Validation of Simplified Dosimetry Approaches in 89 Zr-pet/ct: the Use of Manual versus Semi-automatic Delineation Methods to Estimate Organ Absorbed Doses

PURPOSE
Increasing interest in immuno-positron emission tomography (PET) studies requires development of dosimetry methods which will provide accurate estimations of organ absorbed doses. The purpose of this study is to develop and validate simplified dosimetry approaches for (89)Zirconium-PET (Zr-PET)/computed tomography (CT) studies.


METHODS
Five patients with advanced colorectal cancer received 37.1 ± 0.9 MBq (89)Zr-cetuximab within 2 h after administration of a therapeutic dose of 500 mg m(-2) cetuximab. PET/CT scans were obtained 1, 24, 48, 94, and 144 h post injection. Volumes of interest (VOIs) were manually delineated in lungs, liver, spleen, and kidneys for all scans, providing a reference VOI set. Simplified manual VOIs were drawn independently on CT scans using larger voxel sizes. The transformation of VOIs based on rigid and/or nonrigid registrations of the first CT scan (CT1) onto all successive CT scans was also investigated. The transformation matrix obtained from each registration was applied to the manual VOIs of CT₁ to obtain VOIs for the successive scans. Dice similarity coefficient (DSC) and Hausdorff distance were used to assess the performance of the registrations. Organ total activity, organ absorbed dose, and effective dose were calculated for all methods.


RESULTS
Semi-automatic delineation based on nonrigid registration showed excellent agreement for lungs and liver (DSC: 0.90 ± 0.04; 0.81 ± 0.06) and good agreement for spleen and kidneys (DSC: 0.71 ± 0.07; 0.66 ± 0.08). Hausdorff distance ranged from 13 to 16 mm depending on the organ. Simplified manual delineation methods, in liver and lungs, performed similarly to semi-automatic delineation methods. For kidneys and spleen, however, poorer accuracy in total activity and absorbed dose was observed, as the voxel size increased. Organ absorbed dose and total activity based on nonrigid registration were within 10%. The effective dose was within ±3% for all VOI delineation methods.


CONCLUSIONS
A fast, semi-automatic, and accurate delineation method based on nonrigid registration was developed for determination of organ absorbed and effective dose in (89)Zr-PET/CT studies which may also be applied to other long-lived radionuclide PET/CT studies.


Introduction
Positron Emission Tomography (PET) is a valuable, non-invasive tool for diagnosing, staging and monitoring response to therapy in cancer patients.Recently, 89 Zirconium-PET ( 89 Zr-PET) has also shown its ability to provide the quantitative data necessary to derive an estimate of organ and tumour dose (1), allowing for an estimate of the biodistribution of a non-89 Zr labelled therapeutic analog (2).Images are subsequently used to estimate absorbed dose to critical organs, prior to targeted immunotherapy (3).This may eventually be used to improve patient selection and radionuclide dose prescription.ImmunoPET is based on monoclonal antibodies (mAbs) that are characterized by relatively slow pharmacokinetics and therefore labelled with a long-lived radionuclide.Consequently, multiple PET/Computed tomography (CT) scans per patient are required in order to quantify tumour uptake and targeting over time as well as to perform organ and tumour dosimetry.PET based dosimetry usually starts with manual delineation of organs of interest in the CT scan, followed by mapping these delineated volumes onto the corresponding PET scans.The area under the time-activity curve, extracted for each organ from a set of PET/CT scans, is then used for calculation of the cumulated activity (or residence time) (4).Cumulated activity can be translated into absorbed dose through the use of dose conversions factors (S-values) (5).Although manual delineation of organs is a simple and accurate method to derive organ and tumour dose estimates, it is time consuming and labour intensive.To this end, simplified manual volume of interest (VOI) delineation methods (using various voxel sizes and/or reducing the dimensions of the PET/CT images) may be chosen to obtain a coarser data representation enabling faster delineation.Note that a similar increase in efficiency might be achieved when radiologists are instructed to draw regions with a larger brush size.In addition to manual VOI delineation, rigid or non-rigid registration may enable re-use of the VOI from the first CT scan to all successive CT scans of the same patient, potentially serving as an alternative to manual VOI delineation in each successive scan for organ dose estimation.In a previous study (6) investigating the impact of various image registration strategies on standardized uptake value (SUV) and metabolic volume test-retest variability, nonrigid PET/CT image registration performed equally well as delineating VOI on both scans separately, and with smaller absolute test-retest volume estimates.A recently developed method for automated analysis of small animal PET studies using nonrigid image registration between CT images of the actual mouse and a digital mouse atlas, showed similar performance as manual delineation, with relative errors below 10% in estimated normalized mean activity in most of the regions (7).The purpose of this study was to develop and evaluate simplified dosimetry approaches that would guarantee accurate localization of delineated organ VOIs and subsequent quantification of organ absorbed doses.

Imaging protocol
Five patients with advanced colorectal cancer were included.Patients received 37.1 ± 0.9 MBq 89 Zr-cetuximab within 2 hours after administration of a therapeutic dose of 500 mg⋅m −2 unlabelled cetuximab.Per patient, five PET/CT scans were acquired on a Gemini TF-64 PET/CT scanner (Philips Healthcare, Cleveland, OH, USA) at 1, 24, 48, 96 and 144 h post injection.Data were normalized, corrected for decay, randoms, dead time, scatter and attenuation, and reconstructed using a time-of-flight list-mode orderedsubsets expectation maximization reconstruction algorithm using an image matrix size of 144 × 144 and a voxel size of 4 × 4 × 4 mm.In addition, for each time point, a 50 mAs low dose CT scan was acquired for attenuation correction purposes.The corresponding CT images were reconstructed with an image matrix size of 512 × 512 and a voxel size of 1.17 × 1.17 × 5.00 mm.The study was approved by the Medical Ethics Review Committee of the VU University Medical Centre and informed consent was obtained from each patient prior to inclusion in the study.

VOI delineation methods
VOIs were delineated based on various manual and semi-automatic approaches (Table 5.1):

Manual delineation
Volumes of interest in lungs, liver, spleen and kidneys were drawn independently on all five CT scans of each patient.Computed tomography scans were rebinned (using software developed in-house) with trilinear interpolation using the following voxel sizes: 4 × 4 × 4 mm, 8 × 8 × 8 mm, 4 × 16 × 4 mm, 16 × 16 × 16 mm and 2D (summation of coronal slices of the 3D dataset which is similar to a mean intensity projection in appearance, see Figure 5.1) and VOI delineation was repeated using these lower resolution CT data sets.Subsequently, organ activity was determined by mapping the CT VOI onto the correspondingly resampled PET scans.PET/CT misalignment was compensated (to a certain extent) by visually inspecting the VOIs onto the PET image and by adjusting the VOIs where needed (mainly in the upper part of the liver and the lower part of the right lung).Manual delineation of the organ VOIs using a resampled CT with 4 × 4 × 4 mm voxel sizes was used as the reference method, as it matches the matrix and voxel size of the PET image.

Semi-automatic delineation
CT images were registered using rigid, affine and nonrigid registration strategies implemented in the Elastix software package (8) (University Medical Centre Utrecht, The Netherlands, free of charge).Prior to image registration, CT scans were resampled to PET resolution to increase computational performance and to avoid issues with computer memory.The first CT scan Rigid prior to nonrigid registration: a rigid registration was applied first in order to facilitate the registration process and improve the performance of the nonrigid registration.
Affine prior to nonrigid registration: the rationale for this registration was the same as for the previous, except that first an affine transformation was used instead of a rigid transformation.

Semi-automatic delineation and use of larger voxel size
In addition, CT scans rebinned to a 16 × 16 × 16 mm voxel size were used to manually delineate VOIs which were nonrigidly registered on all successive CT scans.This method enabled the fastest extraction of VOIs and will be referred to as nonrigid [16 × 16 × 16 mm].

Image registration
Two similarity criteria (or cost functions) were used for the registration of two images: one that maximized mutual information and another that maximized normalized cross correlation.The optimization method that maximized the similarity criteria, by varying either the Euler (rigid), affine or B-spline (nonrigid) transformation model parameters, was the adaptive stochastic gradient descent (which combines accuracy with computational efficiency) (9).Three resolution levels were applied for rigid, affine and nonrigid registration by smoothing the fixed image in each dimension with downsampling factors of ×4, ×2 and ×1.At each resolution level, 32 grey level bins and 2,000 spatial samples were used to compute mutual information.In total 500 iterations were used.The rationale of using a CT to CT registration was the low count rate in late 89 Zr-PET scans, which may affect a PET to PET registration.Target (organ) to background ratios in 89 Zr images rapidly decreased over time, especially for kidneys and spleen, resulting in inconsistent registrations between PET scans acquired at different time points.In addition, the biodistribution also changes over time, with for example moderate uptake in kidneys and high uptake in liver in the first PET scan, and with very low uptake in kidneys and moderate uptake in liver at late times.This change in biodistribution is likely to affect the PET to PET registration as well.The registration process per subject (four CT to CT registrations and four VOI transformations) required less than 15 min on a 32-bit desktop PC using an Intel Core 2 Duo 2.80 GHz CPU with 3.2 GB of RAM (in the case of nonrigid registrations).Semi-automatic VOI delineation was approximately 7 times faster than the reference method based on manual VOI delineation.

Evaluation measures
For each patient and scan a set of registered VOIs was obtained on the basis of CT 1 VOIs.The performance of the image registration strategy was assessed by means of Dice similarity coefficient (DSC) (10).This metric computes the volume overlap between the fixed and the registered VOI as: where F and R correspond to the fixed and registered VOI, respectively.The degree of agreement is typically classified into five categories (7), as follows: < 0.20 → poor agreement; 0.20 -0.40 → fair agreement; 0.40 -0.60 → moderate agreement; 0.60 -0.80 → good agreement; 0.80 -1.00 → excellent agreement.Additionally, the maximum possible DSC was calculated, assuming perfect organ overlap between time points, i.e.DSC is then only affected by differences in VOI volume.In this way DSC provides an estimate for tumor delineation reproducibility (without effects of translation/movement).The second evaluation measure was the Hausdorff distance (11) which is the maximum distance one would need to move the boundaries of the fixed region to completely overlap with the registered region.It is given by: where r and f are points in the registered and fixed region, respectively, and d(r,f ) is the Euclidean distance between r and f.However, as HD R→F may not be symmetric and therefore not equal to HD F →R , a generalized equation should be used, which is defined as follows:

Organ dosimetry
The total activity for each visible organ (liver, lungs, kidneys and spleen) was determined using software developed in-house.The normalized cumulated activity was calculated as the area under the curve of the organ time-activity data determined by the trapezoidal rule and assuming physical decay only after the last measurement, divided by the total injected activity.Finally, the OLINDA/EXM software (Version 1.1, ©2007 Vanderbilt University, Nashville, TN, USA) was used for calculation of organ absorbed doses (5).Average percentage changes in both total organ activity and absorbed dose between all VOI delineation methods and the (reference) manual VOI delineation method was calculated for each patient.Individual discrepancies in total organ activity are presented per scan.

Statistical analysis
The results found with any of the tested VOI delineation methods were tested for normality by means of the Shapiro-Wilk test.Similarity of homogeneity of variance between the different results was checked using Levene's test.The 11 VOI delineation methods were compared with each other using analysis of variance (ANOVA).A post-hoc Tukey-Honestly-Significant-Difference (Tukey-HSD) was performed in case of a statistically significant ANOVA test.The type I error cut-off was set to 0.05.

Results
In Fig. 5.2 an example of multiple VOIs obtained following nonrigid image registration of two CT scans is shown.It should be noted that, as a similarity criterion, mutual information performed slightly better (for spleen and kidneys) than normalized cross correlation; however, this difference was not statistically significant (nonrigid; DSC spleen : 0.71 ± 0.07, DSC kidneys : 0.66 ± 0.08 and DSC spleen : 0.68 ± 0.08, DSC kidneys : 0.62 ± 0.11, respectively).Therefore, only mutual information was used for further evaluation of image registration strategies.

Impact of manual/semi-automatic delineation method on derived organ activity
Average percentage changes in total between (reference) manually delineated VOIs and all other methods (Table 5.1) is illustrated in Fig. 5.4.Across all organs investigated, average percentage change in total activity was within ±10% for semi-automatic delineation methods (except for non- rigid [16 × 16 × 16 mm]: -3 to +30% depending on organ) and from -5 to +94% for manual delineation methods depending on organ and voxel size.
In kidneys and spleen, semi-automatic delineation methods showed individual discrepancies ranging from -30 to +80%, whereas in manual delineation methods individual discrepancies ranged from -20 to +270% depending on used voxel size.The large upper limit can be explained by the use of the 2D approach using a summation of all coronal slices.In this case tissues located behind and in the front of the organ of interest will contribute to the estimation of the total organ activity.This effect is particularly pronounced in small organs as the relative contribution from voxels not related to the organ of interest would be higher for small organs than for larger organs.In lungs and liver, however, semi-automatic delineation, using either rigid or deformable registration, seemed to perform similarly (within ±15%) as manual delineation, regardless of voxel size due to these structures' larger volume and relatively simple physical shapes.

Impact of manual/semi-automatic delineation method on calculated absorbed dose
As illustrated in Fig. 5.5, semi-automatic delineation methods showed an average percentage change in absorbed dose within ±10% across all organs (except for nonrigid [16 x 16 x 16 mm] in the spleen: +17%), whereas changes for simplified manual delineations were as high as +35% depending on voxel size and organ.Standard deviation for the manual delineation methods was within ±10% for lungs and liver, whereas for smaller organs, such as kidneys and spleen, it was within ±25% (with the manual 2D method showing the largest discrepancies).Semi-automatic delineation methods based on rigid and nonrigid registration showed similar performance with the smallest standard deviation in average percentage change in absorbed dose (±5%) for all organs.The underestimation on average (<10%) in liver absorbed dose while using semi-automatic methods may be explained by the transformed VOI liver that could be slightly overprojected onto the right lung region.Incorporation of tissue with relatively lower activity concentration than that of the liver will then lead to underestimation of total activity and absorbed dose in the liver.Average percentage change in effective dose, as the weighted sum of organ contributions, was well within ±2% for all delineation methods and organs and standard deviations were at most ±2% (Fig. 5.6).In Table 5.3 organ absorbed dose and effective dose estimates for all 11 manual and semi-automatic delineations methods are summarized.

Discussion
Use of immunoPET with 89 Zr labelled mAbs is an effective tool for investigating mAb therapy and dose optimization.It allows measurement of the uptake in critical organs and the variability in patient kinetics.For that purpose tools need to be developed that allow accurate quantification and absorbed dose estimation.In this study, different image processing and analysis approaches using manual VOI delineation or spatially transformed VOIs (on the basis of CT scans) were developed and their impact on timeactivity data and dose estimates were evaluated.Based on average DSC, performance of semi-automatic VOI delineation methods for nonrigid registration showed excellent agreement in liver and lungs and good agreement in spleen and kidneys.In the case of rigid or affine registrations, good agreement in liver, lungs and spleen and moderate agreement in kidneys were obtained.Significant differences in DSC between nonrigid and both rigid and affine registrations were observed in lungs, spleen and kidneys.Overall, deformable registration strategies yielded the highest average DSC for all organs.Although a rigid or affine registration prior to nonrigid registration would be expected to improve DSC (6), this was not the case in the present study, probably due to absence of significant changes in posture and/or organ geometry of each patient during the time-course of the five PET/CT scans.This suggests that any technique using deformable CT-CT registration will be comparable.When comparing nonrigid to rigid and affine registrations, however, nonrigid registration seemed to perform slightly better for lungs, as it produced the smallest individual discrepancies (-2% / +32% for nonrigid, -18% / +55% for rigid and -20% / +53% for affine).This is further illustrated by the average Hausdorff distance that differed significantly (p < 0.05) between nonrigid and both rigid and affine registrations, indicating a significantly larger displacement of VOI for rigid and affine compared to nonrigid registrations.Maximum DSC in the reference manual method (assuming perfect organ registration or alignment between time points) was >0.90 for all organs.For liver and lungs, the semi-automatic methods (using non rigid registration) showed comparable DSCs to the maximum ones, whereas for spleen and kidneys those differences were substantially larger.Overall, for all organs, both rigid and nonrigid registrations showed smaller average per-  xv centage changes in total activity compared to simplified manual delineation methods.Manual delineation methods, in the case of liver and lungs, seemed to perform equally well as semi-automatic delineation methods.In the case of kidneys and spleen, however, larger individual discrepancies in average percentage change in total activity were observed for increasing voxel size.It is clear that small structures are more prone to errors in measured total activity when increasing voxel size due to the increased partial volume effects near the structure boundaries.It is important to note that the effective dose can be estimated accurately (within ±3%) regardless of (manual or semi-automatic) delineation method used.The use of a semi-automatic delineation method based on nonrigid registration showed the highest accuracy.Because effective dose involves the summation of weighted dose values from multiple structures, it is less susceptible to the delineation method (and its associated errors) used.Analysis of organ absorbed doses showed that all semi-automatic VOI delineation methods produced average percentage changes smaller than when using simplified manual VOI delineation methods.In terms of absorbed dose, similar accuracy and variability for rigid and nonrigid registrations were observed, suggesting that rigid registration might be preferable, as it is faster and less likely to cause artefacts (after visual inspection of all images no artifacts were observed).Nonrigid registration, however, outperforms rigid registration not only with respect to DSC, but also with respect to individual discrepancies in total activity for lungs and kidneys, making it a more robust choice for VOI delineation and more accurate with regards to calculation of total organ activity.Recent findings by Van Velden et al. (6) also suggest that nonrigid image registration might serve as a good surrogate method for accurately delineating VOIs in test-retest 18 F-FDG studies.In a similar study by Jackson et al. (12) image registration of three CT volumes to a first day CT dataset required 4 h per case on a desktop computer, whereas in the current study, coregistration of CT 1 to four successive CT scans and transformation of VOI 1 using the transformation matrices from the first step required at most 15 min.Unlike the study of Jackson et al. (12) in which full resolution CT images were used for registration, in the present study CT images were downsampled to match the PET resolution prior to registration.Although, in theory, downsampling may affect accuracy of CT registrations, it has been shown that this is not the case for rigid registration (13), while at the same time increasing computational speed.xvi

Conclusion
A fast and accurate method for semi-automatic delineation of organs in successive scans was developed enabling time-efficient estimation of organ absorbed doses in 89 Zr PET/CT studies.The method is generally applicable to PET/CT studies with other (long-lived) radionuclides as the method uses the CT data for semi-automated VOI generation.Moreover, VOIs derived semi-automatically (based on nonrigid CT-CT registrations) showed good agreement with the reference manual VOI in terms of total organ activity, organ absorbed dose and effective dose.

Figure 5 . 1 .
Figure 5.1.Effect on PET image resolution of using different binning intervals; from left to right 4 × 4 × 4 mm, 8 × 8 × 8 mm, 4 × 16 × 4 mm, 16 × 16 × 16 mm and 2D.For the 2D projection images only the left kidney was delineated and total volume for both kidneys was based on multiplying the left kidney data by a factor of 2.0 and this was subsequently used to calculate total activity in kidneys)

Figure 5 . 4 .
Figure 5.4.Box plots showing the percentage change in total activity for various manual and semi-automatic delineation methods in (A) liver, (B) lungs, (C) kidneys and (D) spleen as compared with the reference manual VOI delineation.The latter analysis was based on the repeat scans only (the first scan was not included).

Figure 5 . 5 .
Figure 5.5.Percentage change in absorbed dose for various manual and semi-automatic methods in (A) liver, (B) lungs, (C) kidneys and (D) spleen as compared with the reference manual VOI delineation.Error bars illustrate standard deviation.

Figure 5 . 6 .
Figure 5.6.Percentage change in effective dose for various manual and semi-automatic methods using the reference manual VOI delineation as the basis for comparison.Error bars illustrate standard deviation.

Table 5 . 1 .
List of manual and semi-automatic delineation methods CT 1→2 , CT 1→3 , etc) was applied to the reference (manually delineated) VOIs of CT 1 .The spatially transformed VOIs were then projected onto the corresponding PET images in order to obtain mean organ activity.The registration strategies implemented in this study were:Rigid registration: allows for translation and rotation.Affine registration: allows for translation, rotation, scaling and shearing.Nonrigid registration: the high number of degrees of freedom (typically from 100 to 1000) allows for local elastic image deformations.

Table 5 . 3 .
Mean organ absorbed doses and effective dose for various VOI delineation methods. *