divided into groups depending on which factor, whether it be space, time or frequency, the processing is done. In addition to this analysis, spectra may need pre-processing in order, to remove unwanted properties and imperfections from the data caused by the measurement event.

The realistic signal model equation for a FID represents a Lorentzian lineshape in the time domain

n K

k

t f i d j k

n a e e e

y

∑

^k ^k ^k ⁿ

=

+

− +

=

1

) 2

( π

ϕ , [11]

where n = 0…N-1, K refers to the number of signal, ak is proportional to the nuclei that resonates with the frequency fk and dk is the damping of the k^th signal. The sampling time points tn differs from the ideal time points by the receiver dead time t0, which leads to tn=n∆t+t0. The term i2πfk

is the so called first order phase, φk= φ0+φ´k where φ0 is the zero-order phase and φ´k the additional phase factor (which is usually zero). The noise term, en, is assumed to be a complex Gaussian noise without correlation between its real and imaginary parts and equal variance.

However, this is model for ideal NMR signals. Under practical conditions, signals include a variety of imperfections caused by external magnetic field, B0, inhomogeneities, field temporal shifts amongst other causes. Therefore, under practical conditions, the signal can not be perfectly represented using only one damped complex sinusoid and therefore an alternative model to represent imperfect signals is to use Gaussian lineshape model

n K

k

t f i t g j k

n a e e e

y

∑

^k ^k ⁿ ^k ⁿ

=

+

− +

=

1

) 2

( π

ϕ , [12]

where gk is the Gaussian decay factor. The combination of these two has been used as well and the special name for this combination is the Voigt model

n K

k

t f i t g d j k

n a e e e

y

∑

^k ^k ^kⁿ ^k ⁿ

=

+

−

− +

=

1

) 2

( π

ϕ . [13]

Equations [11-13] include a huge number of parameters, which should be estimated in order to characterize the K signals. However, the number of parameters can be enormously decreased using a priori knowledge in the estimation process. In MR spectral analysis the following a priori information can widely include: the known frequency differences between resonances and between peaks included in multiplets, dampings as well as intensity ratios between and inside multiplets and phases. The choice of a priori information to be included depends on some assumptions, which must be evaluated for individual applications. Generally, a prior information can be derived using phantom spectra of a pure metabolite solution under similar measurement conditions used to obtain the experimental spectra (Bartha et al., 1999; Mierisova et al., 1998).

Model spectrum can also be derived from MR spectrum derived from pure metabolite preparations (Provencher, 1993) or performing spectral simulations (Young et al., 1998).

(21)

However, before spectral analysis, with or without a priori knowledge, some preprocessing of the spectra may be need.

2.2.1 Spectral preprocessing

The goal of the spectral preprocessing is to correct for imperfections in the data and thus improve the quality of the quantifiable results. The preprocessing can be performed whether or not the actual quantification is performed in the time or in the frequency domain. The preprocessing contains several independent methods, such as multiplication with non-linear functions in the time domain, and the removal of unwanted components.

Signal-to-noise ratio (SNR) as well as spectral resolution can be enhanced for the purpose of visual inspection of the MR spectrum by multiplying the FID using a non-linear function in the time domain. Multiplication with a decreasing mono-exponential function causes the attenuation of the spectrum points in the end of the FID and increases the FWHM of the peaks in the spectrum. In addition, multiplication makes the spectrum look less noisy, but this is expense of spectral resolution. Opposite results are obtained by multiplying the time domain data with an increasing exponential function. However, after multiplication, an assumption, of the equal variance of Gaussian noise in every data point, is no longer valid. This can be taken into account in the analysis as exemplified in the work done by Bartha and co-workers (Bartha et al., 1999).

However, it should be noted that decreased parameter accuracy, due to line broadening, was reported for this pre-processing technique in the same publication.

Elimination of the unwanted features prior to quantitation can be favourable in some cases. Their removal may reduce the computational burden on the estimation phase and improve the accuracy of the parameter estimates. The earliest way to do this is based on filtration of the spectra, especially in cases where the components of interest are well separated from the unwanted ones.

Usually, the water signal is located at zero frequency in the NMR spectra (or it can be frequency shifted to zero frequency). This feature was applied when first and second order differentiation was used to remove the water signal (Kuroda et al., 1989). With this method, visual inspection of the filtered data reveals that, in most cases, the water signal is successfully removed. However, these filters are very simple and have strong affects on metabolite peaks, as well. Simple filter based methods suffer from the strong influence of user defined filter design and filter order.

Furthermore, the filter influence on the peaks of interest is not taken into account and this limits the accuracy of the estimates. However, more sophisticated methods have been developed based on finite impulse response (FIR) filters. This method, introduced by Sundin and colleagues (Sundin et al., 1999), uses maximum-phase FIR filters to achieve solvent signal suppression and an iterative non-linear least-squares (NLLS) algorithm for parameter estimation. The estimation algorithm takes the filter influence on the metabolites of interest into account and therefore can utilize a large variety of prior knowledge.

An alternative way to do unwanted data removal involves modelling of the unwanted resonances.

The idea is to create a model for the undesired signal (such as that derived from the water) and subtract from the original spectrum. However, this is not without difficulties. For example the water peak is often distorted due to field inhomogeneities and/or partial water suppression and thus, usually, a single peak is not enough to describe the full form of the water resonance. Black- box algorithms with several damped exponentials (typically between 5 and 10 in number) have been successfully used to reconstruct accurate models of water resonance spectra derived from tissues. Algorithms used include the Hankel singular value decomposition (HSVD) (van den Boogaart et al., 1994b), the Hankel total least squares (HTLS) (Leclerc, 1994) and the more

(22)

computationally efficient algorithms such as the Hankel Lanczos singular value decomposition (HLSVD) (Pijnappel et al., 1992). Please note that the principles of these are introduced in paragraph entitled Time domain analysis and will not be discussed here. The fitting method in the black-box approaches uses relatively high model orders and the parameters of damped exponential are found in the frequency region of the water peak and the signal derived from this molecule is reconstructed using these exponentials. The only knowledge, which the user has to specify, is the model order and the region of the water peak that is to be removed. In some cases these choices have been observed to influence the estimated parameters of the peaks of interest (Vanhamme et al., 2000). The clear advantage of these fitting methods, compared to filtering, is that fitting takes into account the signal “tails” under the estimated resonances.

2.2.2 Time domain analysis

The MR data is acquired in the time domain; therefore it would be natural to analyze it in the same domain without swapping into the frequency domain by using a discrete Fourier transform (DFT). However, visual interpretation of the measured and fitted results is the best done in the frequency domain. A frequency domain is useful for adding a priori information for time- domain analysis. In time domain analysis, quantification algorithms are based on non-linear minimization of functional

y 2

y− , [14]

where ⋅ denotes the Euclidean vector norm and y=[y0,…,yN-1]^T is the acquired magnetic resonance spectroscopy (MRS) data and y=

[

y₀,K,yN₋₁

]

^Tis the modelled signal and the superscripted ^T denotes the transpose of the vector. The problem [14] can be divided into linear and nonlinear parts as described in more detailed in the literature (Vanhamme et al., 2001). After separation, the cost function includes fewer parameters because the linear ones are eliminated.

However, the solutions of the separated and original problems are identical. Both, original and separated NLLS problems can be solved by using, either local or global, optimization theories.

The search for the minimum using local minimization methods in multidimensional space often results in a local minimum instead of a global one. Global optimization methods have been applied to in vivo MR spectral analysis in order to circumvent this problem (Metzger et al., 1996;

Weber et al., 1998). However, the problem with these global methods is poor computational efficiency, which has limited their use in practical MRS studies. By choosing suitable starting values (a priori information), reliable results are achieved using local minimization methods in reasonable time intervals. To feed a priori information into the algorithm includes a lot of input from the user and thus, methods of this type, are called interactive methods. However this can be eliminated by using black-box methods to provide starting values (Potwarka et al., 1999). An examples of interactive methods are variable projection (VARPRO) (van der Veen et al., 1988), which minimizes separated problem and uses modified version of Osborne’s Levenberg- Marquardt algorithm (Osborne, 1972). The advanced method for accurate, robust and efficient spectral fitting (AMARES) method improves VARPRO in terms of robustness and flexibility (Vanhamme et al., 1997). AMARES also uses sophisticated NLLS to minimize function [14], as well as allowing the imposition of more types of a priori knowledge and the fitting of echo signals. Gaussian and Lorentzian lineshapes can be chosen for each peak separately as well as for imposing upper and lower boundaries on the parameters.

(23)

Black-box methods can be separated into two main categories: those using a linear prediction principle and others described as state-space formalism. In state-space based methods, the data is arranged in Hankel matrix as follows:

⎥⎥

⎦

⎤

⎢⎢

⎣

⎡

=

−

+

1 1

1 2

1

1 0

N M

N M N

M M

y y

y

y y

y

y y

y H

L M O M M

L L

, [15]

with M ≥K,N −M >K. For the analysis of MRS spectra using Kung’s and co workers method please refer to the work of Barkhuysen and colleagues (Barkhuysen et al., 1987). The singular value decomposition (SVD) of H=UΣV^H, where Σ is a diagonal matrix containing singular values Σ1…Σp with condition Σ1≥K≥Σ_p, p is the rank of matrix H. H is chosen to be as square as possible to maximize parameter accuracy. The matrix H is truncated into rank K

HK=UKΣKVKH, [16]

where UK and VKH are the first K columns of U and V respectively, and ΣK is the K×K submatrix of Σ. Using UK, a set of equations are created by

Q U

U_K^↑ ≈ _K^↓ , [17]

where the up and down arrows stand for deletion of the top and bottom rows, respectively.

Matrix Q can be estimate by solving equation [14] in a least square sense, as described for HSVD (Barkhuysen et al., 1987). The eigenvalues of the latter matrix forms a set of the estimateszˆ , _k k=1,…,K. To calculate the amplitudes and phases, these estimates are back- substituted into the model equation [11] and fitted to the original data with the least square sense.

To enhance the accuracy of HSVD, the HTLS algorithm can be used to compute the total least square (TLS) solution of equation [14]. The HLSVD (Pijnappel et al., 1992) algorithm computes only a part of the SVD by using the Lanczos algorithm. The advantages of black-box methods are the minimal user interaction combined with the fact that all of the parameters are estimated in one step, without the need for starting values. Some variants of these methods can use a priori information as well, such as method introduced by Chen and colleagues (Chen et al., 1996). The inclusion of prior information is a crucial point when the method is applied to in vivo spectral analysis. In general, black-box methods provide a very good mathematical fit to the data.

However, the parameters determined by these methods may lack of physical meaning, especially in instances with low SNR. Non-Lorentzian peaks are represented as a superposition of several Lorentzian lineshapes, as for water signal removal in preprocessing using black-box methods.

2.2.3 Frequency domain analysis

The oldest form of frequency domain analysis is the integration of an area under the peak of interest revealed in the Fourier transformed time-domain signal (Meyer et al., 1988). Integration includes no assumptions concerning the lineshape of the signal, but it suffers from low estimation accuracy. Reliable estimation by integration needs an appropriate phasing and depends on the frequency range around the peak, especially when multiple peaks are not well separated. Integration analysis is also blind to missing time domain data points and acquisition

(24)

artefacts. Unidentified broad resonances can also distort the frequency domain baseline. This may result in decreases in the reliability of the integrated estimates and especially biological systems integration suffers from heavy overlaps in both the signals and the wavy baseline, as well as low SNR. Model signal based methods in the frequency domain are equivalent to the time-domain fitting methods from a theoretical point of view. In frequency domain analyses, Gaussian and Voigt lineshapes, as well trying to account for missing data points and non- uniform sampling are all more difficult to achieve than in time domain analysis.

Typical in vivo NMR data consists of many overlapping peaks which are difficult to resolve due to low SNR and broad peaks. Several studies have shown that inclusion of a priori knowledge into lineshape fitting analysis is essential for reliable quantitation in both frequency and time domains. This improves the reliability of the results especially in case of heavily overlapping peaks (Mierisova et al., 1998; Tkac et al., 1999; van der Veen et al., 1988). In addition, comparisons between time and frequency domain analysis, using model signals, have been shown mostly to yield consistent results (Kanowski et al., 2004; van den Boogaart et al., 1994a).

Model signal analysis uses, in a regular basis, Lorentzian lineshapes to model MRS resonances and an additional term to describe the baseline, which is commonly polynomial in nature.

Inhomogeneities in the sample and instrument imperfections both cause distortion to the ideal lineshape of a MR signal. Therefore lineshapes such as Gaussian and Voigt are incorporated into the fitting paradigm, in a similar fashion to that done in time domain analysis. To estimate the spectral parameters, such as frequencies, line widths, intensities and phases, various algorithms can be applied, including a Levenberg-Marquardt algorithm (Hiltunen et al., 1991; Marquardt, 1963). A modified Gauss-Newton algorithm (Stoer and Bulirsch, 1983) is used in an advanced lineshape fitting program, TLS that is included in the PERCH software package (Laatikainen et al., 1996). TLS performs lineshape fitting in the frequency domain and allows for the estimation of frequencies, intensities, line widths, and the phases of individual peaks with the possibility to add a priori information as well. It has been developed for the purpose of high resolution NMR spectral analysis (Laatikainen et al., 1996), but it has also found applicability in the field of biochemical MRS (Korhonen et al., 1998) and in vivo applications (III). TLS allows for the specification of model lineshapes as combinations of Lorentzian components. A model lineshape can contain several individual Lorentzians with either absolute or relative constraints on the half line widths, intensities, and/or resonance frequencies. Fixed values or relative values can be used to form different kind of model lineshapes. Therefore, in summary, estimation of the half line widths, intensities and resonance frequencies of the specified model resonances can be achieved for every spectrum whilst maintaining the same general lineshapes.

The LCModel (for linear combination model) was first introduced for analysis of in vivo proton MR spectra derived from the human brain (Provencher, 1993). In the LCModel analysis of an in vivo spectrum, maximum information and uniqueness is reached using a basic set of complete model spectra. A constrained regularization method accounts for differences in phase baseline and lineshape metabolite concentrations and their uncertainties (Cramér-Rao lower bounds). The only input for LCModel is the time domain FID signal (when a full set of in vitro metabolite spectra is available) that keeps interaction with the user to a minimum and subsequently leads to less subjective results. Remarkably, even concentrations of minor metabolites can be estimated using the LCModel with high internal precision. High accuracy combined with minimal user interaction, are clear advantages of this type of analysis. A more detailed description of the capabilities of the LCModel is described in literature (Provencher, 2001) as it’s mathematical basis in the original publication (Provencher, 1993).

(25)

2.3 MR signal localization

The simplest experiment for measuring an NMR spectrum from a certain sample is to excite the sample using an rf-pulse and acquire a FID signal. This approach collects a spectrum from whole volume excited and that data from different locations cannot be separated from the acquired FID.

However, using rf-pulses simultaneously with magnetic field gradients, the spin system can be manipulated so that only a desired location is excited. In MR imaging and spectroscopy rf-pulses and gradients are used for signal localization and manipulation. The specific order and temporal arrangement of these components is called a pulse sequence.

As the Larmor equation [3] and the ideal FID [5] shows, the data from a MR experiment is formed by frequency components. These components can be made spatially dependent by adding a gradient into the main external magnetic field, B0. A selective rf-pulse, applied simultaneously with a precise gradient allows for the selection of a particular one-dimensional slice. With the addition of two or even more orthogonal gradients (in the other physical dimensions, relative to the first gradient) with selective rf-pulses, a single volume element, usually called a volume of interest (VOI), can be selected. This approach is used widely in single voxel spectroscopy sequences, such as point resolved spectroscopy (PRESS) (Bottomley et al., 1984) and stimulated echo acquisition mode (STEAM) (Frahm et al., 1989). PRESS consists of one 90˚ pulse and two 180˚ pulses with the scheme 90-TE/4-180-TE/2-180-TE/4-Acq, where TE denotes the time-to- echo. The STEAM pulse sequence consists of three 90˚ rf-pulses with a construction of 90-TE/2- 90-TM-90-TE/2-Acq, where TM is the mixing time.

When a slice selective rf-pulse is applied, both the gradient strength and direction defines the excited slab. If slab selection is based on some reference compound, usually water, all the other specimens with different chemical shifts are shifted in the direction of the applied gradient. The amount of chemical shift displacement can be calculated by dividing the chemical shift in Hertz’s by the gradient strength. Consequently, to minimize chemical shift displacements, as strong as possible gradient strength should be used.

2.3.1 Point resolved spectroscopy

PRESS (Figure 1 A) includes three rf-pulse gradient pairs with orthogonal gradient alignment.

This design forms a VOI at the intersection of three slices. Spins will remain coherent only in the intersection of slices and form a detectable echo, whilst outside spins will either be not exited at all or not refocused. Since individual slices are chosen, by using a selective rf-pulse simultaneous with a gradient, each individual pulse selects one slice. The timing of the rf-pulses in PRESS follows at double spin echo (Hahn, 1950) (Figure 1 A). The first selective rf-pulse (90˚) excites spins in the slice, followed by a gradient refocusing lobe to maintain the excited spins in phase.

After a delay in time equivalent to TE/4, the orthogonal slice is refocused by the second selective rf-pulse (180˚). After this second rf-pulse only the intersection of these two slices is both excited and refocused. To eliminate signals excited by the 180˚ pulse, identical crusher gradients on each side of the refocusing pulse are used. The phase added by the first gradient is rewound by the second crusher gradient and therefore spins which experienced a 180˚ flip angle, keep their phase, but spins outside of bandwidth of the 180˚ pulse experience dephasing by both gradients.

Thus spins excited by the 90˚ pulse, but not refocused by the first 180˚ pulse are dephased and do not contribute to the detectable MR signal. The third pulse is identical to second one, but is again applied orthogonally with respect to the two previous slice selections. After the third pulse, only the intersection of the three planes is excited and refocused and this forms the detectable MR signal derived from the VOI.

(26)

PRESS is a single shot technique, which makes it insensitive to motion compared to signal localization image selected in vivo spectroscopy (ISIS). Applications for PRESS can be found in the clinical spectroscopy of metabolites with relatively long T2 relaxation times such as the choline containing compounds (Cho) creatine and phosphocreatine (Cr) as well as N-acetyl aspartate (NAA). One advantage of PRESS lies in the short delay experienced between dephasing and rephasing crushers. This minimizes the diffusion weighting of the sequence.

However, there are disadvantages associated with this technique as well. For instance, PRESS is sensitive to flip angles, because 180˚ degree pulses are used. If all pulse calibrations (90˚ and two 180˚) are based on a 90˚ pulse calibration, the maximal signal is decayed by sin⁵θ, where θ is the flip angle of calibrated 90˚ pulse. This feature plays a crucial role when an inhomogeneous rf- induced magnetic B1 field is applied, for example when one uses surface coils. In PRESS, the minimal TE value is limited by the time necessary to transmit the two refocusing pulses and crushing gradients, whilst maintaining the appropriate delays for echo formation. Thus it is difficult to achieve short echo times when using PRESS. Longer inter-pulse delays allow transverse relaxation to occur, as well as J-modulation, which distorts both the shape and the amplitude of the signals derived from coupled spin systems.

2.3.2 Stimulated echo acquisition mode

STEAM, like PRESS, includes three rf-pulses but, in STEAM, all three pulses have the same flip angle (Figure 1 B). The signal in STEAM is a stimulated echo instead of a spin echo and at least three rf-pulses are required to form the stimulated echo. To obtain a maximal signal, an optimally flip angle of 90˚ degrees should be used for all three pulses. The first pulse in STEAM is for the excitation of a net magnetization from the longitudinal axis to a transversal plane. After the first rf-pulse, transversal magnetization dephases until time point TE/2, when a second pulse is applied. This pulse returns 25 % portions of the net magnetization to both parallel and antiparallel positions relative to, the external magnetic field, B0. Even though, magnetization is in the longitudinal axe, phase information remains in the spins. The 50 % of the net magnetization not affected by the second rf-pulse is left in transversal plane and these spins either dephase or become suppressed by the crusher gradients. During the TM delay, the signal decays obeying the T1 relaxation time instead of T2 relaxation. This enables water suppression and the addition of a crusher gradient in the TM period (Moonen et al., 1989; Tkac et al., 1999). In diffusion studies, relatively long diffusion times are needed to separate diffusion gradients from each other and the TM period enables this with minimal relaxation effects (T1) (Kärger et al., 1988). The last of the three rf-pulses returns the longitudinal magnetization to the transversal plane. Since phase information has remained in the spins, they are rephased after the last rf-pulse and form a peak of stimulated echoes after time equal to TE/2 has elapsed from the last pulse. In STEAM, selective rf-pulses can be used in the way as they were used in PRESS and the voxel can be defined by using three selective pulses with gradients (Figure 1 B).

Importantly, using STEAM, very short echo times can be used, which reduces T2 effects and the distortion caused by J-modulation. Just like PRESS, STEAM is a single shot technique.

However, STEAM includes three 90˚ pulses which make it less sensitive to flip angle errors compared to PRESS with 90˚ and two 180˚ pulses. The dependence for flip error is sin³θ instead of sin⁵θ in the case of PRESS (Moonen et al., 1989). The biggest disadvantage of STEAM is that only half of the net magnetization can be measured. When neglecting relaxation and flip angle errors, the lack of 50 % in the maximal signal leads to roughly four times more averaging. The reduction of relaxation effects (short TE) leads to an advantage when metabolites with short T2

and coupled spin systems are present, such as lipids. The simultaneous quantification of 18

Molecular magnetic resonance imaging of gene therapy-induced apoptosis and gene transfer: A role for 1H spectroscopic imaging and iron oxide labelled viral particles (Spektroskooppisen magneettikuvantamisen ja rautaoksidi-leimattujen virusten rooli geenit