| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Reviews |
1 From the Department of Radiology, Maastricht University Hospital, P. Debyelaan 25, 6202 AZ Maastricht, the Netherlands; and Biomedical NMR, Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, the Netherlands. Address correspondence to J.F.A.J. (e-mail: j.f.a.jansen{at}tue.nl).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
© RSNA, 2006
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
In nuclear MR, the total area under a metabolite resonance in a 1H MR spectrum is directly proportional to the concentration of the metabolite. Currently, metabolite concentrations are usually expressed as ratios (relative quantification) rather than as absolute concentrations. Ratios can be useful for clinical diagnosis to characterize pathologic tissue. For example, relative quantification has been successfully applied in the diagnosis of cancer (2), leukemia (3), epilepsy (4), dementia (5), and multiple sclerosis (6). A locally obtained reference data set with normal ratios for healthy tissue can be used for most clinical purposes, while for focal pathologic conditions a reference spectrum can be acquired from the contralateral region of the patient's brain. The use of signal ratios has the great advantage that it is very easy to implement because it does not require extra imaging time and time-consuming postprocessing. Furthermore, a number of problems—for example, partial volume effects arising from different amounts of cerebrospinal fluid in the selected volumes—can be largely avoided. In addition, ratios are, in some cases, more sensitive in terms of detecting changes (eg, when one metabolite increases and another decreases) and can be more accurate than absolute concentrations, owing to specific characteristics of the analysis computer program applied (7).
If a change is observed in the ratio of metabolite peaks, however, it remains uncertain which metabolite concentration actually changes. This information can only be obtained from absolute concentrations. Absolute quantification (AQ, also referred to as absolute quantitation) implies that concentrations are expressed in biochemical units, such as millimoles per kilogram wet weight. Furthermore, when spectroscopic results are to be compared in an interdisciplinary context (eg, with biochemically derived concentrations), ratios or arbitrary units might not be appropriate.
Since many factors need to be considered to obtain reliable absolute concentrations with 1H MR spectroscopy, it is hard for a newcomer to comprehend all aspects and potential pitfalls concerned with AQ and to decide which method to use. The latest reviews on AQ were written several years ago (8–11), when AQ was more or less considered to be a basic research tool. However, several groups have recently published clinical trials that incorporate AQ. These include studies of patients with Alzheimer disease (12), epilepsy (13–15), multiple sclerosis (16,17), leukoencephalopathy (18), amyotrophic lateral sclerosis (19), and cancer (20). Furthermore, the reviews published so far have mainly focused on single-voxel spectroscopy, whereas spectra simultaneously recorded from multiple adjoining spatial regions (chemical shift imaging) find increasing use in many clinical trials (13,15,17,21). In addition, new developments, such as spectral editing, macromolecule suppression, tissue segmentation, and quantitative analysis of spectra with dedicated software, can aid in AQ of brain metabolites. This review will address the implementation of these new developments in AQ strategies.
|
ADVANTAGES AND DISADVANTAGES OF ABSOLUTE METABOLITE CONCENTRATIONS |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
The assumption that the concentration of certain reference metabolites (eg, total creatine, choline) remains constant may be incorrect under normal conditions, as well as in many pathologic states. For example, it has been shown in patients with temporal lobe epilepsy (22) that the temporal lobe ipsilateral to the seizure focus displays a significant increase in total creatine and choline; a similar finding has also been reported in patients with frontal lobe epilepsy (23). Furthermore, even the temporal lobe contralateral to the seizure focus exhibits total creatine and choline concentrations that are increased, albeit not significantly (22). In addition, if patients have global metabolic defects, comparisons with contralateral brain regions (which are assumed to be metabolically normal) may not be feasible (24,25).
Clear evidence has recently been provided showing that, in addition to causing possibly ambiguous interpretation, relative quantification may introduce larger errors than does AQ (26,27). Two independent 1H MR spectroscopy studies with healthy volunteers in whom single-voxel spectroscopy (27) and chemical shift imaging (26) were used have shown that metabolite ratios exhibit higher coefficients of variation (up to 1.6-fold) than do AQ concentrations of individual metabolites. Furthermore, authors of a chemical shift imaging study in patients with temporal lobe epilepsy (15) reported that metabolite ratios were less sensitive to abnormalities than were absolute values. It is, therefore, advisable to obtain concentrations expressed in standard units (such as millimoles per kilogram wet weight) by applying AQ. One should realize, however, that AQ requires more time than does relative quantification, and one can benefit from AQ only if all additional referencing steps are executed properly; otherwise, unwanted additional errors may be introduced.
|
THEORETIC BACKGROUND |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
The requirements for signal reception are generally similar to those for transmission. The penetration of the RF pulses of the transmit coil and the signal-to-noise ratio (SNR) from the receive coil are strongly dependent on the coil loading, which is determined by the electric conductivity of the coil and the volume of the object near the coil. Variations in the size and tissue composition of the head positioned in the RF coil affect the amount of RF energy that gets into and the amount of signal detected from the volume of interest (VOI). This is one of the reasons why the electric characteristics of the RF coil are optimized before measurement begins, to achieve the most efficient performance.
Characteristics of MR Spectra
A water-suppressed brain 1H MR spectrum typically displays a number of signals that correspond to several brain metabolites. These signals are characterized by one or more peaks with a certain resonance frequency, line width (full width at half maximum of the peak's height), line shape (eg, lorentzian or Gaussian), phase, and area. The peaks are separated owing to differences in resonance frequency, which are caused by the difference in the chemical environment of the different nuclei. The molecular structure of a particular metabolite is reflected by a typical peak pattern. The area of a peak is directly proportional to the number of nuclei that contribute to it and to the concentration of the metabolite to which the nuclei belong. However, the peak areas are also influenced by other factors, including T1 and T2 relaxation times.
Quantification of 1H MR Spectra
Generally, the quantification procedure is divided into two steps. In the first step, accurate peak areas for the relevant metabolites are determined. In the second step, careful calibration is used to convert peak areas to metabolite concentrations to which the metabolite signals are referenced.
Definitions of Concentration
Different quantification strategies lead to different concentration standards. Two definitions of concentration are commonly used: (a) molarity, which is the number of moles of metabolite per liter of tissue water, and (b) molality, which is the number of moles of metabolite per kilogram of tissue water. In principle, molal concentrations can be converted into molar concentrations (32) by using a conversion that requires knowledge of the specific brain tissue density.
Comparison of in Vivo 1H MR Spectroscopy to Other Quantification Methods
Most alternative quantification methods can only be performed in vitro, after brain autopsy. These methods include high-pressure liquid chromatography and high-field-strength (>7-T) 1H MR spectroscopy. Whereas in vivo 1H MR spectroscopy can only determine metabolites with a concentration on the order of millimoles, both high-pressure liquid chromatography and high-field-strength 1H MR spectroscopy can demonstrate concentrations in the micromolar range. However, several metabolites, including N-acetylaspartate (NAA) and phosphocreatine, are generally detected at lower concentrations by using in vitro methods (up to 10% reduction relative to that detected in vivo in healthy subjects) (33,34) owing to the rapid degradation of these metabolites in brain tissue samples immediately after autopsy (35,36). The assessment of metabolites that are not subject to degradation yields similar concentrations in both in vivo and in vitro methods (37). Typical metabolite concentrations in parietal lobe white matter obtained with in vivo 1H MR spectroscopy are 9.6 mmol per kilogram (mmol/kg) wet weight ± 3.0 (standard deviation) for NAA, 7.0 mmol/kg wet weight ± 2.0 for total creatine, 1.8 mmol/kg wet weight ± 0.5 for choline, and 5.2 mmol/kg wet weight ± 3.3 for myo-inositol (38).
|
QUANTIFICATION STRATEGIES |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
|
|
AQ Methods
External reference method.—A vial with a known reference solution and relaxation properties is positioned near or inside the RF coil (Fig 1, B) (39). Immediately after the acquisition of the in vivo spectrum, a reference spectrum from the calibration sample is obtained while the patient's head is still inside the coil. The reference spectrum can also be obtained simultaneously with the acquisition of patient data (40). An option to account for possible B1 inhomogeneity is to measure (41) or to simulate (42) the B1 distribution of the RF coil used. It is important to realize that the external vial might introduce substantial distortions of the constant magnetic induction field (B0) homogeneity, which will complicate shimming (adjustment of the homogeneity of the local magnetic field) and water suppression.
Replace-and-match method.—The basic principle of the replace-and-match strategy is to replace the human subject with a phantom that simulates (human) tissue as closely as possible and to match the coil load to the load obtained previously in vivo (Fig 1, C) (43,44). Then, a calibration measurement is performed, which is identical with respect to all imager settings (eg, pulse sequence and size and position of the VOI). The phantom usually contains a solution that mimics the electric conductivity of human tissue. Fine adjustment of the coil load can be made by placing the phantom further inside or outside the coil or by inserting a small bottle containing saline. It is important that the load be accurately matched, because it is usually not possible to eliminate the effect of a difference in load by using a post hoc correction factor (44). Unfortunately, the matching of the load by maneuvering the phantom is not always straightforward and can be time consuming. Furthermore, the B1 distribution through a phantom is often considerably different from that in the human head. Therefore, simple matching of the conductivity and overall load might be insufficient, because the human head has structures of varying resistance that alter electric current paths.
Water signal reference method.—Several research groups have used tissue water as an internal standard (32,45–50). The in vivo measurement is performed first. Then, from the same voxel, the signal from unsuppressed tissue water is recorded, which serves as an endogenous concentration reference (Fig 1, A). It is best to record the water-suppressed spectra under conditions that are identical to those for the metabolite spectra. This can be achieved by detuning or switching off the water-suppression RF pulses. The use of pure cerebrospinal fluid (recorded from a voxel that contains only cerebrospinal fluid) as a water reference to obtain molar concentrations is inadvisable, since it is usually impossible to position a voxel of sufficiently large size in an area that contains only cerebrospinal fluid (eg, ventricles).
Alternatively, the water content can be derived from proton-density–weighted images (51,52). The obtained metabolite concentrations, in molal units, can be converted into molar concentrations by using the density of brain tissue (32). Because the determination of brain density and brain water content can be quite elaborate, one usually applies a value from the literature. However, one has to be cautious since several pathologic conditions are known to affect the water content. For instance, it has been determined for conditions such as multiple sclerosis (53,54), brain tumors (55), and hydrocephalus (55) that the brain water content can be increased up to 12%, 50%, and 120%, respectively, over the water content in healthy subjects.
Principle of reciprocity.—From the principle of reciprocity (55,56), it can be deduced that (if the coil is a combined transmit-receive coil) the external voltage needed to produce a certain B1 at a given location is inversely related to the voltage induced in the same RF coil by a predefined B1. The calibration strategy (also referred to as the phantom-replacement strategy) that is based on this notion involves the determination of the local B1 at the VOI, which is accomplished by measuring the voltage amplitude required to obtain the maximum signal response (Vmax) (Fig 1, A). Absolute standard units can then be obtained by dividing the recorded in vivo MR signals by Vmax and by performing additional calibration measurements with a solution of known concentration (57–61). The principle of reciprocity applies only to transmit-receive coils. For MR systems with a receive-only coil, an alternative approach based on the same principle has been designed in which the transmit-receive capabilities of the body coil are used (62).
Institutional units.—In this quantification approach, one of the previously mentioned absolute reference methods is used, but the final calibration to standard units is not performed. Therefore, this method shares the robustness of the applied AQ strategy but is only useful for diagnostic comparison within one institution.
Important Considerations for AQ
Echo time and repetition time.—When the repetition time of a pulse sequence is relatively short (generally shorter than five times the longitudinal relaxation time, T1), the magnetization cannot totally recover before the next excitation, which leads to a reduction of the signal (ie, signal saturation).
At 1.5 T, typical T1 relaxation times (± standard deviation) for several metabolites are 1430 msec ± 165 for NAA, 1330 msec ± 199 for choline, 1460 msec ± 270 for total creatine, and 1140 msec ± 308 for myo-inositol (32,63,64). Given these values, a repetition time of 2000 msec will lead to a decrease of the NAA signal of approximately 25%, whereas a repetition time of 7000 msec will cause a reduction of only 0.7%. Because the T1 values for most metabolites are not the same, the acquired signal intensity of each resonance should be corrected separately for partial saturation. Therefore, the use of a long repetition time is advisable since it reduces systematic errors caused by T1 signal saturation, even though this increases the duration of the MR acquisition.
The use of a short echo time (eg, <20 msec) minimizes signal losses caused by T2 relaxation. At 1.5 T, typical T2 relaxation times for several metabolites are 422 msec ± 48 for NAA, 356 msec ± 35 for choline, 214 msec ± 23 for total creatine, and 200 msec ± 20 for myo-inositol (32,51,64). The reduction of the NAA signal will be approximately 21% if an echo time of 100 msec is used, while the signal loss will only be approximately 5% if an echo time of 20 msec is used (the effect of varying echo times is indicated in Table 2). In contrast, macromolecules (compounds with a high effective molecular weight) have a very short T2 relaxation time (<50 msec at 2.1 T [65]), so the use of even the shortest echo time leads to a large reduction in signal. For instance, an echo time of 20 msec will lead to a reduction in the macromolecule signal of approximately 30%. On the other hand, spectra acquired with a long echo time (>150 msec) benefit from a less complicated appearance (mainly resonances from uncoupled spins remain visible, such as singlets from NAA, total creatine, and choline), improved water suppression, and a flatter baseline (due to signal reduction of components with short T2, such as water and macromolecules).
|
Receiver gain instability.—In any quantification strategy that requires an additional calibration measurement of a phantom, this measurement should be recorded with exactly the same amplifier gain settings as were used for the in vivo measurement, to prevent systematic errors. Therefore, a calibration experiment is preferably performed immediately after the in vivo experiment. However, if time restrictions prevent this measurement, one has to monitor the receiver stability on a regular basis, since it is not easily corrected for afterwards. An assessment of receiver gain stability can be performed by measuring the background noise level in a region of the spectrum "downfield" from the water resonance, where no signals are expected (61). The background noise level should exhibit minimal variations if the spectrometer has good long-term stability (56).
Compartmentation.—Because in vivo MR is performed on tissue at a macroscopic level, various brain compartments (gray matter, white matter, blood, cerebrospinal fluid, or lesions) with different metabolite concentrations might contribute to the metabolite signal measured. It is hard to determine the "true" concentration level of a metabolite in a specific homogeneous tissue. For example, metabolite concentrations that are uncorrected for the contribution of cerebrospinal fluid are generally underestimated, since the concentration of 1H MR spectroscopically detectable metabolites in cerebrospinal fluid is very low (70). Therefore, several tissue segmentation approaches have been proposed, all of which rely on differences in relaxation properties as determined with a separate MR imaging measurement. These methods include the use of calculated T1- or T2-weighted images (17) or spectra in which the T2 decay of water is measured as a function of echo time (50,71). However, since pathologic conditions may affect relaxation times, one should always carefully interpret results from any segmentation procedure.
MR visibility.—Not all metabolite molecules contribute equally to the MR signal from a certain VOI. For example, creatine has an invisible metabolite pool (2.5%) that is bound to macromolecular structures and, therefore, has low mobility; this results in very short T2 relaxation times and thus broad resonances, which can be unobservable on conventional MR spectra (72). A larger underestimation (up to 10% for creatine [73] and up to 30% for lactate [74]) of the true concentration can be caused by magnetization transfer effects if a water-suppression technique such as presaturation is used. However, most metabolites are not susceptible to magnetization transfer effects and show a signal change only slightly above the limits of experimental error (75). Furthermore, in most measurement methods magnetization transfer effects can generally be avoided.
|
DATA ANALYSIS |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
|
Peak Fitting
In this approach, all important peaks are initially selected and coarse estimations of the resonance frequency, line width, and peak intensity are performed, either by the operator or by an algorithm. Subsequently, a fit is performed by using a least-squares optimization algorithm, which iteratively fits all peaks to a line-shape model function, so that the fitted spectrum resembles the experimental spectrum as closely as possible (80). In general, this method proves to be fairly robust with respect to spectral overlap. However, if the actual line shapes deviate substantially from Gaussian or lorentzian model functions, the algorithm will not be able to fit the peaks accurately. To prevent this problem, other line shapes have occasionally been used, of which the Voigt profile is the most common (81). However, the use of a Voigt profile introduces more degrees of freedom in the fitting procedure, which can lead to ambiguous results.
Prior Knowledge
This method allows the incorporation of prior knowledge about the metabolites that contribute to the 1H MR spectroscopic signal (82). All known signal parameters, such as relative frequencies, amplitude ratios, scalar coupling, and phases of resonances that are characteristics of a certain metabolite, can be implemented as constraints in the fitting routine. Prior knowledge is the only way of enhancing, by reducing the degrees of freedom, the accuracy of fitted model parameters for a given data set. The calculation time is also reduced. An example of a method that enables incorporation of prior knowledge is the AMARES method (Department of Electrical Engineering, Katholieke Univesiteit Leuven, Leuven, Belgium) (83).
Advanced prior knowledge obtained through metabolite basis set.—A decade ago, algorithms were proposed (82,84) that implemented a strategy where the maximum of prior knowledge is used. Those methods were based on the assumption that there are a limited number of 1H MR spectroscopically detectable metabolites present in the human brain that had already been identified and analyzed in earlier studies. The approach aimed to individually determine the exact response of all metabolites possibly present to a specific pulse sequence. One has to take into account the type of imager, B0, pulse sequence, echo time, and repetition time, as well as the pH, ion composition, and temperature of the solution in which the individual metabolites are dissolved. In this way, the prior knowledge of each metabolite—including chemical shifts, signal intensity (influenced by relaxation effects), amplitude ratios, splitting patterns, and J evolution—matches the in vivo conditions. The in vivo MR spectra are then analyzed as a linear combination of the separately recorded in vitro spectra of the individual metabolites.
There are two commonly used approaches for obtaining a basis set: simulation and in vitro measurement. In the simulation approach (85,86), the response is numerically simulated on the basis of molecular and quantum mechanical characteristics of each metabolite (87,88). The in vitro approach (89) requires that for each metabolite, a spectrum is acquired with exactly the same conditions (eg, pulse sequence and timing parameters) as those used during the in vivo measurements. The advantage of the simulation approach is that, as long as the exact molecular structure is known, each metabolite can be included, whereas the in vitro approach requires carefully prepared metabolite phantoms. The advantage of the in vitro approach is that, as long as the phantoms are adequately prepared, the response of the metabolites in the in vitro measurement will be identical to the response of the metabolites in the in vivo 1H MR spectroscopy examination.
The high information content of 1H MR spectra leads to an advantage regarding the fitting accuracy, because overlapping resonances at one chemical shift position can be directly related to other nonoverlapping resonances from the same metabolite. Currently used computer software includes the metabolite basis set fitting program LCModel (Stephen Provender, PhD, http://s-provender.com/pages/lcmodel.shtml) (84), possibly incorporating the molecular simulation library GAMMA (Department of Radiology, University of California, San Francisco) (87); the quantification package jMRUI (The MRUI Project, http://sermn02.uab.es/mrui/) (90), which includes the quantum mechanical simulation algorithm NMR-SCOPE (Laboratoire de RMN, Universite Lyon I-CPE, Villeurbanne, France) (88); and the metabolite basis set fitting routine QUEST (Laboratoire de RMN, Universite Lyon I-CPE) (85) and the algorithm TDFDFit (Department of MR Spectroscopy and Methodology, University and Inselspital, Berne, Switzerland) (91). A detailed description (92) and a critical assessment (7) of both AMARES and LCModel can be found elsewhere. Examples of LCModel output and TDFDfit output are given in Figures 2 and 3, respectively.
|
|
Important Considerations for Data Analysis
Quantification accuracy.—It is important to study and report the error estimates of the quantification method. Most of the fitting routines present the so-called Cramer-Rao minimum variance bounds (CRMVBs), which reflect the theoretic standard of precision for the model parameter estimates obtained from the data (94). The parameter estimation must not contain systematic errors (eg, incorrect prior knowledge), which may lead to underestimation errors. It is important to realize that the CRMVBs provide a measure of quality of the spectral fit and do not necessarily reflect the quality of the original data. Furthermore, SNR degradation and increases in line width, which may lead to systematic errors, are not necessarily reflected in CRMVB estimates (95).
Macromolecules
Macromolecules are compounds with a relatively high molecular weight and include proteins and polypeptides. Owing to their lower mobility, macromolecules generally have a short T2 and are thus characterized by broad spectral lines on 1H MR spectra. Macromolecules are the main contributors to the fluctuating baseline that generally hinders accurate quantification of small metabolites. Some macromolecular resonances even overlap completely with sharp metabolite resonances and, therefore, cannot even be accounted for with conventional 1H MR spectroscopy (65,96). Several methods can be used to separate the macromolecular baseline from the wanted metabolite signal.
The best way to eliminate macromolecular contamination is to use acquisition schemes that exploit the different relaxation properties of macromolecules and small metabolites. Removal based on T1 relaxation can be achieved by so-called inversion-recovery (65) or saturation-recovery (97) sequences (Fig 4), whereas removal based on T2 relaxation can be obtained by increasing the echo time (32).
|
Recently, it has become popular to include the macromolecules baseline in the fitting routine (20,89,99). With the LCModel approach, it is possible to extend the metabolite basis set with a macromolecules spectrum. Since the macromolecular contribution to in vivo spectra cannot be simulated with simple model solutions, one has to determine the macromolecules baseline experimentally. Both inversion recovery (89) and saturation recovery (100) have been used to obtain macromolecules baseline spectra, which were averaged over several subjects and then parameterized to be included in the basis set. The incorporation of macromolecules into the basis set generally improves quantification accuracy and precision (101).
Macromolecules themselves have been the subject of recent research (65,102). It has been reported that several conditions, such as stroke (103), brain tumors (101), and multiple sclerosis (99), show an altered macromolecular profile. For example, stroke (103) and multiple sclerosis (99), respectively, can cause the signal occurring from the macromolecules resonance at 1.3 ppm to increase by 86% and 60%. Therefore, one should always carefully consider the incorporation of macromolecules spectra in the basis set when such a condition is of concern.
Motion
Physiologic motion can have a large effect on 1H MR spectroscopic quantification (104). For example, repeated small body motions or pulsatile cerebral motion (eg, due to cardiac activity), can substantially affect the characteristics of the MR signal, even with motion-compensated 1H MR spectroscopy sequences. These effects may include an increase in line width, reduced peak intensities, a diminished quality of water suppression, or a doubling of all peaks. Moreover, if quantification is based on the summed signal of reference spectra of the unsuppressed water peak, motion-induced phase effects can cause a substantial overestimation of metabolite concentrations, a problem that can be circumvented by storing and phasing all spectra individually (104). For chemical shift imaging studies that include subcutaneous lipid signals, which have an enhanced sensitivity to subject motion, an in-plane motion correction can be applied (105). The reduction of motion effects in 1H MR spectroscopy has been achieved in several studies (eg, by using cardiac triggering [104] or postprocessing [106]).
VOI Shape and Location
In the process of spectral quantification, one should always keep in mind that the true spatial voxel profile might deviate from a perfect rectangular profile. The voxel dimensions are, in principle, identical for all resonances; however, the voxel position is dependent on the chemical shift of individual resonances. When a section-selective gradient is applied to a volume containing metabolites with different chemical shifts, there will be a displacement of the sensitive volume for each resonance of the metabolite. The displacement is regulated by the bandwidth of the RF pulse for voxel selection (38). For example, the methyl protons of NAA and the methylene protons of total creatine are separated by 1.90 ppm, or 122 Hz, at 1.5 T. Use of an RF bandwidth of 2000 Hz and a voxel size of 2 x 2 x 2 cm results in a 1.22-mm displacement of the 2 voxels, which results in an 82.8% overlap of the 2 voxels.
Chemical shift artifacts can be reduced by increasing the bandwidth of the section-selective RF pulses and, to keep the same voxel dimensions, by increasing the strength of section-selective gradients. Obviously, safety margins should not be exceeded. It should be noted that chemical shift imaging does not suffer from the chemical shift artifact for the spatial phase-encoding gradients, which are applied in the absence of RF pulses. However, if the chemical shift imaging study involves preselection of a VOI (eg, to exclude skull lipids) then the position of that VOI is again subject to the chemical shift artifact, while the chemical shift imaging voxels within the VOI are not shifted with respect to each other. Chemical shift imaging data are also influenced by the point-spread function, which will be treated in the section entitled Sequence-Specific Issues.
Line Shape and Eddy Currents
Very often, the line shape of signals obtained from metabolites deviates substantially from the ideal lorentzian line shape. If left uncorrected, these line shape deviations severely complicate peak fitting. Generally, line shape deviations are caused by variations in B0, such as magnetic field inhomogeneities and eddy currents. Rapid switching of magnetic field gradients (which occurs in every localization pulse sequence) induces time-varying eddy currents in the magnetic cryostat. As a result, the magnetic field is transiently distorted. Several methods have been developed to correct irregular line shapes, including the QUALITY deconvolution (107) and eddy current correction procedure (108). These correction methods use a single reference peak that was affected by the same distortions. The unsuppressed water signal, acquired by using the same spoiler gradients as the metabolite spectrum, is usually used for that purpose.
Water Suppression
As indicated earlier, water-suppression pulses might saturate parts of the metabolite spectrum and can thus influence the measured concentration of several metabolites (a reduction of up to 21% has been reported [109]). Numeric calculations can be employed for accurate corrections for these effects (109). In addition, the residual water signal that is still present in water-suppressed spectra can substantially degrade the performance of metabolite fitting routines. Therefore, it is advisable to eliminate the unwanted water signal by using a high-pass filter (110) or an algorithm such as the Hankel-Lanczos singular-values decomposition filter (111) prior to data analysis (Fig 5).
|
|
SEQUENCE-SPECIFIC ISSUES |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
-aminobutyric acid, glutamine, lactate, and glutathione. It is important that the pulse sequence be carefully optimized and that the specificity of the editing procedure be critically assessed. If unwanted signals remain in the final spectrum, simple peak fitting will not suffice, although one could resort to the advanced prior knowledge approach (115).
Chemical Shift Imaging
Quantification of chemical shift imaging data generally requires extra considerations, because chemical shift imaging data can be different from single-voxel data in some respects. For example, chemical shift imaging data usually have a lower SNR, since these imaging data are subject to field inhomogeneities (due to the large VOI). In addition, chemical shift imaging data are also subject to artifacts such as imperfect water-suppression and lipid-contamination artifacts. On the other hand, the spatial information of chemical shift imaging offers several opportunities to improve postprocessing and analysis. Knowledge of the spatial dependence of parameters may yield extra information with regard to prior knowledge or constraints. For instance, the notion that voxel-to-voxel variations in metabolite concentrations can be expected to change smoothly can be used to set constraints on the limits of these variations (116).
Not all AQ strategies are necessarily compatible with chemical shift imaging data. For example, it is often too time consuming to obtain fully relaxed water reference spectra with the same resolution (117). A commonly used referencing scheme for chemical shift imaging is the time-efficient reciprocity principle (21,61). An absolute reference can also be obtained by means of proton-density–weighted MR imaging (40).
Point-spread function.—Quantification of chemical shift imaging data is complicated by the point-spread function, which describes the spread of signal from one voxel to surrounding voxels. The point-spread function originates from Fourier transformation of a signal sampled in the time domain at a limited number of discrete points. Although the point-spread function influences all Fourier-transformed data, in chemical shift imaging the effect may become prominent since chemical shift imaging data are generally acquired with a relatively low spatial resolution, corresponding to a small number of phase-encoding increments.
The point-spread function, and therefore the spatial resolution, can be improved by applying apodization functions in the spatial frequency or k-space domain (118). One approach is to apply different k-space sampling schemes to improve the point-spread function and the SNR (119,120). Another approach is k-space weighting during signal acquisition, which also yields a better SNR (121). In chemical shift imaging of the brain, the point-spread function effect will typically lead to contamination of 1H MR spectra of voxels within the brain by intense extracranial lipid signals. k-Space weighting or application of apodization functions to improve the point-spread function is usually insufficient in this case; therefore, other methods, such as outer volume suppression, are used to prevent lipid signals from contaminating brain spectra.
Other methods reduce lipid contamination during postacquisition processing (122). In an AQ strategy, coregistered MR images are generally used for tissue segmentation (T1- or T2-weighted MR imaging [17]) or for water referencing (proton-density–weighted MR imaging [40,52]). It is important to note that apodization filters, used to improve the point-spread function of the metabolite spectra, should also be applied to the MR imaging data that are used for the segmentation procedure.
Fully automated spectral analysis.—Analysis of spectra from large chemical shift imaging data sets is time consuming, and because usually only spectra from a few selected voxels are analyzed, the complete data set is often underutilized. Therefore, automated methods have been developed (87,116,117) that enable analysis of large numbers of spectra and that are relatively insensitive to the low SNR and spectral distortions commonly associated with in vivo chemical shift imaging. Furthermore, the introduction of a standardized processing and analysis protocol makes comparisons between different studies possible.
|
QUALITY ASSESSMENT |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
SNR Considerations
The SNR can be used as a measure in the rejection of spectra (eg, the SNR should generally be at least 4). The SNR is usually defined in the frequency domain as the ratio of the highest metabolite peak intensity to the standard deviation of the noise amplitude in a metabolite-free part of the spectrum (118). The SNR can be improved by increasing the size of the VOI and the number of signal averages acquired. However, an increase in the size of the VOI generally degrades the line shape, increases the line width, and may decrease the sensitivity to detect abnormalities, whereas an increase in the number of signals acquired adds to the examination time.
Line Width
The line width is commonly defined as the full width of the peak at half its maximum height, or full width at half maximum, in the frequency domain (118). A small line width generally implies a high spectral resolution, which will improve the quality of the fitting routines. Better shimming, the use of a smaller VOI, and the choice of a VOI at a sufficient distance from tissue interfaces can improve the resolution. It has been recommended that the full width at half maximum value should always be treated as a covariate in the statistical analysis, to help correct for single-subject intersession variations (27).
Criteria for Rejection of Data
In a recent insightful review, Kreis (125) proposed the following acceptance criteria for spectral data and their fitting results: (a) The full width at half maximum of the metabolites should be less than 0.1 ppm; (b) the Cramer-Rao minimum variance bound should be smaller than 50%; (c) the fitting residual should not contain unexplained features; and (d) the spectra should not contain artifacts (eg, doubled peaks or asymmetric line shapes).
|
CONCLUSIONS AND FUTURE PERSPECTIVES |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
|
ESSENTIALS |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
|
FOOTNOTES |
|---|
Abbreviations: AQ = absolute quantification • NAA = N-acetylaspartate • RF = radiofrequency • SNR = signal-to-noise ratio • VOI = volume of interest
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONADVANTAGES AND DISADVANTAGES OF...THEORETIC BACKGROUNDQUANTIFICATION STRATEGIESDATA ANALYSISSEQUENCE-SPECIFIC ISSUESQUALITY ASSESSMENTCONCLUSIONS AND FUTURE...ESSENTIALSReferences |
|---|
