Technical aspects of amyloid imaging for Alzheimer's disease

[11C]Pittsburgh Compound B positron emission tomography has now been extensively used to evaluate the amyloid load in different types of dementia and has become a powerful research tool in the field of neurodegenerative diseases. In the present short review we discuss the properties of amyloid imaging agent [11C]Pittsburgh Compound B, the different modalities of molecular imaging, image processing and data analysis, and newer amyloid imaging agents.


Introduction
Alzheimer's disease (AD) is the leading neuro degenerative disorder and accounts for approximately two-thirds of dementia. AD aff ects around 10% of people above the age of 75, and in the United States approximately 4 million people suff er from AD-related dementia with annual associated costs estimated to be approximately $100 billion.
Th e diagnosis of AD currently depends on patients having impairments in memory function and at least one other cognitive domain, to the extent that it impairs daily function. AD is a clinical diagnosis mainly based on NINCDS-ADRDA criteria [1] and DSM-IV criteria [2]. Th ese criteria were revised recently by the international working group for new research criteria for the diagnosis of AD [3]. Th e cardinal features are late-onset impairment of short-term recall [4,5] associated with deterioration of language and visuo-spatial functions [6,7] in the absence of disturbance of con scious ness and systemic disorders. A postmortem diag nosis of AD is based on the presence of extracellular senile plaques positive for β-amyloid protein (Aβ), dystrophic neurites, and intracellular neurofi brillary tangles in the brain. Neurofi brillary tangles contain a hyperphosphorylated form of the microtubule-associated protein, tau, and also ubiquitin [8]. Amyloid plaques are composed of 40 to 42 amino acid Aβ-folded peptides. Two types of amyloid plaques are present in neuro degenerative diseases: one with a central core and radiating fi brils, the fi brillar amyloid; and second, the diff use (amorphous) amyloid. A small degree of AD-like pathology can be found in cognitively normal individuals over the age of 75 but large deposits of amyloid suggest AD.
Until a few years ago, magnetic resonance imaging (MRI) and cerebral blood fl ow studies were the only methods by which we could gain information about the changes in the living human brain. Structural MRI and blood fl ow studies, however, depend on gross changes in brain structure and function that suggest gross atrophy or a functional alteration that has already occurred. Th ese markers are thus secondary phenomena and are therefore not the primary targets for following patients over a period of time or for diagnosing very early and subtle changes. Clinico-pathological studies suggest that neuronal loss has already occurred by the time gross atrophy is detected by MRI. Positron emission tomography (PET) with [ 18 F]fl uorodeoxyglucose adds to the diag nostic and prognostic accuracy in the clinical evalu ation of AD [9], but the technique still detects an indirect measure of disease presence or progression.
Modifi cation of the polar amyloid binding histological dye, thiofl avin T, led to the fi nding that neutral benzothiazoles bind to amyloid with high affi nity and additionally cross the blood-brain barrier [10]. Th e benzothiazole amyloid binding agent 2-(4'-methyl-aminophenyl)-benzothiazole and related compounds bind to amyloid with low nanomolar affi nity, enter the brain in amounts suffi cient for imaging with PET and clear rapidly from normal brain tissue [11,12]. At the low nanomolar concentrations typically administered during PET studies, 2-(4'-methyl-amino-phenyl)-benzothiazole binds to extracellular amyloid plaques in postmortem brain slices but not to intracellular neurofi brillary tangles. In vitro studies suggest that, while 2-(4'-methyl-amino-phenyl)-benzothiazole binds to fi brillar Aβ deposits found in the cortex Abstract [ 11 C]Pittsburgh Compound B positron emission tomography has now been extensively used to evaluate the amyloid load in diff erent types of dementia and has become a powerful research tool in the fi eld of neurodegenerative diseases. In the present short review we discuss the properties of amyloid imaging agent [ 11 C]Pittsburgh Compound B, the diff erent modalities of molecular imaging, image processing and data analysis, and newer amyloid imaging agents.  A hydroxylated derivative of benzothiazole, N-methyl-[11-C]2-(4'-methylaminophenyl)-6-hydroxybenzothiazole -widely known as [ 11 C]Pittsburgh Compound B ([ 11 C]PIB) -has been synthesised as a PET tracer and studied in humans. An initial study showed that [ 11 C]PIB shows a signifi cant uptake in AD subjects compared with the control subjects [13]. Following this initial study, several other studies have shown similar fi ndings in AD and in other dementia [14][15][16]. Postmortem studies have shown a direct correlation between Aβ plaque and in vivo [ 11 C]PIB retention measured by PET imaging. Studies have shown that the fl uorescent PIB ana logue 6-CN-PIB labelled Aβ-containing structures, includ ing compact/ cored, diff use, neuritic and non-neuritic Aβ plaques. 6-CN-PIB also labelled vascular amyloid, but no signal was detected in neurophil threads or dystrophic neurites and a signal was only detected sporadically on an extracellular ghost tangle. A patient who underwent postmortem neuropathological examinations 10 months after [ 11 C]PIB PET imaging demonstrated that in vivo retention corre lated directly with postmortem quantifi cation of PIB and Aβ plaque load but not with neurofi brillary tangle or other neurofi brillary pathology [17].
In the present article, a review of the technical aspects of amyloid imaging for AD will be presented. We therefore fi rst introduce the general procedure for in vivo molecular imaging in man using PET. We then go on to describe image processing and data analysis.

Molecular imaging in man
Th e molecular imaging process is shown in Figure 1. In the fi rst step (top right of the fi gure), radioisotopes are produced either in dedicated cyclotrons -for example, the positron emitters 11 C or 18 F -or in special radionuclide generators -for example, the single-photon emitter 99m Tc. After the production of the radioisotope, the radio labelled compounds for the imaging study are produced in a radiochemistry system. Various routes for the radiochemistry usually exist; for example, for 18 Flabelled tracers by nucleophilic or electrophilic reaction. Quality control is performed to check and confi rm that the radiopharmaceutical product is within predetermined specifi cations; for example, for specifi c activity, radiochemical purity or sterility.
Th e basis of tracer imaging is the detection by external devices of the radiation emitted from the radiolabel attached to the tracer injected into humans. In most cases, the nuclear disintegration of the radioisotope is detected via registering photons that are either uncorrelated (hence single-photon imaging) or that are paired as the result of positron annihilation. Th is imaging technique is then known as coincidence imaging or positron imaging, and is schematically shown in the scan box of Figure 1. A pair of photons resulting from the annihilation of a positron generated by the conversion of a proton in the nucleus is recorded as a line of response by the PET scanner. Tomographic image reconstruction techniques are subsequently employed to reconstruct a dynamic sequence of emission images from the brain scan [18]. Today, PET scanners operate in three-dimensional mode, covering the entire brain in a single scan with an axial fi eld of view of at least 15 cm. Modern PET scanners are usually combined PET/computer tomography systems as they include an integrated X-ray computer tomograph for the acquisition of photon attenu ation images required for the correction of the PET images for the eff ects of photon attenuation and scatter. Combined PET and MRI systems are just emerging that combine the excellent contrast between grey and white matter provided by the magnetic resonance acquisition with the molecular images of the PET study.
Th e most widely used method for the reconstruction of quantitative brain images using PET is fi ltered back projec tion, an analytical image reconstruction method. Filtered back projection is computationally fast and has linear proper ties, which means the precision of the reconstructed images is independent of the location within the image and of the intensity of the object. Iterative image reconstruction, on the other hand, is computationally expensive and often only slowly converging and in an object-dependent manner. Iteratively reconstructed images are visually more appealing, because they do not exhibit the streak artefacts typical for fi ltered back projection reconstructed images. As the iterative image reconstruction allows one to more accurately model the entire imaging process, resolution eff ects can be included in the system descriptionimages with higher spatial resolution than those from fi ltered back projection can therefore be obtained.
Th e absolute quantifi cation of the radiotracer kinetics in the tomographic images normally requires an input function. Th e input function is the time course of the radiotracer in the supply stream that drives the tissue response. Th e time course of the concentration of the radiolabelled compound in arterial plasma therefore has to be measured. In contrast to the acquisition of the images, which is performed by a single instrument (a SPECT or PET camera), the measurement of the plasma input function requires the combination of several laboratory devices. Online blood detector systems are used to provide whole blood activity measurements of continuously withdrawn blood with excellent temporal resolution but with limited sensitivity due to their relatively small counting volume. Well counters or automated gamma counters are used to measure with very high sensitivity the activity of discrete blood samples or, after centrifugation, of plasma samples. For radiotracers that undergo metabolism in the body, quantitative assays of the plasma samples to determine the fraction of radioactivity that is due to unmetabolised parent compound is required to obtain the input function. Th e laboratory setup for these analyses often involves solid-phase extraction followed by HPLC or thin-layer chromatography. Figure 2 shows an example chromatogram obtained from an HPLC analysis run of a scan with [ 11 C]PIB.

Image processing and data analysis
After completion of the PET scan and the reconstruction of the dynamic images from the emission scan, an integral or sum image is generated in order to perform a co-registration with a structural volumetric magnetic resonance image as shown in Figure 3. Th is then allows the accurate defi nition of volumes of interest (VOIs) on the brain imaging data. Automated methods using atlases defi ned on brain templates are used as well as VOIs being manually outlined, particularly in cases with high levels of brain atrophy.
Using the set of VOIs defi ned, the dynamic PET images are then sampled and tissue time-activity curves generated. As is illustrated in Figure 4, these tissue timeactivity curves represent the measured activity concentration averaged across the VOI -which implies that, in the ensemble of volume elements (voxels) comprising a VOI, all voxels share the same parameters of the underlying physiological and biochemical processes, and only diff er as they are diff erent realisations of the same random process. Th is means that special attention has to be paid to tissue heterogeneity when VOIs are defi ned, and often magnetic resonance images segmented in diff erent tissue classes (grey matter, white matter and cerebrospinal fl uid) are employed for the defi nition of homogeneous VOIs. As an example, the diff erence in the [ 11 C]PIB signal between cerebellar grey matter and cerebellar white matter is highlighted in Figure 3.
Th e purpose of tracer kinetic analysis (centre top box in Figure 1) is to disentangle the diff erent processes that jointly result in the tissue response curves obtained [19]. For amyloid imaging studies, the contributions to consider are as follows. First, tracer delivery and washoutas the amyloid imaging markers currently used (for ) are thought to cross the blood-brain barrier by passive diff usion, the delivery to and washout from brain tissue of these radiotracers is governed by cerebral blood fl ow. A second contribution is specifi c binding -the association, and dissociation for reversibly binding ligands, of the radioligand with the target (that is, Aβ). Another contribution is nonspecifi c binding -referring to any nonsaturable binding that occurs to other sites than Aβ (for example, to membranes or lipid fractions). A fourth contribution is radiolabelled metabolites -several of the [ 18 F]-labelled radiotracers used for amyloid imaging have been reported to form radiolabelled metabolites in vivo that are also able to cross the blood-brain barrier [20]; part of the radioactivity signal measured in brain tissue is therefore due to the contamination with radiolabelled metabolites. Finally, vascular activity -owing to the spatial resolution of the positron cameras of several millimetres, any VOI defi ned in the brain contains a few per cent blood volume; the spill in of activity from the vasculature therefore needs to be accounted for.
Using mathematical modelling and parameter estimation methods, system parameters such as rate constants, volumes of distribution or binding potentials [21] can be estimated from the dynamic imaging data. Th e outcome parameters chosen should, of all the contributions listed above, refl ect the specifi c binding to the maximum possible extent and should be insensitive to the other confounders. For example, it has been shown for [ 11 C]PIB that the accumulation rate did not correlate with cerebral blood fl ow [22]. Th e parameter estimates can either be obtained regionally for each VOI, or they can be calculated for each voxel individually and then again represented as an image that is often referred to as a parametric map [23] (Figure 5). Parametric maps can then be interrogated for parameter changes that do not correspond to the anatomically predefi ned VOIs.
A variety of modelling approaches exists and they have been applied to amyloid imaging studies with [ 11 C]PIB. Th ese approaches range from compartmental models [24], through graphical analyses such as Logan plots [25] or Patlak plots [26,27], to spectral analysis [28] and reference tissue models [29,30]. In reference tissue models, the tissue time-activity curve of a region without specifi c binding is used as a substitute for the plasma   input function. For studies of sporadic AD the cerebellar grey matter is widely used as a reference region because postmortem investi gations confi rmed negligible concentrations of Aβ in cerebellar grey matter in this disease. However, careful validation of the reference region is required for each disease population.
One of the commonly used methods of analysis is the target to cerebellar ratio, commonly referred to as the RATIO method. Diff erent groups have used diff erent time points to create RATIO images from 40 to 60 minutes, from 40 to 70 minutes and from 60 to 90 minutes. Th ese diff erent time points largely give comparable results, however -and at later time points, while the signal increases, the noise also increases. Again, in diff erent studies, diff erent RATIO values have used as cut-off points [31,32]. Some studies have used a strict cut-off value of two standard deviations above the control mean for individual regions, while other studies have used much more liberal cut-off values and a RATIO value of 1.4 or even 1.5. Even though diff erent scanners diff er   Figure 3: top images, control subject; bottom images, Alzheimer's disease (AD) patient. Images generated with spectral analysis; colour bars for each column indicate the image minimum and maximum values. The impulse response function (IRF) at 1 minute in the left column gives a refl ection of tracer delivery. The cerebral blood fl ow in most grey matter areas of the AD subject's brain is lower than that in the elderly control subject. The IRF 75-minute images in the centre give an indication of [ 11 C]PIB retention in tissue. In the control subject, [ 11 C]PIB is retained predominantly in the white matter; whereas in the AD patient, widespread [ 11 C]PIB retention in cortical grey matter occurred with association cortical areas, in particular, being targeted. Note that the cerebellar grey matter in both the control subject as well as the AD patient shows only very little [ 11 C]PIB retention. The total volume of distribution (VD) parametric images in the right column also refl ect [ 11 C]PIB uptake in tissue, but with a slightly lower contrast between grey matter and white matter than the IRF 75-minute parametric images. slightly, it is generally accepted that a RATIO value above 1.5 is clearly abnormal.
Less sophisticated but technically much simpler to perform than dynamic scans are static acquisitions. In this procedure, the time-course of activity is not measured but only an integral (sum) image of tissue activity of a certain period after tracer injection is acquired. By then it is assumed that the tracer has reached a state of pseudo-equilibrium so that the tissue activity-concentration ratio can be used as an apparent volume of distribution ratio [33] (Figure 6).

Statistical parametric mapping
Th e statistical parametric mapping (SPM) method of analysis is an approach used to localise signifi cant changes in brain physiology or pharmacology. Th is localisation is achieved by applying the general linear model to all independent voxels and creating an image of a statistic. SPM refers to the construction of spatially extended statistical processes to test hypotheses about regionally specifi c eff ects. SPMs are image processes with voxel values that are, under the null hypothesis, distributed according to a known proba bility density function (usually Gaussian). Th ese statis tical parametric maps are three-dimensional projec tions of statistical functions that are used to characterise signifi cant regional brain diff erences in imaging para meters. Th e theory of Gaussian fi elds is used to provide P values that are corrected for the brain volume analysed [34,35].
To implement voxel-based analysis of imaging data, the data from diff erent subjects must derive from homologous parts of the brain. Spatial transformations are therefore applied that move and warp the images such that they all conform (approximately) to some standard brain. Th is normalisation facilitates intersubject averaging. Th e normalising transformations can be computed on the basis of the PET data themselves on the basis of co-registered high-resolution anatomical MRI. Convolving the data with a smoothing kernel has several important objectives. First, it generally increases the signal relative to noise. Second, convolving with a Gaussian kernel conditions the data to conform more closely to a Gaussian fi eld model, even though this reduces spatial resolution.
Th e aim of the analysis is to identify areas of cerebral [ 11 C]PIB binding that were signifi cantly diff erent in patients compared with the controls. Signifi cant diff erences between patients and control subjects were estimated according to the general linear model at each and every voxel [36]. Comparison between patients and controls as groups was performed without using analysis of covariance. Linear contrasts were used to test the hypotheses for specifi c focal eff ects. Th e resulting set of voxel values for each contrast constitutes a statistical parametric map of the t statistic SPM[t].
Spatially normalised RATIO images could be interrogated using diff erent thresholds. Here we have used a threshold of P <0.00001 with an extent threshold of 200  Figure 7 shows [ 11 C]PIB uptake was high in AD patients com pared with the healthy control subjects in frontal, temporal, parietal and occipital cortices. SPM was not able to interrogate the parametric images at a lower threshold for signifi cance. [ 11 C]PIB also showed a signifi cant increase in mild cognitive impairment subjects, as 60% of these patients had signifi cantly increased levels compared with those of AD.

Newer amyloid tracers
A fl uorinated derivative of PIB, [ 18 F]fl utemetamol, is now being developed and studied [37]. Since the half-life of this compound is considerably higher than carbon-11 com pounds, a cyclotron is not necessary for routine studies as this could be manufactured at one site and could be distributed. [ 18 F]Flutemetamol is analysed using the target to cerebellar ratio similar to [ 11 C]PIB [38].
Other fl uorinated compounds widely tested for amyloid imaging include [ 18 F]AV-45 [39,40] and fl orbetaben [41]. A head-to-head comparison of [ 18 F]AV-45 against [ 11 C]PIB is now underway. Several other imaging agents are also in development.