Comparisonofmanualandautomated quantificationmethodsof123I-ADAM

205 © 2005

Comparison of manual and automated quantification methods of 123I-ADAM T. Kauppinen1,2, A. Koskela2, M. Diemling3, A. Keski-Rahkonen4,5, E. Sihvola4,5, A. Ahonen2 1 HUS Helsinki Medical Imaging Center, 2Division of Nuclear Medicine, 5Department of Psychiatry, Helsinki University Central Hospital, 4Department of Public Health, University of Helsinki, Finland 3 Hermes Medical Solutions, Stockholm, Sweden Keywords

Schlüsselwörter

Summary

Zusammenfassung

Brain quantification, single photon emission tomography, serotonin receptor, SERT 123I-ADAM

is a novel radioligand for imaging of the brain serotonin transporters (SERTs). Traditionally, the analysis of brain receptor studies has been based on observer-dependent manual region of interest definitions and visual interpretation. Our aim was to create a template for automated image registrations and volume of interest (VOI) quantifications and to show that an automated quantification method of 123I-ADAM is more repeatable than the manual method. Patients, methods: A template and a predefined VOI map was created from 123I-ADAM scans done for healthy volunteers (n = 15). Scans of another group of healthy persons (HS, n = 12) and patients with bulimia nervosa (BN, n = 10) were automatically fitted to the template and specific binding ratios (SBRs) were calculated by using the VOI map. Manual VOI definitions were done for the HS and BN groups by both one and two observers. The repeatability of the automated method was evaluated by using the BN group. Results: For the manual method, the interobserver coefficient of repeatability was 0.61 for the HS group and 1.00 for the BN group. The intraobserver coefficient of repeatability for the BN group was 0.70. For the automated method, the coefficient of repeatability was 0.13 for SBRs in midbrain. Conclusion: An automated quantification gives valuable information in addition to visual interpretation decreasing also the total image handling time and giving clear advantages for research work. An automated method for analysing 123I-ADAM binding to the brain SERT gives repeatable results for fitting the studies to the template and for calculating SBRs, and could therefore replace manual methods.

Nuklearmedizin 2005; 44: 205–12

Received: January 3, 2005; in revised form: March 21, 2005

Gehirnquantifizierung, Single-Photonenemissions-Computertomographie, Serotoninrezeptor, SERT 123I-ADAM

ist ein neuer Radioligand für die bildgebende Diagnostik der zerebralen Serotonintransporter (SERTs). Bisher beruhte die Analyse zerebraler Rezeptorstudien auf der Beobachter-abhängigen Definition der Bereiche besonderen Interesses und der visuellen Interpretation. Unser Ziel war, eine Schablone für automatische Bilddeckungen und für Quantifizierungen des interessierenden Volumens (VOI) zu erstellen und zu zeigen, dass eine automatische Methode zur Quantifizierung von 123I-ADAM reproduzierbarer ist als die manuelle Methode. Patienten, Methoden: Aus 123I-ADAM-Scans, die bei Gesunden durchgeführt wurden (n = 15), wurden eine Schablone und eine vordefinierte VOI-Karte erstellt. Scans einer anderen Gruppe Gesunder (GP, n = 12) und von Patienten mit Bulimia nervosa (BN, n = 10) wurden automatisch in die Schablone eingepasst, und es wurden spezielle Bindungsquotienten (specific binding ratios – SBRs) unter Verwendung der VOIKarte berechnet. Die manuellen VOI-Definitionen wurden für die GP- und BN-Gruppe sowohl von einem als auch von zwei Beobachtern vorgenommen. Die Reproduzierbarkeit der automatischen Methode wurde anhand der BN-Gruppe ausgewertet. Ergebnisse: Bei der manuellen Methode betrug der Interbeobachter-Koeffizient der Reproduzierbarkeit bei der GP-Gruppe 0,61 und bei der BN-Gruppe 1,00. Der Intrabeobachter-Koeffizient der Reproduzierbarkeit für die BN-Gruppe betrug 0,70. Bei der automatischen Methode betrug der Koeffizient der Reproduzierbarkeit für die SBRs im Mittelhirn 0,13. Schlussfolgerung: Eine automatische Quantifizierung liefert zusätzlich zur visuellen Interpretation wertvolle Informationen und verkürzt gleichzeitig die Gesamt-Bildbearbeitungszeit und bietet eindeutige Vorteile für wissenschaftliche Arbeiten. Eine automatische Methode für die Analyse der 123I-ADAM-Bindung an den zerebralen SERT liefert reproduzierbare Ergebnisse für die Einpassung der Studien in die Schablone und für die Berechnung der SBRs; sie könnte manuelle Methoden ersetzen.

Schattauer GmbH

123

I-ADAM: Vergleich manueller und automatischer Methoden

T

he serotonin transporter (SERT) has an important role in the modulation of the serotonergic neuronal function. Iodine-123 labelled 2-(2-dimethylamino-methyl-phenyl-thio)-5-iodophenylamine (ADAM) which binds the central nervous system serotonin transporters with high affinity and specificity has been shown to be a very promising ligand to detect SERT with single photon emission tomography (SPECT) in both human brain and nonhuman primates (1, 5, 6, 8, 16, 21, 25). In preliminary studies with humans specific regional uptake of 123I-ADAM was detected in brain regions such as midbrain and thalamus-hypothalamus, which have a high density of SERT. Also pretreatment with citalopram showed a clear effect for 123I-ADAM binding to SERT rich regions (e.g. in midbrain) (5, 6, 8, 21). ADAM is selected because this study is focusing to brain imaging of serotonergic system (transporter) with SPECT. ADAM is one of the most potential tracer for this purpose (21). The advantage of ADAM over nor-beta CIT is that ADAM is selective for the SERT. Nor-beta-CIT is not selective for SERT because it binds to other monoamine transporters, too. There are also other similar iodine-123 labelled serotonin transporter ligands which are used in SPECT, such as 123I-IDAM and 123I-ODAM. However, these ligands have either non-selective properties, e.g. lower affinity or lower specificity to the SERT (2, 3, 22). Biodistribution and dosimetry of 123 I-ADAM has been studied by different groups and 123I-ADAM has been shown to be a suitable imaging agent for studying the serotonin transporters in the brain (20, 23). Traditionally, the analysis of brain images is based on visual interpretation and binding ratio calculations. Visual methods Nuklearmedizin 5/2005

206 Kauppinen et al.

are highly subjective and the observer should be capable to use independent numerical values in addition to visual images. However, manually defined and drawn regions of interests (ROI) are sensitive to inter- and intraobserver variation caused by different interpretation and strongly dependent on various parameters, such as image quality, investigation setup. Therefore, an automated method for quantifying binding ratios and localizing regions of interests could be helpful (4, 26). An automated quantification method based on three-dimensional (3D) reference templates for normal values and variation (27) has been developed for inter-subject registration and quantification of brain SPECT images. This objective assessment of abnormalities includes relative quantification of the counts. Such template-based quantification requires automated alignment and sizing of images and has been demonstrated to yield reproducible quantitative results due to reliably and accurately registered images (26, 27). Automated registration algorithms and quantifications are mainly used for the brain perfusion SPECT studies instead of receptor, especially SERT, SPECT studies (11). The aim of this study was to compare manual and automated quantification methods and evaluate repeatability of inter- and intraobserver manual quantification compared to an automated method by using SERT radioligand 123I-ADAM. The purpose was also to produce a normal reference template and a 3D brain model for analysing SERT distribution by using 123I-ADAM studies done for healthy volunteers and then to find out whether automated quantification will help in the analysis of patient studies.

Patients, material, methods Study participants The subjects took part in a larger twin-study on 123I-ADAM binding amongst twins discordant or concordant for eating disorders and healthy control twins. The study participants were recruited from FinnTwin16, a population-based, longitudinal study inNuklearmedizin 5/2005

cluding virtually all Finnish twins born 1975–1979 (19). Persons with potential eating disorder were identified by a questionnaire including 30 eating disorder screening questions (e.g. 3 subscales of the Eating Disorder Inventory-1) (14) and SCID-I (13) telephone interviews, and further evaluated semi-structured EATATE- (7) and SSAGAinterviews (31) to establish DSM-IV (Diagnostic and Statistical Manual of Mental Disorders, American Psychiatric Association) diagnoses and to obtain a comprehensive history of eating disorders and other psychiatric morbidity. For this methodological analysis we used three groups: ● Group 1 (so called template-group): 15 healthy subjects (co-twin also healthy) whose 123I-ADAM -studies were used for producing the normal reference template, ● Group 2: 12 healthy subjects (co-twin had an eating disorder), ● Group 3: 10 subjects with actual BN (fulfilling the DSM-IV-criteria for BN). Because the current paper focuses on methodological issues and repeatability of quantification methods, no twin analyses or comparison between patients and controls are presented. Clinical aspects will be addressed in future works. In groups 1 and 3 all participants were women. Group 2 consisted of nine women and three men. One woman in group 2 had to be excluded because of an unsuccessful (extra vasated) radioligand injection. Thus, the final number of volunteers in this group was 11.The mean ages were 25.3 ± 1.9 years (group 1), 25.2 ± 1.9 years (group 2), and 24.9 ± 1.9 years (group 3). Respectively, the body mass indexes were 20.67 ± 2.18, 24.05 ± 3.47, and 23.49 ± 3.36 kg/m2. None of the study participants had any other medical disorder nor medications known to affect the serotonin transporter binding. They gave informed consent for their participation in the study. The Ethics Committee of the Kuopio University Hospital and Helsinki University Central Hospital approved the study and the work was performed in accordance with the Declaration of Helsinki. To all study participants 400 mg potassium perchlorate was given orally 30 minutes before the radioligand injection in

order to reduce 123I uptake in the thyroid and salivary glands.

Radiopharmaceutical, SPECT The brain imaging studies were done using 123I-labelled ADAM (MAP Medical Technologies Oy, Tikkakoski, Finland), the radiosynthesis of which has been described earlier (20). The amount of radioactivity in each syringe containing 123I-ADAM was measured in a dose calibrator before and after injection. Injected radioactivity of 123 I-ADAM was 177-231 MBq. The administered mean (± SD) dose was 207 ± 15 MBq (group 1), 205 ± 12 MBq (group 2), and 206 ± 15 MBq (group 3). Image acquisition was carried out 10 minutes and 5 hours after injection with a Philips Picker Prism3000XP three-headed gamma camera with ultra-high-resolution fan-beam collimators (Philips Medical Systems, Cleveland, OH, USA). The fan-beam focus of the collimator was 535 mm and the radius of rotation, measured from the surface of the collimator, varied within 130 and 160 mm, depending on the patient. SPECT acquisitions were performed using a 120° orbit in a stepwise mode. The person’s head was positioned to the centre of rotation in the head locker using a crossed laser beam system for repositioning. Positioning information (position and height of the bed) was recorded and used for the following acquisition. A symmetrical energy window (159 keV; 20% wide, 143-175 keV) was used for 123I, and the SPECT imaging was carried out with a 128 × 128 matrix size using 120 projection angles (40 projections/ detector). The scans were obtained at 45 s per projection angle, resulting in an average of 50 kcts per projection in the acquisitions acquired 10 min after injection and about 20 kcts in 5 h acquisitions. Images acquired 5 h after injections were used for quantifications. Selection of the time points was based on previous studies by the same group (5, 6). All three study groups underwent an identical acquisition protocol.

207 Manual and automated methods of 123I-ADAM

Data reconstruction

Quantification and data analysis

After acquisitions, the data were transferred into a HERMES software system (Hermes Medical Solutions, Stockholm, Sweden), for all reconstructions and image analyses. An iterative reconstruction program HOSEM (OS-EM V5.201 by R. Larkin) in the HERMES software system was used for iterative reconstructions (17). Transverse slices oriented on the orbitomeatal line were reconstructed. The number of subsets was chosen to be 8 with 6 iterations. Attenuation correction was performed during the reconstruction using Chang’s first-order approximation with the linear attenuation correction (m = 0.110 cm-1), which was based on an ellipse contour of the brain. The images were post-filtered using a Butterworth filter with cut-off frequency of 1.2 cm-1 and order 15.

Manual quantification

Reference regions The 15 persons of group 1 (template group) underwent SPECT 123I-ADAM scans using the protocol mentioned. In addition, a magnetic resonance imaging (MRI) was taken from every subject (Siemens Vision 1.5T with MP_RAGE sequence: TR 9.7 ms, TE 4.0 ms and flip angle 12°; Siemens AG, Erlangen, Germany). This MRI was further used as an anatomical reference for finding volumes of interest (VOI). One arbitrarily selected normal 123 I-ADAM scan was then co-registered with the same person’s MRI. Registration was performed with the application MultiModality on HERMES using an automatic algorithm with 6 degrees of freedom (size changes were restricted, 3 for rotation and 3 for translation) and “Mutual Information” method as measure of similarity (24). Next, all 123I-ADAM scans of the template group were spatially co-registered with this first one using BRASS (Brain Registration and Analysis of SPECT Studies) software (Hermes Medical Solutions, Stockholm, Sweden) on a HERMES (27). For this registration we used 9 degrees of freedom and Mutual Information (Fig. 1). Finally, these 15 co-registered SPECT scans were normalized to the total counts and averaged to build

Fig. 1 Definition and orientation of adjustment parameters: three for translation, three for scaling and three for rotation

a template containing mean values and distributions within the template group in every pixel. Anatomically standardized normal reference templates for the template group were created using the Modelgen software (Hermes Medical Solutions, Stockholm, Sweden). This module intrinsically generates a mean and variance 3-dimensional composite brain template. The magnitude of activity differences over extended predefined regions was evaluated through automatic calculation of activity in 14 predefined VOIs. For each individual scan, the VOI activity counts were calculated per voxel and normalized to the total number of counts of the complete VOI set. This normal template was then used in conjunction with the – intrinsically co-registered – sample MRI to define a set of volumetric regions for all further analysis. This set of VOIs contains the following regions: pons, midbrain, left and right temporal lobe, hypothalamus, putamen, caudate nucleus, anterior gyrus cingulate, thalamus, and as reference region the cerebellum. The anatomically standardized (stereotactic) images were used for automated VOI quantification. All image registrations were done by rigid body transformations to keep count rate information as accurate as possible and the user interaction to a minimum to reduce the sources of variability, .

Two reconstructed and attenuation corrected transverse slices were consecutively summed together. ROIs were drawn onto the cerebellum, left and right putamen, caudate nucleus and striatums, thalamus, and midbrain. ROIs (except for midbrain) were defined onto images acquired 10 min after injection and then the same ROIs were used for images 5 h after injection. The midbrain ROIs were drawn onto the images obtained after 5 h and transferred onto those of 10 min, because midbrain area is generally more clearly visualized from its surroundings after 5 h . For the midbrain ROI a lower threshold (30%) was used. Manually defined ROIs were drawn by two separate observers so that the other observer repeated her ROI definitions after 6 months after the initial ROI drawings. One of the observers was a physician while the second observer was a medical physicist. Observers were blinded and patient names were coded. Both inter- and intraobserver repeatability was tested in group 3 (patients with BN). Only interobserver repeatability was tested in group 2. The specific binding ratios (SBR) of each ROI were calculated by the following formula: SBR =

(target activity - reference activity in brain) reference activity

which can be operationally expressed by the formula: SBR =

(mean counts in midbrain - mean counts in cerebellum) mean counts in cerebellum

Half-life correction was done for calculated SBR values. SBRs were calculated for all defined ROIs, but only midbrain SBRs were used for methodological comparisons between manual and automated methods. Cerebellum served as a reference region. Even if it contains some SERTs, they are so much less than inother brain areas, that this is the best region for reference (1, 12).

Nuklearmedizin 5/2005


direction, 10% in scaling and 10° in rotation. Because YZ rotation might vary more depending on patient's positioning this parameter was defined to vary up to 20°. The amount of misalignment was based on typical magnitudes occurring in realistic clinical situations. These misalignments were re-registered onto to the same reconstructed datasets, which were the basis for the first registration. SBR of midbrain was re-calculated after misalignments.

Statistical analyses

Fig. 2 The 123I-ADAM mean template based on the data of 15 healthy volunteers acquired 5 h after injection. Example of the template aligned to a MRI dataset (A) and matched stereotactic VOI map (B) with transverse, coronal and sagittal views. The midbrain area (green cross) is shown in all three views. Other VOIs starting from the bottom are cerebellum, pons, left and right side of the temporal lobe, hypothalamus, putamen, caudate nucleus, anterior gyrus cingulate, thalamus.

Automated quantification The images of the groups 2 and 3 were registered to the template using a Mutual Information algorithm (24) to have common orientation, position, and size using the BRASS software. The accuracy of the registration was assessed by visual inspection of the overlaid images. Furthermore, the accuracy of the fitting was verified by checking the nine transformation parameters (three for translation, anisotropic scaling, and rotation), which were calculated automatically by the software. Altogether, maximum 1000 iterations were used for the registration of the datasets. Registration was terminated at an accuracy level of 0.1% changes of mutual information by subsequent iterations. The distribution of SERT in the brain was determined by the predefined region map. After spatial co-registration, the patient's study was normalized to have the same number of total counts as the template. By means of this region map, the counts in each VOI region were quantified. Finally, SBRs were calculated from counts of predefined VOI areas of the 3D region map. Nuklearmedizin 5/2005

Every scan obtained from the persons in groups 2 and 3 was loaded into BRASS. The scans were spatially co-registered to the normal database (template group) and the underlying MRI, and normalized to the total count rate and quantified using the predefined VOI map. One of the purposes of this work was to create a unified set of VOIs. By a unified, predefined set of VOIs, we could rule out inter-observer variability and badly defined ROIs due to lack of anatomical information because of low resolution of the images (26, 30).

Repeatability of automated technique The precision of the registration was verified by the group 3. After the first SBR quantification, we misaligned the original images of the persons and then re-registered them with the template and quantified them another time. Alignments were changed for each patient using randomly selected values generated in Excel2000 (Microsoft Corporation, USA) for every registration parameter. Misalignment parameters varied between 10 pixels in translation in X, Y, and Z

The inter- and intraobserver repeatability of the manual quantifications and the repeatability of the automated quantifications were investigated by using a Bland-Altman plot, where the difference between measurements is plotted against their mean value. Because the true value is unknown and no so-called “golden standard” is available to compare to, the mean value of the repeated measurements is the best estimate we have (9). It is also possible to evaluate repeatability of different methods by using numerical values calculating coefficient of repeatability (s), which has been described in the paper by Bland and Altman (9). Coefficient of repeatability was determined as twice the standard deviation of differences between analyses (9, 10). Repeatability of registrations was tested performing the misalignment for group 3. The difference between both registrations for each patient was tested by the program, comparing two registrations on a voxel-byvoxel basis and calculating the mean variation of the registered studies, which describes the percentage difference or change in voxels between both registrations.

Results Template The mean template (based on scans of subjects in group 1) registered with MRI dataset and defined VOIs are shown in figure 2. The mean SBR in midbrain at 5 h after radioligand injection was 1.85 ± 0.33 (± SD).


a) Fig. 3

b) Inter- and intraobserver repeatability in SBR of midbrain in bulimia patients: Repetition of one (A) and two different observers (B)

Correspondingly the manually defined and calculated SBR value was 2.09 ± 0.42.

Repeatability

When interobserver variation was done for the group 2, it turned out to amount to s = 0.61.The Bland-Altman plot of the interobserver repeatability for this group is shown in the figure 4.

Manual quantification

Registration and automated quantification

The inter- and intra-observer variations of SBRs in midbrain in group 3 are shown in figure 3. The coefficient of repeatability for one observer was s = 0.70, when analysis was repeated six months after the initial ROI drawing. Respectively, the coefficient of repeatability for two blinded analysers was s = 1.00.

Repeatability of automated registration and quantification was tested performing the misalignment described above for group 3. After misalignment both registration datasets were loaded onto BRASS and overlaid. The difference between both registrations for each patient was tested by the program, comparing two registrations on a voxel-by-

a)

voxel basis.All repeatedly registered studies were fitted correctly. In 3 of 10 cases the program could not find any differences. In the other cases only slight differences were noticed. The mean variation of the registered studies was 0.38 ± 0.15% (mean ± SD). For every study, repeated automated quantification yielded exactly the same results if repeated without the misalignment described. If the studies were misaligned prior to the quantification, the coefficients of repeatability for study group 3 are shown in table 1. Repeatability in SBR of midbrain is depicted in figure 5. Results were neither dependent of the person running the program nor the time it was run.

b)

Fig. 4 Repeatability in SBR of midbrain a) in the group of the healthy volunteers b) calculated by using automated technique (repetition performed after misaligned registration)



Tab. 1

Repeatability for study group 3 in brain areas

brain area

coefficient of repeatability, s

midbrain

0.13

oons

0.17

right temporal lobe

0.23

left temporal lobe

0.25

hypothalamus

0.20

right putamen

0.22

left putamen

0.22

right caudatus

0.23

left caudatus

0.23

right gyrus gingulus

0.11

left gyrus gingulus

0.19

right thalamus

0.24

left thalamus

0.22

mean ± SD

0.20 ± 0.04

SBR values for each defined VOI are presented in table 2, where are SBR results of template group and group of healthy volunteers at 10 min and 5 h after injection.

Discussion It has been demonstrated that 123I-ADAM with SPECT imaging is a potential tracer for imaging the brain SERTs (5, 6, 8). Manual brain area

group I: SBR

quantification techniques are sensitive not only to the “handwriting” of the observer but also for experience and reproducibility (28). Results are further highly dependent on the acquisition technique and even more on general image quality. It has also been published that automated techniques are clinically feasible, accurate and less time consuming than operator dependent methods (29). Compared to manual techniques, automated methods are more standardized and objective alternatives for brain quantification. This kind of software programs enable statistical comparisons between different patient groups not only in brain perfusion studies but also in the brain receptor studies (11, 26). A quantitative analysis of brain SPECT is important for clinical evaluations, especially when researching new radioligands, and thus observer-independent quantification techniques are required. In this study we present an automated quantification technique for the novel radioligand 123I-ADAM and compare it with an established manual technique. The automated technique had a better reproducibility than manual methods no matter whether the quantifications were repeated by one or two observers. This was quantified using the coefficient of repeatability. The inter- and intra-observer variability is remarkable contrary to automated method, which produced repeatedly same results. However,

group II: SBR

10 min

5h

10 min

5h

midbrain

0.11 ± 0.07

1.85 ± 0.33

0.15 ± 0.06

1.96 ± 0.45

pons

0.03 ± 0.05

0.69 ± 0.22

0.02 ± 0.05

0.78 ± 0.22

right temporal lobe

0.00 ± 0.05

0.72 ± 0.18

0.00 ± 0.04

0.69 ± 0.27

left temporal lobe

0.00 ± 0.03

0.62 ± 0.22

0.00 ± 0.06

0.66 ± 0.25

hypothalamus

0.00 ± 0.11

0.42 ± 0.32

0.00 ± 0.09

0.29 ± 0.29

right putamen

0.17 ± 0.11

1.16 ± 0.34

0.08 ± 0.11

1.11 ± 0.40

left putamen

0.20 ± 0.07

1.15 ± 0.41

0.13 ± 0.13

1.22 ± 0.44

right caudatus

0.37 ± 0.08

1.35 ± 0.27

0.38 ± 0.07

1.49 ± 0.41

left caudatus

0.29 ± 0.06

1.17 ± 0.31

0.32 ± 0.08

1.19 ± 0.40

right gyrus gingulus

0.07 ± 0.10

0.69 ± 0.26

0.02 ± 0.09

0.75 ± 0.29

left gyrus gingulus

0.00 ± 0.07

0.80 ± 0.25

0.00 ± 0.09

0.89 ± 0.29

right thalamus

0.20 ± 0.07

1.32 ± 0.27

0.16 ± 0.06

1.39 ± 0.41

left thalamus

0.29 ± 0.08

1.47 ± 0.24

0.24 ± 0.11

1.60 ± 0.31


Tab. 2 SBR values for each defined VOI of template group and healthy volunteers

there are some possible pitfalls for automatic quantification. Firstly, accuracy of the positioning of VOIs due to the fitting procedure can be questioned. However, the software used in our study has been proven with other tracers to give reliable results (26, 28-30). Secondly, normalization may play a role when quantitatively comparing a patient or a volunteer study with the template. This is true, and great care has been taken to ensure correct results. BRASS allows different normalization methods (by a reference region, by maximum counts or by total counts) and total counts were chosen here because of the most stable results. Template based methods were earlier used mainly for statistical comparison or group comparisons, but nowadays they offer advantages for SPECT quantifications (11, 28). It was noticed that a reliable template enables registrations between different study groups and thus template based techniques might be useful, e.g. for specific binding ratio calculations. True registration parameters of clinical data are never exactly known and thus there is no golden standard for image registration. It is difficult to give an exact estimate of the true registration error. In this study, we demonstrated that misaligned and re-registered images were fitted reproducibly and correctly as judged by visual inspection of the images. The calculated mean variation between these was less than half percent. For the goals of ● an exact quantification and ● to keep user interaction and thus user variability to a minimum, we chose a registration algorithm that gives full flexibility by allowing rigid transformations in nine degrees of freedom, but does not involve image warping. Non-linear registration techniques are among the most up-to-date research topics in medical imaging. New algorithms for various parts of the body are published almost weekly. However, all these algorithms involve a priori assumptions that determine the behaviour in those areas that are warped. Visually, these algorithms give highly satisfactory results, as demonstrated on many occasions. Numerically, all these models do not ad-


equately fulfil the need to keep the counts in the regions warped. Furthermore, it has been shown that warping accuracy is significantly increased by manual definition of landmarks (18). However, in this work we are presenting a fully automated technique for image quantification, which on purpose does not allow any user interaction to keep user dependent variability low. SBR is dependent on the size or volume of both the reference region and target areas. Defined normal values are valid only for one material or study group and if the acquisition or image handling parameters, especially filtering, change, they might affect SBR values. SBR values between different patient materials are valid only if defined regions are equal between the compared materials. Hence registration and determination of the VOIs are important for both manual and automated techniques. However, manual ROI and VOI definitions are more sensitive for “handwriting” and other individual variations than predefined constant ROIs or VOIs. If a reliable registration of the study under investigation to the template can be achieved, then it is advisable to use predefined ROIs or VOIs. Only then it can be assured that subjective interpretation of image imperfections will not influence the size and position of ROIs. It has be shown that the following factors influence the manual positioning of the ROIs (15): ● Partial volume effects in SPECT may make regions of higher uptake appear larger than they are. ● Poor count statistics or target-to-background ratio in the image makes it sometimes difficult to define the edges of anatomical structures. Possible reasons might be the radioligand itself and its specificity, the injected dose, medication, which might influence brain metabolism, and the disease. ● Thresholding will be different for different observers and thus the size of the regions might be different, too. Only by careful definition of the regions under investigation on high-resolution anatomical scans, it can be assured that outlines are drawn correctly. Provided a sufficient

registration, these exact regions will be applied correctly and in the same manner to all patient and control studies independent of any of the above factors. In practise, however, it may sometimes be necessary to slightly correct manually the positioning of the ROIs due to individual variations in anatomy. In that case, keeping the brain level and the size of ROI/VOI unchanged, repeatability should not suffer significantly.

Conclusion A key purpose of this study was to illustrate an automated method for quantification of brain SERT-binding of 123I-ADAM and to show that an automated method is more objective and not as sensitive for individual variation as the manual technique. Automated quantification techniques allow better repetition of the calculated results. It was noticed by the study presented here that automated quantification gives valuable information in addition to visual interpretation and will help in the analyses of patient studies using 123I-ADAM. Acknowledgement This work was granted partly by the US-PHS National Institute of Alcohol Abuse and Alcoholism (AA12502),Academy of Finland (44069 and 201461) & European Union Fifth Framework Program (QLRT-1999–00916, QLG2-CT-2002–01254) and the special governmental subsidy for research and education at the Helsinki University Central Hospital. The authors thank MAP MedicalTechnologies and especially Ph.D. Kim Bergström for their help and support with 123I-ADAM. The authors wish to express their sincere thanks also to Prof. Aila Rissanen, Prof. Jaakko Kaprio and Mr. Timo Lukkarinen for their kind help.

References

1. Acton PD, Choi SR, Hou C et al. Quantification of serotonin transporters in nonhuman primates using 123I-ADAM and SPECT. J Nucl Med. 2001; 42: 1556–62. 2. Acton PD, Kung MP, Mu M et al. Single-photon emission tomography imaging of serotonin transporters in the non-human primate brain with the selective radioligand 123I-IDAM. Eur J Nucl Med 1999; 26: 854–61. 3. Acton PD, Mu M, Plossl K et al. Single-photon emission tomography imaging of serotonin transporters in the nonhuman primate brain

with 123I-ODAM. Eur J Nucl Med 1999; 26: 1359–62. 4. Acton PD, Pilowsky LS, Kung HF et al.Automatic segmentation of dynamic neuroreceptor singlephoton emission tomography images using fuzzy clustering. Eur J Nucl Med 1999; 26: 581–90. 5. Ahonen A, Koskela A, Kauppinen T et al. Serotonin transporter availability in patients with symptomatic bulimia using a novel SPET ligand 123 I-ADAM. Eur J Nucl Med 2003; 30 (suppl): S214. 6. Ahonen AK, Kauppinen TA, Heikman P et al. I-123 labeled ADAM – a selective novel radioligand for imaging of serotonin transporters in the human brain. J Nucl Med 2002; 43 (suppl): 232P. 7. Anderluh MB, Tchanturia K, Rabe-Hesketh S et al. Childhood obsessive-compulsive personality traits in adult women with eating disorders: defining a broader eating disorder phenotype. Am J Psychiatry 2003; 160: 242–7. 8. Asembaum S, Füger B, Diemling M et al. 123IADAM SPECT – a new method for investigating serotonin transporter in humans. Eur J Nucl Med 2003; 30 (suppl): S309. 9. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986; 1: 307–10. 10. British Standards Institution. Precision of the methods. 1: Guide for the determination and reproducibility of a standard test method (BS5497, part 1), London, UK: BSI 1979. 11. Catafau AM. Brain SPECT in Clinical Practice. Part I: Perfusion. J Nucl Med 2001; 42: 259–71. 12. Cortes R, Soriano E, Pazos A et al. Autoradiography of antidepressant binding sites in the human brain: localization using 3H-imipramine and 3H-paroxetine. Neuroscience 1988; 27: 473–96. 13. First MB, Spitzer RL, Gibbon M et al. Structured Clinical Interview for DSM-IV-TR Axis I Disorders, Research Version, Nonpatient Edition. NewYork, USA: Biometrics Res., NewYork State Psychiatr. Inst. 2002. 14. Garner DM, Olmsted MP, Polivy J et al. Comparison between weight-preoccupied women and anorexia nervosa. Psychosom Med 1984; 46: 255–66. 15. Habraken JBA, Booij J, Slomka P et al. Quantification and visualization of defects of the functional dopaminergic system using an automatic algorithm. J Nucl Med 1999; 40: 1091–7. 16. Huang WS, Ma KH, Chou YH et al. 123I-ADAM SPECT in healthy nonhuman primates: a preliminary report.Ann Nucl Med Sci 2002; 15: 27–32. 17. Hudson HM, Larkin RS. Accelerated image reconstruction using ordered subsets of projection data. IEEE Trans Med Imaging 1994; 13: 601–9. 18. Kang X, Bertrand O, Alho K et al. Local landmark-based mapping of human auditory cortex. Neuroimage 2004; 22: 1657–70. 19. Kaprio J, Pulkkinen L, Rose RJ. Genetic and environmental factors in health-related behaviors: studies on Finnish twins and twin families. Twin Res 2002; 5: 366–71.



20. Kauppinen TA, Bergstrom KA, Heikman P et al. Biodistribution and radiation dosimetry of 123 I-ADAM in healthy human subjects: preliminary results. Eur J Nucl Med Mol Imaging 2003; 30: 132–6. 21. Kuikka JT. Quantitative accuracy of serotonergic neurotransmission imaging with high-resolution 123 I SPECT. Nuklearmedizin 2004; 43: 185–9. 22. Kung MP, Hou C, Oya S et al. Characterization of 123 I-IDAM as a novel single-photon emission tomography tracer for serotonin transporters. Eur J Nucl Med 1999; 26: 844–53. 23. Lin KJ,Ye XX,Yen TC et al. Biodistribution study of 123I-ADAM in mice: correlation with whole body autoradiography. Nucl Med Biol 2002; 29: 643–50. 24. Maes F, Collignon A, Vandermeulen D et al. Multimodality image registration by maximization of mutual information. IEEE Trans Med Imaging 1997; 16: 187–98. 25. Oya S, Choi SR, Hou C et al. 2-(2-dimethylaminomethyl-phenyl-thio)-5-iodophenylamine


(ADAM): an improved serotonin transporter ligand. Nuc Med Biol 2000; 27: 249–54. 26. Radau PE, Linke R, Slomka PJ et al. Optimization of automated quantification of 123I-IBZM uptake in the striatum applied to Parkinsonism. J Nucl Med 2000; 41: 220–7. 27. Slomka P, Stephenson J, Reid R et al. Automated template-based quantification of brain SPECT. In: De Deyn PP, Dierckx RA, Alavi A et al. (eds). SPECT in neurology and psychiatry. London: John Libbey Company Ltd. 1997: 507–12. 28. Van Laere K, de Sadeleer C, Dobbelair A. Quantification of 99mTc-HMPAO brain SPET in two series of healthy volunteers using different tripleheaded SPET configurations: normal databases and methodological considerations. Nuc Med Commun 1999; 20:1031–40. 29. Van Laere K, Koole M, D’Asseler Y et al. Automated stereotactic standardization of brain SPECT receptor data using single-photon transmission images. J Nucl Med 2001; 42: 361–75.

30. Van Laere KJ, Warwick J, Versijpt J et al. Analysis of clinical brain SPECT data based on anatomic standardization and reference to normal data: An ROC-based comparison of visual, semiquantitative, and voxel-based methods. J Nucl Med 2002; 43: 458–69. 31. Wade T, Heath AC, Abraham S et al. Assessing the prevalence of eating disorders in an Australian twin population. Aust N Z J Psychiatry 1996; 30: 845–51.

Correspondence to: Tomi Kauppinen, Ph.D. HUS Helsinki Medical Imaging Center Helsinki University Central Hospital P.O. Box 750, 00029 Helsinki, Finland E-mail: [email protected] Tel.+358/9/47 18 00 49 Fax+358/9/47 17 13 54