Improved dosimetry for targeted radionuclide therapy using nonrigid registration on sequential SPECT images Edwin C. I. Ao Biomedical Imaging Laboratory, Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Macau SAR, China
Nien-Yun Wu Department of Biomedical Imaging and Radiological Sciences, National Yang Ming University, Taipei 112, Taiwan and Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei 112, Taiwan
Shyh-Jen Wang Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei 112, Taiwan
Na Song Department of Nuclear Medicine, Albert Einstein College of Medicine, Yeshiva University, Bronx, New York 10461
Greta S. P. Moka) Biomedical Imaging Laboratory, Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Macau SAR, China
(Received 26 June 2014; revised 26 November 2014; accepted for publication 3 January 2015; published 29 January 2015) Purpose: Voxel-level and patient-specific 3D dosimetry for targeted radionuclide therapy (TRT) typically involves serial nuclear medicine scans. Misalignment of the images can result in reduced dosimetric accuracy. Since the scans are typically performed over a period of several days, there will be patient movement between scans and possible nonrigid organ deformation. This work aims to implement and evaluate the use of nonrigid image registration on a series of quantitative SPECT (QSPECT) images for TRT dosimetry. Methods: A population of 4D extended cardiac torso phantoms, comprised of three In-111 Zevalin biokinetics models and three anatomical variations, was generated based on the patient data. The authors simulated QSPECT acquisitions at five time points. At each time point, individual organ and whole-body deformation between scans were modeled by translating/rotating organs and the body up to 5◦/voxels, keeping ≤5% difference in organ volume. An analytical projector was used to generate realistic noisy projections for a medium energy general purpose collimator. Projections were reconstructed using OS-EM algorithm with geometric collimator detector response, attenuation, and scatter corrections. The QSPECT images were registered using organ-based nonrigid image registration method. The cumulative activity in each voxel was obtained by integrating the activity over time. Dose distribution images were obtained by convolving the cumulative activity images with a Y -90 dose kernel. Dose volume histograms (DVHs) for organs-of-interest were analyzed. Results: After nonrigid registration, the mean differences in organ doses compared to the case without misalignment were improved from (−15.50 ± 5.59)% to (−2.12 ± 1.05)% and (−7.28 ± 2.30)% to (−0.23 ± 0.71)% for the spleen and liver, respectively. For all organs, the cumulative DVHs showed improvement after nonrigid registration and the normalized absolute error of differential DVHs ranged from 6.79% to 22.70% for liver and 26.00% to 39.70% for spleen with different segmentation methods. Conclusions: These results demonstrated that nonrigid registration of sequential QSPECT images is feasible for TRT and improves the accuracy of 3D dosimetry. C 2015 American Association of Physicists in Medicine. [http://dx.doi.org/10.1118/1.4906242] Key words: dosimetry, nonrigid registration, quantitative SPECT, targeted radionuclide therapy
1. INTRODUCTION In targeted radionuclide therapy (TRT), accurate estimates of organ dose and 3D dose distribution are important in delivering maximum tumor doses and reducing normal organ toxicity for optimal dosing regimens. The 3D dose distribution depends on various factors including patient-specific pharmacokinetics 1060
Med. Phys. 42 (2), February 2015
and anatomy. The dosimetric planning based on the 3D dose distribution can potentially improve the treatment efficacy as compared to the standardized regimen that was only based on the patient weight or body surface area.1–3 The 3D patient-specific dosimetry requires an estimate of the voxel-based time-integrated activities which are typically obtained by quantitative nuclear medicine imaging at multiple
0094-2405/2015/42(2)/1060/11/$30.00
© 2015 Am. Assoc. Phys. Med.
1060
1061
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
time points, using either conventional planar scintigraphy, SPECT, or PET. However, only SPECT or PET can provide the spatial information needed for voxel-level dosimetry. With appropriate compensation for image degrading factors, many groups have already proven that SPECT can provide accurate activity quantification for TRT.4–8 Previously, a comprehensive quantitative SPECT method called QSPECT was proposed and proven to be accurate for activity quantification and making 3D dosimetry feasible.6,9–11 Once the activity distributions at different time points are obtained, one can convert them to dose-rate images using either a dose point kernel12 or Monte Carlo simulations.13,14 Voxel-by-voxel integration of the dose-rate images over time generates 3D voxel-based dose images. Misregistration error among serial images is an important factor limiting the dosimetric accuracy in TRT.15 Potential sources of misregistration are changes in patient position, organ deformation, and tumor progression/regression between different scans. These errors will lead to misestimation of organ dose distributions, which can in turn lead to a nonaccurate prediction of treatment response. To address the misregistration problem, research groups employed different image registration methods among sequential SPECT images with positive results. Papavasileiou et al.16 developed a 4D rigid image registration scheme. They generated the temporal information by fitting the maximum count of the tumor at each time point to a polynomial function. Then, the temporal information and
the voxel-based similarity criteria were combined for simultaneous SPECT registration. A uniform radial deformation was applied by Dewaraja et al.17 to register the tumor on serial quantitative SPECT images. Sjogreen-Gleisner et al.18 registered serial CT images and applied the resultant deformation field to the SPECT images for mutual information-based rigid or nonrigid registration. This method avoided using the time-varying activity distributions on SPECT images which may affect the image registration process. However, it may change the SPECT intensity distribution at different time points.19 To the best of our knowledge, most of the existing studies relied on the use of CT images for registration or applied only rigid transformation for quantitative SPECT dosimetry. However, nonrigid deformation in organ/tumor is more likely in the clinical cases20 and rigid registration may not be adequate especially for small organs or lesions. Nonrigid registration on CT images was previously performed to aid the registration of SPECT images. However, direct nonrigid registration on SPECT images is still warranted as CT images are usually not acquired at every time point for TRT. The aim of this work was to improve the TRT dosimetry using organby-organ nonrigid image registration on QSPECT images. The potential improvement of using nonrigid registration as compared to rigid registration on sequential SPECT images was assessed. We also evaluated the role of CT for potential improvement on segmentation accuracy.
F. 1. Sampled coronal slice of the activity maps of nine XCAT phantoms. Medical Physics, Vol. 42, No. 2, February 2015
1061
1062
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1062
2. METHODS AND MATERIALS 2.A. 4D extended cardiac torso (XCAT) phantom population
The 4D XCAT (Ref. 21) phantom with realistic modeling of anatomy and In-111 Zevalin activity distribution of a male patient was used in this study. We modeled a phantom population of three anatomical variations with three respective activity uptake ratios, i.e., a total of nine phantoms (Fig. 1). Uniform activity was simulated in the background and four organs, i.e., heart, liver, kidneys, and spleen,10 while the activity distribution for lungs was not uniform since the activity in the airways was set to zero. The activity variations among the phantoms were obtained from a set of clinical patient data.10 These patients underwent sequential scans at 0–1, 4–6, 24, 72, and 144 h after injection of In-111 Zevalin. The activity in each organ was measured by conventional planar method while QSPECT was performed at 24 h. For each organ, the time-activity curves (TACs) were generated with exponential fitting to calculate the effective half-life and then the activity level was extrapolated back to the injection time. We used these time-varying activity data and the effective halflives of each target organ to simulate SPECT scans acquired at 1, 12, 24, 72, and 144 h postinjection (Table I). The random local deformation of the liver, kidneys, spleen, and stomach was modeled in 1, 12, 72, and 144 h time point to model the organ deformation between scans, while the 24 h time point was set as the reference. These organs were randomly translated and rotated within 5 pixels (11.05 mm) or degrees, respectively, and the change of the total volume of each organ was held within 5% except for the stomach. The boundaries of lungs were defined by the deformation of liver and heart. Whole body rigid transformation within 5 pixels (11.05 mm) or degrees of translation or rotation was also modeled randomly to simulate the whole body interscan movement.20 The nonrigid movement was then obtained by combining the local organ nonrigid and whole body rigid deformation (Fig. 2). Phantoms without any misalignment were used as the gold standard for evaluation. 2.B. Simulation
An analytical projector including attenuation, scatter, and geometric collimator-detector-response (GCDR) was used to generate realistic projections. In both forward and back projection processes, the attenuation was modeled by the XCAT attenuation maps. The scatter was modeled by the effective T I. Organ activity and effective half-life for Phantom #1C (Ref. 10).
Heart Lungs Liver Spleen Kidneys
Activity (MBq/ml) (×10−3)
Effective half-life (h)
1h
12 h
24 h
72 h
144 h
43.5 43.5 85.1 63.1 76.9
7.71 3.61 13.74 27.81 7.78
6.47 3.01 12.28 24.94 6.67
5.35 2.50 11.42 21.60 5.95
2.49 1.18 7.58 13.13 3.99
0.79 0.38 4.26 5.92 2.16
Medical Physics, Vol. 42, No. 2, February 2015
F. 2. Sampled noise-free projections of Phantom #1C at 72 h time point for (a) with no misalignment, (b) with local, and (c) with total nonrigid deformation. The yellow line indicates the position of the horizontal image profiles as shown in (d).
source scatter estimation (ESSE) method,22 which models the scatter and down scatter from the two separate photopeaks into the two windows based on scattering physics by using kernels estimated from Monte Carlo simulations. It models the object dependent, spatially varying scatter response and takes into account the effects of the nonuniform attenuation distribution. The GCDR was computed using an analytic formulation proposed by Metz et al.23 We simulated a GE Discovery VH Hawkeye SPECT/CT system with crystal thickness of 2.54 cm mounted with a medium energy general purpose (MEGP) collimator. The intrinsic resolution of the system was 0.4 cm. Both photopeaks of In-111 (171 and 245 keV) were considered with appropriate abundances (90.2% and 94%, respectively)24 and probability of detection to generate the abundance weighted average energy of 192.6 keV for attenuation modeling and correction in reconstruction.25 The projections had 128 transaxial and 170 axial bins with 128 views and 30 s/view over 360◦ acquisition. In order to model the continuous nature of the activity distribution, projections were generated using isotropic phantoms with voxel size of 0.221 cm. The projections were then collapsed to a bin size of 0.442 cm for reconstruction. A system calibration factor of 1.43 × 10−4 counts s−1 Bq−1 was used to scale the noise-free projection to a clinical count level and then modeled with Poisson noise to obtain realistic noisy projections. Figure 3 shows the noisy projections with local and global deformation for Phantom #1C at different time points. 2.C. Quantitative reconstruction
We reconstructed the data using QSPECT method which utilized a rotation-based projector26 compensated with attenuation, ESSE scatter model, and GCDR in the OS-EM
1063
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1063
F. 3. Sampled noisy projections at different time points with (a) local deformation and (b) global deformation.
algorithm (8 iterations and 16 subsets, i.e., 128 updates).27 A more detailed description of QSPECT could be found in the work of He et al.6 The reconstructed matrix was 128 × 128 × 170. No postfiltering was performed on the reconstructed images. 2.D. Image segmentation
Before registration, each organ-of-interest was segmented out at each time point. There are three kinds of volume-ofinterest (VOI) used in this study: (i) “exact VOI” (E-VOI) where VOIs were mapped out from the phantoms directly with no segmentation errors. (ii) “SPECT VOI” (S-VOI) where VOIs were obtained directly on the QSPECT images. (iii) “CT VOI” (C-VOI) where VOIs were drawn on the attenuation maps that served as the CT images. For both S-VOI and CVOI methods, VOIs were segmented using the semiautomatic region growing segmentation (snake) method28 from the open source program “ITK-SNAP” (Ref. 29) and adjusted based on the interpretation of an experienced operator (Fig. 4).
with 3 and 2 resolution levels were used in affine and B-spline registration, respectively. Normalized mutual information32,33 was used to measure the image similarity between the fixed and “moving” images based on the 64 histograms bins for the 2 registration steps. Optimization of the registration process was performed by the standard nonadaptive Robbins–Monro (RM) algorithm.34 Numerical voxel-by-voxel integration was later performed on the registered images over 5 time points followed by convolution with a 33 × 33 × 33 Y -90 dose point kernel to generate 3D dose distribution images.12 2.F. Data analysis
The organ dose, normalized absolute error (NAE) of the differential dose volume histograms (DDVHs) and cumulative dose volume histograms (CDVHs) for each target organ was evaluated. The segmentation results of the reference images acquired at 24 h were used to generate the dose volume histograms. Data with and without nonrigid registration were compared to the result of phantoms with no misalignment. Each figure-of-merit (FOM) was defined as the following.
2.E. Image registration
We employed organ-by-organ affine plus B-spline image registration30 on sequential QSPECT images using the open source “” program.31 We chose the 24 h time point images as the reference (“fixed”) images.20 The registration framework started with the affine registration for global realignment, where images can be translated, rotated, scaled, and sheared. The resultant images from affine registration would later be served as the starting images for the Bspline registration. The B-spline transformation generated a deformation field where the deformation vectors were computed on a B-spline grid. The grid spacing was 4 mm in x-, y-, and z-directions, respectively. For saving computation time and lowering data complexity, multiresolution approach Medical Physics, Vol. 42, No. 2, February 2015
2.F.1. Organ dose
%Error =
dosew or w/o registration − dosew/o misalignment × 100%. dosew/o misalignment
Positive values indicate over estimation and negative values indicate under estimation of the absorbed dose. 2.F.2. NAE of DDVHs
We calculated the DDVH by determining the volume for each of the 20 dose intervals, i.e., dose bins (i), for each target organ. The volume for each bin (Vi ) represented the organ
1064
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1064
F. 4. Sample segmentation results at 72 h time point for (a) liver using S-VOI method and (b) kidneys using C-VOI method.
volume receiving a specific dose in the corresponding bin. We calculated the NAE in each bin for with and without registration using phantom with no misalignment as the reference. The total volume of the organ (VT ) was used as a normalization factor,
optimizing their registration parameters individually, based on a single core 2.4 GHz Intel CPU.
%normalized absoulte error 20 Vi_w/ or w/o registration −Vi_w/o misalignment i=1 = × 100%. VT Although organ movement between scans is probably nonrigid in nature, the merit of nonrigid registration over rigid registration is not fully addressed, especially with the consideration of the potential increased implementation complexity and computational time from nonrigid registration. Therefore, on top of nonrigid registration, we performed Euler transformation31 as the rigid registration on Phantom #1C. We compared the results of the above FOMs between rigid and nonrigid registrations. The average computational time for all organs was recorded for both registrations after
Tables II–V show the FOMs comparing with and without nonrigid registration for liver, spleen, kidneys, and lungs, respectively. All tables represented the average results of nine phantoms. Both organ dose and %NAE improved after registration regardless of the segmentation methods for all organs. For the liver results in Table II, the mean absorbed dose error improved from −7.28% ± 2.33% to −0.23% ± 0.71% by nonrigid registration based on E-VOI method. The %NAE in DDVHs improved to 6.79%–22.7% after registration using different segmentation methods. When using C-VOI for registration, the %NAE in DDVHs reduced about 10% as compared to those using S-VOI (22.70% ± 9.00% vs 12.10% ± 6.13%). Although the average errors of total absorbed dose between S-VOI and C-VOI were not obvious, the standard deviation in
3. RESULTS
T II. Dose analysis for liver. Nonrigid registration %error in total dose ± std %NAE of DDVH ± std
W W/O W W/O
a
E-VOI/E-VOIa (%)
S-VOI/E-VOIb (%)
C-VOI/E-VOIc (%)
S-VOI/S-VOId (%)
−0.23 ± 0.71 −7.28 ± 2.30 6.79 ± 1.96 28.70 ± 8.03
−1.79 ± 2.56 −11.3 ± 2.67 22.70 ± 9.00 26.30 ± 5.72
−2.40 ± 1.18 −7.01 ± 2.05 12.10 ± 6.13 28.70 ± 8.82
−2.40 ± 1.92 −7.49 ± 2.34 18.40 ± 9.69 29.80 ± 8.38
“E-VOI/E-VOI”: results based on E-VOI segmentation method as compared to results with no misalignment obtained also from E-VOI. “S-VOI/E-VOI”: results based on S-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. c “C-VOI/E-VOI”: results based on C-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. d “S-VOI/S-VOI”: results based on S-VOI segmentation method as compared to results with no misalignment obtained from S-VOI. b
Medical Physics, Vol. 42, No. 2, February 2015
1065
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1065
T III. Dose analysis for spleen. Nonrigid registration %error in total dose ± std %NAE of DDVH ± std
W W/O W W/O
E-VOI/E-VOIa (%)
S-VOI/E-VOIb (%)
C-VOI/E-VOIc (%)
S-VOI/S-VOId (%)
−2.12 ± 1.05 −15.50 ± 5.59 36.80 ± 13.20 42.10 ± 13.80
−8.64 ± 5.64 −25.60 ± 6.42 35.00 ± 6.17 38.30 ± 13.20
−4.22 ± 1.39 −16.60 ± 5.64 39.70 ± 5.46 42.00 ± 14.80
1.55 ± 2.35 −17.30 ± 5.98 26.00 ± 8.25 49.60 ± 16.80
a
E-VOI/E-VOI: results based on E-VOI segmentation method as compared to results with no misalignment obtained also from E-VOI. S-VOI/E-VOI: results based on S-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. c C-VOI/E-VOI: results based on C-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. d S-VOI/S-VOI: results based on S-VOI segmentation method as compared to results with no misalignment obtained from S-VOI. b
C-VOI was relatively smaller. For spleen (Table III), similar results in different segmentation groups were observed. For kidneys (Table IV) and lungs (Table V), SPECT segmentation was not feasible for images at later time points due to the significant activity wash out from these organs and the low count level on the images, thus, only C-VOI and E-VOI results were shown. The results from C-VOI approached to those from E-VOI. For the lungs using E-VOI method, the total absorbed dose was overestimated for 4.7% but was underestimated for 1.4% when using C-VOI before registration (Table V). The errors of total absorbed dose reduced after registration with both segmentation methods. Figure 5 showed the sample CDVHs for different organs in different phantoms. Dose underestimation was observed for all organs except lungs before registration. Nonrigid registration improved the CDVH distribution to be closer to the results of without misalignment. The difference of total organ dose between rigid and nonrigid was not very significant, i.e.,
Nien-Yun Wu Department of Biomedical Imaging and Radiological Sciences, National Yang Ming University, Taipei 112, Taiwan and Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei 112, Taiwan
Shyh-Jen Wang Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei 112, Taiwan
Na Song Department of Nuclear Medicine, Albert Einstein College of Medicine, Yeshiva University, Bronx, New York 10461
Greta S. P. Moka) Biomedical Imaging Laboratory, Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Macau SAR, China
(Received 26 June 2014; revised 26 November 2014; accepted for publication 3 January 2015; published 29 January 2015) Purpose: Voxel-level and patient-specific 3D dosimetry for targeted radionuclide therapy (TRT) typically involves serial nuclear medicine scans. Misalignment of the images can result in reduced dosimetric accuracy. Since the scans are typically performed over a period of several days, there will be patient movement between scans and possible nonrigid organ deformation. This work aims to implement and evaluate the use of nonrigid image registration on a series of quantitative SPECT (QSPECT) images for TRT dosimetry. Methods: A population of 4D extended cardiac torso phantoms, comprised of three In-111 Zevalin biokinetics models and three anatomical variations, was generated based on the patient data. The authors simulated QSPECT acquisitions at five time points. At each time point, individual organ and whole-body deformation between scans were modeled by translating/rotating organs and the body up to 5◦/voxels, keeping ≤5% difference in organ volume. An analytical projector was used to generate realistic noisy projections for a medium energy general purpose collimator. Projections were reconstructed using OS-EM algorithm with geometric collimator detector response, attenuation, and scatter corrections. The QSPECT images were registered using organ-based nonrigid image registration method. The cumulative activity in each voxel was obtained by integrating the activity over time. Dose distribution images were obtained by convolving the cumulative activity images with a Y -90 dose kernel. Dose volume histograms (DVHs) for organs-of-interest were analyzed. Results: After nonrigid registration, the mean differences in organ doses compared to the case without misalignment were improved from (−15.50 ± 5.59)% to (−2.12 ± 1.05)% and (−7.28 ± 2.30)% to (−0.23 ± 0.71)% for the spleen and liver, respectively. For all organs, the cumulative DVHs showed improvement after nonrigid registration and the normalized absolute error of differential DVHs ranged from 6.79% to 22.70% for liver and 26.00% to 39.70% for spleen with different segmentation methods. Conclusions: These results demonstrated that nonrigid registration of sequential QSPECT images is feasible for TRT and improves the accuracy of 3D dosimetry. C 2015 American Association of Physicists in Medicine. [http://dx.doi.org/10.1118/1.4906242] Key words: dosimetry, nonrigid registration, quantitative SPECT, targeted radionuclide therapy
1. INTRODUCTION In targeted radionuclide therapy (TRT), accurate estimates of organ dose and 3D dose distribution are important in delivering maximum tumor doses and reducing normal organ toxicity for optimal dosing regimens. The 3D dose distribution depends on various factors including patient-specific pharmacokinetics 1060
Med. Phys. 42 (2), February 2015
and anatomy. The dosimetric planning based on the 3D dose distribution can potentially improve the treatment efficacy as compared to the standardized regimen that was only based on the patient weight or body surface area.1–3 The 3D patient-specific dosimetry requires an estimate of the voxel-based time-integrated activities which are typically obtained by quantitative nuclear medicine imaging at multiple
0094-2405/2015/42(2)/1060/11/$30.00
© 2015 Am. Assoc. Phys. Med.
1060
1061
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
time points, using either conventional planar scintigraphy, SPECT, or PET. However, only SPECT or PET can provide the spatial information needed for voxel-level dosimetry. With appropriate compensation for image degrading factors, many groups have already proven that SPECT can provide accurate activity quantification for TRT.4–8 Previously, a comprehensive quantitative SPECT method called QSPECT was proposed and proven to be accurate for activity quantification and making 3D dosimetry feasible.6,9–11 Once the activity distributions at different time points are obtained, one can convert them to dose-rate images using either a dose point kernel12 or Monte Carlo simulations.13,14 Voxel-by-voxel integration of the dose-rate images over time generates 3D voxel-based dose images. Misregistration error among serial images is an important factor limiting the dosimetric accuracy in TRT.15 Potential sources of misregistration are changes in patient position, organ deformation, and tumor progression/regression between different scans. These errors will lead to misestimation of organ dose distributions, which can in turn lead to a nonaccurate prediction of treatment response. To address the misregistration problem, research groups employed different image registration methods among sequential SPECT images with positive results. Papavasileiou et al.16 developed a 4D rigid image registration scheme. They generated the temporal information by fitting the maximum count of the tumor at each time point to a polynomial function. Then, the temporal information and
the voxel-based similarity criteria were combined for simultaneous SPECT registration. A uniform radial deformation was applied by Dewaraja et al.17 to register the tumor on serial quantitative SPECT images. Sjogreen-Gleisner et al.18 registered serial CT images and applied the resultant deformation field to the SPECT images for mutual information-based rigid or nonrigid registration. This method avoided using the time-varying activity distributions on SPECT images which may affect the image registration process. However, it may change the SPECT intensity distribution at different time points.19 To the best of our knowledge, most of the existing studies relied on the use of CT images for registration or applied only rigid transformation for quantitative SPECT dosimetry. However, nonrigid deformation in organ/tumor is more likely in the clinical cases20 and rigid registration may not be adequate especially for small organs or lesions. Nonrigid registration on CT images was previously performed to aid the registration of SPECT images. However, direct nonrigid registration on SPECT images is still warranted as CT images are usually not acquired at every time point for TRT. The aim of this work was to improve the TRT dosimetry using organby-organ nonrigid image registration on QSPECT images. The potential improvement of using nonrigid registration as compared to rigid registration on sequential SPECT images was assessed. We also evaluated the role of CT for potential improvement on segmentation accuracy.
F. 1. Sampled coronal slice of the activity maps of nine XCAT phantoms. Medical Physics, Vol. 42, No. 2, February 2015
1061
1062
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1062
2. METHODS AND MATERIALS 2.A. 4D extended cardiac torso (XCAT) phantom population
The 4D XCAT (Ref. 21) phantom with realistic modeling of anatomy and In-111 Zevalin activity distribution of a male patient was used in this study. We modeled a phantom population of three anatomical variations with three respective activity uptake ratios, i.e., a total of nine phantoms (Fig. 1). Uniform activity was simulated in the background and four organs, i.e., heart, liver, kidneys, and spleen,10 while the activity distribution for lungs was not uniform since the activity in the airways was set to zero. The activity variations among the phantoms were obtained from a set of clinical patient data.10 These patients underwent sequential scans at 0–1, 4–6, 24, 72, and 144 h after injection of In-111 Zevalin. The activity in each organ was measured by conventional planar method while QSPECT was performed at 24 h. For each organ, the time-activity curves (TACs) were generated with exponential fitting to calculate the effective half-life and then the activity level was extrapolated back to the injection time. We used these time-varying activity data and the effective halflives of each target organ to simulate SPECT scans acquired at 1, 12, 24, 72, and 144 h postinjection (Table I). The random local deformation of the liver, kidneys, spleen, and stomach was modeled in 1, 12, 72, and 144 h time point to model the organ deformation between scans, while the 24 h time point was set as the reference. These organs were randomly translated and rotated within 5 pixels (11.05 mm) or degrees, respectively, and the change of the total volume of each organ was held within 5% except for the stomach. The boundaries of lungs were defined by the deformation of liver and heart. Whole body rigid transformation within 5 pixels (11.05 mm) or degrees of translation or rotation was also modeled randomly to simulate the whole body interscan movement.20 The nonrigid movement was then obtained by combining the local organ nonrigid and whole body rigid deformation (Fig. 2). Phantoms without any misalignment were used as the gold standard for evaluation. 2.B. Simulation
An analytical projector including attenuation, scatter, and geometric collimator-detector-response (GCDR) was used to generate realistic projections. In both forward and back projection processes, the attenuation was modeled by the XCAT attenuation maps. The scatter was modeled by the effective T I. Organ activity and effective half-life for Phantom #1C (Ref. 10).
Heart Lungs Liver Spleen Kidneys
Activity (MBq/ml) (×10−3)
Effective half-life (h)
1h
12 h
24 h
72 h
144 h
43.5 43.5 85.1 63.1 76.9
7.71 3.61 13.74 27.81 7.78
6.47 3.01 12.28 24.94 6.67
5.35 2.50 11.42 21.60 5.95
2.49 1.18 7.58 13.13 3.99
0.79 0.38 4.26 5.92 2.16
Medical Physics, Vol. 42, No. 2, February 2015
F. 2. Sampled noise-free projections of Phantom #1C at 72 h time point for (a) with no misalignment, (b) with local, and (c) with total nonrigid deformation. The yellow line indicates the position of the horizontal image profiles as shown in (d).
source scatter estimation (ESSE) method,22 which models the scatter and down scatter from the two separate photopeaks into the two windows based on scattering physics by using kernels estimated from Monte Carlo simulations. It models the object dependent, spatially varying scatter response and takes into account the effects of the nonuniform attenuation distribution. The GCDR was computed using an analytic formulation proposed by Metz et al.23 We simulated a GE Discovery VH Hawkeye SPECT/CT system with crystal thickness of 2.54 cm mounted with a medium energy general purpose (MEGP) collimator. The intrinsic resolution of the system was 0.4 cm. Both photopeaks of In-111 (171 and 245 keV) were considered with appropriate abundances (90.2% and 94%, respectively)24 and probability of detection to generate the abundance weighted average energy of 192.6 keV for attenuation modeling and correction in reconstruction.25 The projections had 128 transaxial and 170 axial bins with 128 views and 30 s/view over 360◦ acquisition. In order to model the continuous nature of the activity distribution, projections were generated using isotropic phantoms with voxel size of 0.221 cm. The projections were then collapsed to a bin size of 0.442 cm for reconstruction. A system calibration factor of 1.43 × 10−4 counts s−1 Bq−1 was used to scale the noise-free projection to a clinical count level and then modeled with Poisson noise to obtain realistic noisy projections. Figure 3 shows the noisy projections with local and global deformation for Phantom #1C at different time points. 2.C. Quantitative reconstruction
We reconstructed the data using QSPECT method which utilized a rotation-based projector26 compensated with attenuation, ESSE scatter model, and GCDR in the OS-EM
1063
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1063
F. 3. Sampled noisy projections at different time points with (a) local deformation and (b) global deformation.
algorithm (8 iterations and 16 subsets, i.e., 128 updates).27 A more detailed description of QSPECT could be found in the work of He et al.6 The reconstructed matrix was 128 × 128 × 170. No postfiltering was performed on the reconstructed images. 2.D. Image segmentation
Before registration, each organ-of-interest was segmented out at each time point. There are three kinds of volume-ofinterest (VOI) used in this study: (i) “exact VOI” (E-VOI) where VOIs were mapped out from the phantoms directly with no segmentation errors. (ii) “SPECT VOI” (S-VOI) where VOIs were obtained directly on the QSPECT images. (iii) “CT VOI” (C-VOI) where VOIs were drawn on the attenuation maps that served as the CT images. For both S-VOI and CVOI methods, VOIs were segmented using the semiautomatic region growing segmentation (snake) method28 from the open source program “ITK-SNAP” (Ref. 29) and adjusted based on the interpretation of an experienced operator (Fig. 4).
with 3 and 2 resolution levels were used in affine and B-spline registration, respectively. Normalized mutual information32,33 was used to measure the image similarity between the fixed and “moving” images based on the 64 histograms bins for the 2 registration steps. Optimization of the registration process was performed by the standard nonadaptive Robbins–Monro (RM) algorithm.34 Numerical voxel-by-voxel integration was later performed on the registered images over 5 time points followed by convolution with a 33 × 33 × 33 Y -90 dose point kernel to generate 3D dose distribution images.12 2.F. Data analysis
The organ dose, normalized absolute error (NAE) of the differential dose volume histograms (DDVHs) and cumulative dose volume histograms (CDVHs) for each target organ was evaluated. The segmentation results of the reference images acquired at 24 h were used to generate the dose volume histograms. Data with and without nonrigid registration were compared to the result of phantoms with no misalignment. Each figure-of-merit (FOM) was defined as the following.
2.E. Image registration
We employed organ-by-organ affine plus B-spline image registration30 on sequential QSPECT images using the open source “” program.31 We chose the 24 h time point images as the reference (“fixed”) images.20 The registration framework started with the affine registration for global realignment, where images can be translated, rotated, scaled, and sheared. The resultant images from affine registration would later be served as the starting images for the Bspline registration. The B-spline transformation generated a deformation field where the deformation vectors were computed on a B-spline grid. The grid spacing was 4 mm in x-, y-, and z-directions, respectively. For saving computation time and lowering data complexity, multiresolution approach Medical Physics, Vol. 42, No. 2, February 2015
2.F.1. Organ dose
%Error =
dosew or w/o registration − dosew/o misalignment × 100%. dosew/o misalignment
Positive values indicate over estimation and negative values indicate under estimation of the absorbed dose. 2.F.2. NAE of DDVHs
We calculated the DDVH by determining the volume for each of the 20 dose intervals, i.e., dose bins (i), for each target organ. The volume for each bin (Vi ) represented the organ
1064
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1064
F. 4. Sample segmentation results at 72 h time point for (a) liver using S-VOI method and (b) kidneys using C-VOI method.
volume receiving a specific dose in the corresponding bin. We calculated the NAE in each bin for with and without registration using phantom with no misalignment as the reference. The total volume of the organ (VT ) was used as a normalization factor,
optimizing their registration parameters individually, based on a single core 2.4 GHz Intel CPU.
%normalized absoulte error 20 Vi_w/ or w/o registration −Vi_w/o misalignment i=1 = × 100%. VT Although organ movement between scans is probably nonrigid in nature, the merit of nonrigid registration over rigid registration is not fully addressed, especially with the consideration of the potential increased implementation complexity and computational time from nonrigid registration. Therefore, on top of nonrigid registration, we performed Euler transformation31 as the rigid registration on Phantom #1C. We compared the results of the above FOMs between rigid and nonrigid registrations. The average computational time for all organs was recorded for both registrations after
Tables II–V show the FOMs comparing with and without nonrigid registration for liver, spleen, kidneys, and lungs, respectively. All tables represented the average results of nine phantoms. Both organ dose and %NAE improved after registration regardless of the segmentation methods for all organs. For the liver results in Table II, the mean absorbed dose error improved from −7.28% ± 2.33% to −0.23% ± 0.71% by nonrigid registration based on E-VOI method. The %NAE in DDVHs improved to 6.79%–22.7% after registration using different segmentation methods. When using C-VOI for registration, the %NAE in DDVHs reduced about 10% as compared to those using S-VOI (22.70% ± 9.00% vs 12.10% ± 6.13%). Although the average errors of total absorbed dose between S-VOI and C-VOI were not obvious, the standard deviation in
3. RESULTS
T II. Dose analysis for liver. Nonrigid registration %error in total dose ± std %NAE of DDVH ± std
W W/O W W/O
a
E-VOI/E-VOIa (%)
S-VOI/E-VOIb (%)
C-VOI/E-VOIc (%)
S-VOI/S-VOId (%)
−0.23 ± 0.71 −7.28 ± 2.30 6.79 ± 1.96 28.70 ± 8.03
−1.79 ± 2.56 −11.3 ± 2.67 22.70 ± 9.00 26.30 ± 5.72
−2.40 ± 1.18 −7.01 ± 2.05 12.10 ± 6.13 28.70 ± 8.82
−2.40 ± 1.92 −7.49 ± 2.34 18.40 ± 9.69 29.80 ± 8.38
“E-VOI/E-VOI”: results based on E-VOI segmentation method as compared to results with no misalignment obtained also from E-VOI. “S-VOI/E-VOI”: results based on S-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. c “C-VOI/E-VOI”: results based on C-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. d “S-VOI/S-VOI”: results based on S-VOI segmentation method as compared to results with no misalignment obtained from S-VOI. b
Medical Physics, Vol. 42, No. 2, February 2015
1065
Ao et al.: Radionuclide dosimetry using nonrigid registration on sequential SPECT
1065
T III. Dose analysis for spleen. Nonrigid registration %error in total dose ± std %NAE of DDVH ± std
W W/O W W/O
E-VOI/E-VOIa (%)
S-VOI/E-VOIb (%)
C-VOI/E-VOIc (%)
S-VOI/S-VOId (%)
−2.12 ± 1.05 −15.50 ± 5.59 36.80 ± 13.20 42.10 ± 13.80
−8.64 ± 5.64 −25.60 ± 6.42 35.00 ± 6.17 38.30 ± 13.20
−4.22 ± 1.39 −16.60 ± 5.64 39.70 ± 5.46 42.00 ± 14.80
1.55 ± 2.35 −17.30 ± 5.98 26.00 ± 8.25 49.60 ± 16.80
a
E-VOI/E-VOI: results based on E-VOI segmentation method as compared to results with no misalignment obtained also from E-VOI. S-VOI/E-VOI: results based on S-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. c C-VOI/E-VOI: results based on C-VOI segmentation method as compared to results with no misalignment obtained from E-VOI. d S-VOI/S-VOI: results based on S-VOI segmentation method as compared to results with no misalignment obtained from S-VOI. b
C-VOI was relatively smaller. For spleen (Table III), similar results in different segmentation groups were observed. For kidneys (Table IV) and lungs (Table V), SPECT segmentation was not feasible for images at later time points due to the significant activity wash out from these organs and the low count level on the images, thus, only C-VOI and E-VOI results were shown. The results from C-VOI approached to those from E-VOI. For the lungs using E-VOI method, the total absorbed dose was overestimated for 4.7% but was underestimated for 1.4% when using C-VOI before registration (Table V). The errors of total absorbed dose reduced after registration with both segmentation methods. Figure 5 showed the sample CDVHs for different organs in different phantoms. Dose underestimation was observed for all organs except lungs before registration. Nonrigid registration improved the CDVH distribution to be closer to the results of without misalignment. The difference of total organ dose between rigid and nonrigid was not very significant, i.e.,
