An automated technique for SPECT marker-based ... - Semantic Scholar

INSTITUTE OF PHYSICS PUBLISHING

PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 46 (2001) 2085–2097

PII: S0031-9155(01)20337-6

An automated technique for SPECT marker-based image registration in radionuclide therapy Periklis Papavasileiou, Glenn D Flux, Maggie A Flower and Matthew J Guy Joint Department of Physics, Institute of Cancer Research and Royal Marsden NHS Trust, Royal Marsden Hospital, Downs Road, Sutton SM2 5PT, UK E-mail: [email protected]

Received 2 January 2001, in final form 11 May 2001 Published 6 July 2001 Online at stacks.iop.org/PMB/46/2085 Abstract An automated technique for marker-based image registration in radionuclide therapy is described. This technique is based on localization of the centroids of external markers and on establishing correspondence between the individual markers of the two studies to be registered. Localization of the centroids of markers relies on segmenting the markers using iterative thresholding. Thresholding is locally adaptive in order to account for study-dependent conditions (e.g. crossover between adjacent markers and markers with varying radioactive concentrations). Following marker segmentation, the centroids of the markers are computed based on an intensity-weighted method. Finally, prior to the least-squares fit registration, the markers of the two sets are matched to achieve one-to-one correspondence. The technique was applied to both simulated and patient studies resulting in mean residual three-dimensional registration errors (±1 SD) of 1.7±0.1 mm and 3.5±0.3 mm respectively. The technique was compared with a semi-automated approach and no significant difference was found between the mean residual three-dimensional registration errors (t = 0.281, p = 0.8). This automated marker-based image registration technique provides robust and accurate registration. Although it was developed as part of a programme to generate three-dimensional dose distributions for radionuclide therapy, it could be useful for other clinical applications.

1. Introduction In three-dimensional (3D) dosimetry for 131 I radionuclide therapy, administration of a 131 Ilabelled tumour-specific agent is followed by a series of sequential SPECT (single photon emission computed tomography) scans to characterize the dose distribution to both tumour sites and normal organs (Flux et al 1997, 1999). Registration of the sequential 131 I SPECT scans is essential to allow 3D dose distributions to be generated. 0031-9155/01/082085+13$30.00

© 2001 IOP Publishing Ltd

Printed in the UK

2085

2086

P Papavasileiou et al

Image registration involving nuclear medicine studies has been mainly concerned with the brain (Pelizzari et al 1989, Alpert et al 1990, Woods et al 1992, 1993). Dey et al (1999) performed CT–SPECT registration of an anthropomorphic phantom using a radionuclide transmission scan. The automated Woods’ algorithm was validated on SPECT–SPECT registration using an anthropomorphic phantom and proved to be very accurate. In the thorax, Yu et al (1995) used surface matching to register CT and PET scans. Eberl et al (1996) and Bacharach et al (1993) have described methods for SPECT–PET and PET emission– transmission scans based on the sum of absolute differences and the correlation coefficient respectively. In radionuclide therapy, the lack of anatomical information as well as the changes in the radiotracer uptake between serial SPECT scans, due to the biological washout of the administered activity and the physical decay of the isotope, constitute the main limitations of non-marker registration approaches. The use of external markers enables image registration irrespective of any changes in radioactive uptake between the scans of interest. Another approach, similar to that of the external markers, was described by Scott et al (1995) for SPECT–CT thoracic and abdominal registration using surface matching. These authors fitted an external fiduciary band containing tubing which was filled with 99m Tc around the patient’s body. The fiduciary band was visible on the CT scan thus allowing the generation of stacks of 2D contours on both CT and SPECT scans. At our centre, radionuclide therapy SPECT–SPECT image registration is performed using external fiducial markers filled with 99m Tc. Marker-based image registration requires the localization of the centroids of the markers. Localization can be carried out either manually or by using a semi-automated approach. The latter requires an initial estimate of the marker centroid and then employs intensity-weighted centroiding. In previous work, one-to-one correspondence between the two marker centroid sets (marker matching) was established interactively (Flux 1995). In this paper, the registration process is improved by automating both the localization of the centroids and the marker matching processes. Automated localization of the centroids requires segmentation of the markers which is based on iterative thresholding combined with a Gaussian profile criterion (i.e. profiles through the centroids should be Gaussians). The centroids of the markers are then calculated and one-to-one correspondence between the two sets of markers is automatically established using geometry-based rules. The automated marker-based registration technique was validated using both simulated and patient studies. The algorithms were written in C and IDL (Research Systems Inc., Colorado, USA) and have been included in the Royal Marsden Dosimetry Package (RMDP) (Guy 2000). 2. Methods 2.1. Preprocessing of the SPECT marker scan The automated technique for localization of the centroids of the markers described in this paper assumes that the markers’ 99m Tc count density is higher than that due to any scattered 131 I gamma rays which can be present in the 99m Tc energy window. This assumption may be invalid in very early scans following administration of a therapeutic activity (e.g. first day scan), as shown in figure 1. Such early scans are rarely acquired due to dead-time correction and radiation protection problems. In this rare situation, to avoid the down-scatter from 131 I interfering with the segmentation process, the tumour volume is subtracted from the marker scan. The subtraction is carried out automatically exploiting the fact that the 131 I and 99m Tc scans are inherently registered. The maximum voxel value in the 131 I scan is determined and an empirical threshold of 30% of this value is set. All voxels with values above this threshold

Automated SPECT image registration using markers

(a)

2087

(b)

Figure 1. Illustration of the effects of downscatter from the 131 I into the 99m Tc energy window. (a) Transverse slice through a 99m Tc marker scan. (b) A transverse slice through the 131 I scan, corresponding to (a). The 99m Tc marker scan was acquired immediately prior to the 131 I scan. Both scans were acquired 24 h after the therapy activity administration.

are considered as tumour voxels. Then, the corresponding voxels on the 99m Tc scan are set to zero. The possible existence of multiple tumour sites is also taken into consideration. Hence, the above process is repeated as long as 131 I therapy scan voxels with an intensity greater than 30% of the initial maximum are found. 2.2. Automated localization of marker centroids Marker localization is an iterative process. It relies on the fact that the intensity of the 99m Tc markers is sufficiently high so that down-scatter from the 131 I-labelled therapy agent does not interfere with the segmentation process. All voxels with image intensities above an upper threshold of 70% of the maximum are considered to be potential marker voxels, whereas all voxels below a lower threshold of 20% are considered to be background. These initial values were chosen empirically following analysis of several different 99m Tc SPECT marker studies. Both thresholds are allowed to fluctuate within a predefined range in order to allow all the markers to be identified, irrespective of study-dependent conditions such as ‘crossovers’ between markers positioned very close to each other and/or any difference in marker intensities due to varying levels of total radioactivity. The automated process for localization of marker centroids is described as a flow-chart in figure 2 and consists of the following steps: 1. A voxel with an intensity above the upper threshold is identified and growing of the 3D region is carried out to generate the potential marker. 2. A Gaussian profile criterion is formulated exploiting the fact that intensity profiles through the markers are Gaussian distributions. Three orthogonal profiles through the arithmetic centroid are taken which will be Gaussian distributions unless a crossover between adjacent markers has occurred due to their proximity (figure 3). The goodness of fit of a Gaussian to each profile is determined using the Kolmogorov–Smirnov test (Armitage and Berry, 1994). If a crossover is detected the lower threshold is increased locally by 10% in order to ‘break the bridge’ between the two markers (locally adaptive thresholding) and the

2088

P Papavasileiou et al START Compute image maximum #iterations=0 ; #identified_markers=0 upper threshold=70% ; lower threshold=20%

N

while #iterations < 3

Y N

while not EOF

identify voxel with intensity above upper threshold

Increment #iterations

N

#iterations = 3

generate 3D connected component for each marker

if #identified_markers < #expected_markers

Y

3D region growing - 2D radial growing regulated by the lower threshold - Neighbourhood connectivity to allow 2D region growing in neighbouring slices.

Y

decrease upper threshold by 10%

Gaussian profile constraint - compute arithmetic centroid of the 3D marker - sample radial profiles through the centroid ; all should be Gaussian distributions, otherwise a crossover between adjacent markers is identified N

marker centroid computation

END

if crossover

Y

accept marker Delete marker voxels from file. Compute new image maximum lower threshold = 20%.

discard marker increase lower threshold by 10%

Increment #identified_markers

Figure 2. Flowchart of the process of SPECT marker centroid localization (# = number of, EOF = end of marker scan file). The upper and lower threshold values are given as percentages of the maximum image intensity, which varies during the process as each marker is identified.

process returns to step 1. Otherwise, the potential marker is accepted as a correct marker, the corresponding marker voxels are deleted to avoid being resampled and the new image maximum is computed. 3. If following the first iteration the number of segmented markers is less than expected, it is assumed that there is a varying radioactive concentration in the markers. To allow for this discrepancy, the upper threshold is decreased by 10% and the process returns to step 1. The upper threshold is not allowed to decrease below 20% of the initial maximum voxel intensity. In almost all cases two iterations are sufficient to identify all the markers and thus the maximum number of iterations is set to three. 4. Generation of a marker 3D connected component. Following segmentation, the segmented markers’ voxels are copied to a temporary binary file where they are assigned an intensity of 1. 3D six-neighbourhood connectivity is employed to allow generation of the marker 3D connected components. The four immediate 2D voxel neighbours within the slice and the two immediate ones in the neighbouring slices are considered to establish 3D connectivity. 5. The intensity-weighted marker centroid is determined using the intensities in the original image of the voxels comprising the marker 3D connected component.

Automated SPECT image registration using markers

2089

220 200 180

Voxel intensity

160 140 120 100 80 60 40 20 0 -40

-30

-20

-10

0

10

20

30

40

Dis tance of profile voxels from arithmetic centroid of potential marker (mm)

(a)

(b)

Figure 3. (a) Slice through a potential marker (crossover). (b) Profile through the arithmetic centroid of the potential marker. The position of this profile is indicated on (a).

2.3. Marker matching Marker matching is essential to establish one-to-one correspondence between the markers in the two scans to be registered. The scan that remains unaltered is referred to as the reference scan, whereas the scan to be registered to the reference is called the mobile scan. In order to uniquely identify any number of markers a 3D coordinate system is defined by three different and noncoplanar markers. These are referred to as principal markers. The matching process is divided into two subprocesses. The first subprocess identifies three corresponding principal-marker pairs on the reference and mobile marker sets, whilst the second maps the remaining markers based on the coordinate system defined by the principal markers. The idea of principal markers was studied previously by Boesecke et al (1990); these authors used anatomical landmarks, selected by an operator, instead of external markers to form the basis of matching any additional landmarks whose coordinates on the mobile scan were unknown. The underlying assumption of the matching process is that the order in which the markers are segmented on both reference and mobile scans is the same. This assumption holds for abdominal and thoracic scans where the patient is lying flat on the couch and therefore rotation around the x- and y-axes is minimal. Rotation about the z-axis will not significantly affect the order of marker segmentation. If the marker centroids are considered as two arrays ‘sorted’ according to the order of marker segmentation the algorithm allows for two corresponding markers to differ by ±1 position in their respective marker centroid sets. A change in relative position may occur when two markers are placed at the same axial position, so that their centroid appears on the same slice. A change in the order in which the two markers are identified might result from slight mispositioning of one of the markers. The matching process consists of the following steps: 1. The centroids of the reference and mobile marker distributions are computed. 2. The distances of the individual marker centroids to the centroids of their respective marker distributions are computed. 3. The mobile marker centroid furthest away from its respective distribution centroid is determined. 4. A subset of the reference marker distribution is defined based on the assumption that corresponding markers are allowed to differ by ±1 position in their respective marker centroid sets. The search for establishing correspondence is confined to this subset.

2090

P Papavasileiou et al

5. The criterion for establishing correspondence between two marker centroids is that the distances to their corresponding distribution centroids differ by less than 15 mm; this threshold of 15 mm was set empirically. If more than one reference marker centroid satisfies the above criterion, the mobile marker centroid determined in step 3 is excluded from further analysis and the process returns to step 3. If the above criterion is satisfied, the two marker centroids constitute a potential principal marker pair. 6. Steps 1–5 are repeated with the roles of reference and mobile marker sets reversed. This ensures that the algorithm is robust and invulnerable to the order of registration. Ideally, the same potential principal-marker pairs should be identified irrespective of the order of registration. However, because of the established threshold of 15 mm in determining the potential principal markers (step 5) and the possibility of a change in the order in which the markers have been segmented, different potential principal markers can be identified when the order of reference and mobile sets is reversed. Potential principal-marker pairs resulting from matching reference to mobile and mobile to reference sets are compared and common pairs are selected. 7. The number of common potential principal-marker pairs might be more than three. Therefore each triplet of potential principal-marker pairs is studied in order to establish that they can define a unique coordinate system. Prior to testing the uniqueness of the 3D coordinate system, a further criterion is set for the triplet of common principal markers. A vector is defined by each potential principal-marker and the respective marker distribution centroid. The angle between each pair of vectors on both the reference and mobile sets is computed. These angles should not differ by more than 6◦ . This angular constraint is set to ensure that markers, whose centroid is localized on the same axial position (z plane) and on symmetrical transaxial positions (x–y plane) with respect to the marker distribution centroid, are correctly matched. 8. A non-coplanarity criterion is established to ensure that the principal-marker centroids define a 3D coordinate system (figure 4). Consider a plane defined by the marker distribution centroid and two of the principal-marker centroids. If the distance of the third principal-marker centroid to that plane is close to zero a plane of symmetry exists and hence marker identification based on that triplet of principal-marker centroids is nonunique. The non-coplanarity criterion is satisfied when the above distance is greater than 15 mm; this value is set because of the threshold of 15 mm imposed when determining the potential principal-marker centroids. 9. If the analysed triplet satisfies the non-coplanarity criterion the algorithm continues to step 10; otherwise, it returns to step 7 to analyse the next triplet of potential principal-marker centroids. In the case of failure of the algorithm to identify principal-marker centroids the technique is transformed into a semiautomated one where the matching of the segmented markers is performed interactively. Thus, the process of identification of principal markers becomes a ‘sensor’ for either marker mispositioning and/or for severe rotation, which can result in significant change in the order of marker segmentation. 10. Correspondence is established between the remaining non-principal reference and mobile marker centroids based on their distances from the reference and mobile principal-marker centroids respectively. The requirement for establishing correspondence is that the distances should differ by no more than 15 mm. 2.4. Registration The least-squares fit algorithm developed by Arun et al (1987) was employed to compute the six transformation parameters (X, Y , Z translation, XY , XZ and Y Z rotation) required to

Automated SPECT image registration using markers

2091

m1 pmD

P

pmB m2

pm A z

pmC

κ

Figure 4. Graphical representation of the non-coplanarity criterion. The curved lines represent the slice contours and are included to facilitate the understanding of marker positioning and the non-coplanarity criterion. A plane P is defined by the centroid of the marker distribution, κ, and two of the potential principal-marker centroids, pmA and pmB . If the distance of the third potential principal-marker centroid, pmC , to the plane P is close to zero the three-dimensional coordinate system defined by pmA , pmB and pmC is non-unique. However, if pmA is replaced by pmD , coplanarity is avoided and unique identification of the remaining non-principal markers, m1 and m2 , would be possible.

register the mobile to the reference marker sets. The transformation matrix thus derived was used to register the mobile to the reference 131 I scans using trilinear interpolation. 2.5. External markers The external markers consisted of Perspex discs 12 mm in diameter and 9 mm deep, containing a spherical cavity of 5 mm diameter which was stopped with a small nylon screw (figure 5). This cavity was filled with 99m Tc (∼400 MBq cm−3 ) for SPECT marker acquisition for all the patient studies described in this paper. The patient’s skin was marked, using ink, prior to the administration of the therapeutic activity, to allow accurate and reproducible placement of the external markers. The markers were taped on the patient’s skin before each scan so that they were stable for the duration of the acquisition. The number and the distribution of markers were chosen to define a volume such that all regions for which dose distributions had to be generated were included. Ideally, the least-squares fitting algorithm requires only three markers. However, for the transformation to be reliable and ‘representative’ of the whole patient volume, more than three markers were used. In practice, at our centre, eight to ten markers with sufficient spacing are used to ensure that the whole patient volume is covered. 2.6. SPECT acquisition Two sequential SPECT scans (64 views, acquisition matrix 64 × 64), one for the 99m Tc (marker scan) and one for the 131 I (therapy scan), were acquired on a General Electric STARCAM camera (General Electric Medical Systems, Milwaukee, WI) with a high-energy, high-resolution collimator. The 140 keV (20% window) peak of the 99m Tc was used for the marker scan and the 364 keV (12% window) peak of the 131 I for the therapy scan. The 99m Tc scan was acquired prior to the 131 I. A concentration of ∼400 MBq cm−3 allowed an acquisition time of 3 s per view for the 99m Tc scan, whilst the 131 I scan was acquired with 25 s per view. Transaxial images were reconstructed into 3D matrices (64 × 64 × 64, voxel dimensions 6.4 mm × 6.4 mm × 6.4 mm) by filtered backprojection using a Ramp–Hanning filter with a

2092

P Papavasileiou et al

Figure 5. Picture of an external fiducial marker.

cut-off frequency of 0.7 cm−1 . The 99m Tc marker concentration of ∼400 MBq cm−3 provided sufficient marker-to-background contrast for the segmentation process while avoiding pixel overflow in the reconstructed images. Following acquisition and reconstruction, the SPECT images were transferred to the RMDP dosimetry package running on a SUN-SPARC 450 Enterprise workstation (Sun Microsystems Inc., Palo Alto, CA). All the marker SPECT studies were interpolated to 256×256×64 matrices (voxel dimensions of 1.6 mm×1.6 mm×6.4 mm) to match the digital resolution of the diagnostic CT and/or MR studies employed to enable tumour localization. 2.7. Simulated study In order to validate the automated registration process, a simulated study was generated. A Tc marker scan, acquired as part of a patient therapy study, was used to generate nine ‘copies’ of the original scan employing a range of transformation parameters (0–25.6 mm and 0–15◦ for translation and rotation respectively). The transformation parameters were not strictly random but were chosen to cover the values expected in routine clinical studies. The transformation, based on rigid-body trilinear interpolation, was applied to the whole marker scan. Each of the 10 scans of the simulated study was registered to the remaining nine. Theoretically, the number of possible pairs within a set is given by the formula n1 × (n1 − 1), where n1 is the size of the set, thus the simulated study should result in 90 registration pairs. However, during the marker matching process, the roles of reference and mobile scans are reversed in order to make the algorithm robust to the order of the registration. As a result, the simulated study consisted of 45 instead of 90 registration pairs. 99m

2.8. Patient studies Three different patient studies (A, B and C) were used to test the automated marker-based image registration algorithm described in this paper. All three patients received therapeutic activities of iodine-131 metaiodobenzylguanidine (131 I-MIBG) for treatment of neuroblastoma. The three individual patient marker studies consisted of four, three and five sequential SPECT scans, thus resulting in six, three and ten registration pairs respectively (see section 2.7 for

Automated SPECT image registration using markers

2093

Table 1. Details of patient studies. Patient study

Day following administration of therapeutic amount of radioactivity 1st

A B C

2nd

3rd

• •

4th

5th

6th

7th

•

• • •

• • •

• •

Number of registration pairs

8th

•

6 3 10

how the number of registration pairs was determined). Data on the relative timing of each scan and the number of registration pairs of each study are summarized in table 1. Ten markers were used for each scan, thus 120 markers were to be identified in all of the patient studies (12 marker scans in the three studies). The preprocessing step described in section 2.1 was only necessary for the first scan of patient study A. 3. Results 3.1. Simulated study The automated registration process resulted in all 10 markers being segmented and matched on each registration pair. In order to test the accuracy of the registration process, the 3D residual error was computed. The 3D residual error was determined by measuring the root mean square (rms) distance between coordinates of pairs of corresponding marker centroids following image transformation. The mean 3D residual error (±1 SD) for the 45 registration pairs was 1.7 ± 0.1 mm. 3.2. Patient studies The automated localization process identified all 10 markers in all marker scan combinations. The range of values for the upper and lower thresholds that were required to segment the markers in the individual scans was 40–70% and 6–30% of the initial maximum value respectively. The above range is a reflection of the fact that the threshold for marker segmentation might vary significantly both between different markers and between markers of the same study. Lower threshold values above 20% meant that a crossover between adjacent markers was identified whereas upper and lower threshold values of 40% and 6% respectively implied varying levels of marker 99m Tc radioactive concentration in the markers. Ideally, only the lower threshold should be required to segment the markers since there is no other functional information. The inclusion of the upper threshold constrains the search for potential marker voxels in the neighbourhood of potential markers, thus making the algorithm more efficient and robust. The matching process was evaluated by the number of identified marker pairs in all different scan combinations within each study. For study A, the matching process resulted in all marker pairs being matched on all study combinations. For study B all markers were matched in all but one study combination (nine markers matched) whereas in study C, the algorithm failed to match all ten markers in two combinations (nine markers matched in these two combinations). A slight change in the relative position of the markers might result in a change in the distances of the marker centroids to the marker distribution centroid and this could explain the inability of the algorithm to match all ten markers. The change in relative position was a result of either mispositioning of the marker and/or that the marker was not taped properly and therefore ‘moved’ during the acquisition.

P Papavasileiou et al

Residual Error - AUT (mm)

2094

4 3.5 3 2.5 2 1.5 1 0.5 0 0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

Residual Error - MAN (mm)

Figure 6. Comparison of 3D residual errors resulting from semiautomated (MAN) and automated (AUT) marker-based registration methods of the three patient studies (19 registration pairs). The mean 3D residual errors are 3.50±0.35 mm (mean±SD) for both the semiautomated and automated methods. The line of identity is marked.

The patient studies were registered using both the automated and the semiautomated approach (Flux 1995) and the corresponding 3D residual errors computed. Figure 6 shows the comparison of the residual error values calculated using the semiautomated and automated approaches. The mean 3D residual errors were 3.50 ± 0.35 mm (mean ±SD) for both methods. The mean 3D residual error values resulting from the two approaches were compared using the two-sample t-test and no statistically significant difference was found (t = 0.281, p = 0.8). The registration accuracy of 3.5 mm attained with the automated method showed that the Gaussian profile criterion employed to segment the markers and the intensity-weighted centroiding result in accurate registration despite the poor spatial resolution of the imaging system (spatial resolution of ∼20 mm for 131 I SPECT) (Guy 2000). 4. Discussion The automated segmentation algorithm described in this paper relies on the assumption that the Tc SPECT scan only contains information about the 99m Tc markers. However, down-scatter from 131 I may necessitate the subtraction of counts due to scattered radiation, mainly in the early scans. We employ an automated subtraction based on the inherent registration between the marker and therapy scans. Acquisition of the 99m Tc marker scan using a triple-energy window (TEW) method (Takashi et al 1993) could enable subtraction of the scattered 131 I photons contaminating the 99m Tc marker scan. However, our imaging system did not allow TEW acquisition. Implementation of the TEW method would, therefore, require an extra scan, which would significantly prolong the scanning time. Several semiautomated SPECT segmentation approaches have been developed for accurate segmentation, for example of tumour volumes. These include histogram-derived thresholds (Mortelmans et al 1986, Chiu et al 1994, Alaamer et al 1993), fixed and variable thresholds (Erdi et al 1995, Fleming and Alaamer 1998), and two-dimensional and three-dimensional gradient edge-detection methods (Long et al 1992). The marker segmentation process described in this work, even though it employs iterative thresholding, aims at marker centroid localization rather than accurate quantification of the marker volume. Most of the segmentation algorithms in nuclear medicine require information on the specific imaging system (e.g. the

99m

Automated SPECT image registration using markers

2095

point spread function) to obtain accurate segmentation. In contrast, the approach we adopted, exploiting the Gaussian property of profiles through the markers, overcame the necessity for information on the specific system, thus making it more robust and easily applicable to any other system. The accuracy of marker centroid localization has a direct influence on the performance of the matching process since the centroid of the segmented marker set distribution is used for establishing correspondence between markers. Another important factor is that the position of each marker relative to the others might not be preserved between scans irrespective of the accuracy with which the markers are placed on the patient. Respiration and other patient movements are among the reasons for the process failing to match all the segmented markers. If significant marker misplacement has occurred, the entire matching process becomes unstable since the marker distribution centroid will undoubtedly change. The first step of the matching process, which identifies the principal-marker centroids, attempts to determine which markers have preserved their position relative to each other, assuming that only slight misplacements, if any, have taken place. The computed distances between localized marker-centroids have an error associated with them, as discussed above, and as a result thresholds have to be established when comparing theoretically equal distances. A threshold of 15 mm was used for both identification of the three principal-marker centroids and matching of the remaining marker centroids. The value of 15 mm assigned to both thresholds reflects the spatial resolution of SPECT images. In the studies analysed in this work, principal markers were identified on all marker–scan combinations. The tolerance of the matching process to rigid transformations, e.g. rotation around the z-axis, was tested on the simulated study by generating severe misalignment between some of the simulated scans. On the 99m Tc scan used to generate the simulated study several markers were on the same axial position which might result in a change in the order with which they were segmented in case of significant rotation. The residual error of 1.7 mm in the simulated study showed that the registration process was accurate in ideal cases. The residual error should be close to zero since the simulated study consisted of multiple copies of the same scan and since the copies were generated using rigid-body transformation (no changes in the relative position between markers). The only cause of error was the two steps of trilinear interpolation (one step transforming the initial marker scan and a second registering the transformed to the initial). The tolerance of the matching process to a certain level of rotation was dependent on the distance between the markers and this problem is outside the scope of this work. The automated registration process described in this paper has been shown to result in very accurate registration with a residual error of 3.5 mm. The transformation parameters calculated were then applied to register the corresponding 131 I-SPECT therapy scans to enable generation of 3D dose distributions. The underlying assumption in every marker-based registration technique is that the position of the skin markers to internal organs is kept constant between scans. As mentioned above, respiration and patient movement might affect that relationship thus affecting the reliability and usefulness of the registration process. However, the difficulty in voxel-based registration because of changes in radioactivity distributions between subsequent 131 I-SPECT scans and the lack of anatomical information renders markerbased registration necessary. The registration process described in this paper can be applied to other clinical applications involving both intramodality (e.g. registration of sequential cardiac image frames during dynamic PET studies) and intermodality registration (e.g. registration of CT/MRI to SPECT scans in radionuclide therapy to enable delineation of tumour sites). The segmentation subprocess is applicable in all cases where the Gaussian property of profiles across markers is valid. The marker-matching algorithm is a generic one, meaning that it can be applied to

2096

P Papavasileiou et al

all landmark-based registration methods (i.e. those using anatomical landmarks and/or skin markers). 5. Conclusions An automated marker-based image registration technique has been developed and validated for radionuclide therapy. Iterative, adaptive thresholding combined with a priori information on the markers is employed to localize and segment the markers. Marker matching is carried out relying on 3D geometry-based criteria with the underlying assumption that the distances and angles are preserved between the two scans to be registered. Automated marker matching was shown to enable identification of severe mispositioning of markers between scans, thus underlying the importance of a criteria-based automated segmentation and marker-matching process. The technique has been applied to patient studies resulting in a 3D residual registration error of 3.5 mm. Acknowledgments We thank Professor Robert J Ott for his useful comments on the manuscript. This work was funded by the Association of International Cancer Research, the Medical Research Council and the Royal Marsden NHS Trust Charitable Funds. References Alpert N M, Bradshaw J F, Kennedy D and Correia J A 1990 The principal axes transformation—a method for image registration J. Nucl. Med. 31 1717–22 Alaamer A S, Fleming J S and Perring S 1993 Evaluation of a technique for the quantification of radioactivity and volume of SPECT Nucl. Med. Commun. 14 1061–70 Armitage P and Berry G 1994 Statistical Methods in Medical Research 3rd edn (Oxford: Blackwell Scientific) Arun K S, Huang T S and Blostein S D 1987 Least-squares fitting of two 3-D point sets IEEE Trans. Pattern Anal. Machine. Intell. 9 698–700 Bacharach S L, Douglas M A, Carson R E, Kalkowski P J, Freedman N M T, Perrone-Filardi P and Bonow R O 1993 Three-dimensional registration of cardiac positron emission tomography attenuation scans J. Nucl. Med. 34 311–21 Boesecke R, Bruckner T and Ende G 1990 Landmark based correlation of medical images Phys. Med. Biol. 35 121–6 Chiu N-T, Sun Y-N, Chou Y-L, Lin P-W and Yao W-J 1994 Measurement of organ volume by single photon emission computed tomography Nucl. Med. Commun. 15 871–7 Dey D, Slomka P J, Hahn L J and Kloiber R 1999 Automatic three-dimensional multimodality registration using radionuclide transmission CT attenuation maps: a phantom study J. Nucl. Med. 40 448–55 Eberl S, Kanno I, Fulton R R, Ryan A, Hutton B F and Fulham M J 1996 Automated interstudy image registration technique for SPECT and PET J. Nucl. Med. 37 137–45 Erdi Y E, Wessels B W, Loew M H and Erdi A K 1995 Threshold estimation in single photon emission computed tomography and planar imaging for clinical radioimmunotherapy Cancer Res. 55 5823s–5826s Fleming J S and Alaamer A S 1998 A rule based method for context sensitive threshold segmentation in SPECT using simulation Phys. Med. Biol. 43 2309–23 Flux G D 1995 Multimodality image registration and its application to the dosimetry of intralesional radionuclide therapy PhD Thesis University of London Flux G D, Prior M, Chittenden S J, Guy M J, Cody K J and Flower M A 1999 A study of dosimetry errors in I-131 mIBG radionuclide therapy [abstract] J. Nucl. Med. 40 (suppl) 983 Flux G D, Webb S, Ott R J, Chittenden S J and Thomas 1997 Three-dimensional dosimetry for intralesional radionuclide therapy using mathematical modelling and multimodality imaging J. Nucl. Med. 38 1059–66 Guy M J 2000 The application of quantitative single photon emission tomography to targeted radionuclide therapy PhD Thesis University of London Long D T, King M A and Sheehan J 1992 Comparative evaluation of image segmentation methods for volume quantification in SPECT Med. Phys. 19 483–9

Automated SPECT image registration using markers

2097

Mortelmans L, Nuyts J, van Pamel G, van den Maegdenbergh, de Roo M and Suetens P 1986 A new thresholding method for volume determination by SPECT Eur. J. Nucl. Med. 12 284–90 Pelizzari C A, Chen G T Y, Spelbring D R, Weichselbaum R R and Chen C-T 1989 Accurate three-dimensional registration of CT, PET and/or MR images of the brain J. Comput. Assist. Tomogr. 13 20–6 Scott A M et al 1995 Image registration of SPECT and CT images using an external fiduciary band and threedimensional surface fitting in metastatic thyroid cancer J. Nucl. Med. 36 100–3 Takashi I, Ogawa K, Motomura N, Kubo A and Hashimoto S 1993 Compton scatter compensation using the tripleenergy window model for single- and dual-isotope SPECT J. Nucl. Med. 34 2216–21 Woods R P, Cherry S R and Mazziota J C 1992 Rapid automated algorithm for aligning and reslicing PET images J. Comput. Assist. Tomogr. 16 620–33 Woods R P, Mazziota J C and Cherry S R 1993 MRI-PET registration with automated algorithm J. Comput. Assist. Tomogr. 17 536–46 Yu J N, Fahey F H, Gage H D et al 1995 Intermodality, retrospective image registration in the thorax J. Nucl. Med. 36 2333–8