Usefulness of hybrid deformable image registration algorithms in prostate radiation therapy.

To evaluate the accuracy of commercially available hybrid deformable image registration (DIR) algorithms when using planning CT (pCT) and daily cone-beam computed tomography (CBCT) in radiation therapy for prostate cancer. The hybrid DIR algorithms in RayStation and MIM Maestro were evaluated. Contours of the prostate, bladder, rectum, and seminal vesicles (SVs) were used as region-of-interest (ROIs) to guide image deformation in the hybrid DIR and to compare the DIR accuracy. To evaluate robustness of the hybrid DIR for prostate cancer patients with organs with volume that vary on a daily basis, such as the bladder and rectum, the DIR algorithms were performed on ten pairs of CT volumes from ten patients who underwent prostate intensity-modulated radiation therapy or volumetric modulated arc therapy. In a visual evaluation, MIM caused unrealistic image deformation in soft tissues, organs, and pelvic bones. The mean dice similarity coefficient (DSC) ranged from 0.46 to 0.90 for the prostate, bladder, rectum, and SVs; the SVs had the lowest DSC. Target registration error (TRE) at the centroid of the ROIs was about 2 mm for the prostate and bladder, and about 6 mm for the rectum and SVs. RayStation did not cause unrealistic image deformation, and could maintain the shape of pelvic bones in most cases. The mean DSC and TRE at the centroid of the ROIs were about 0.9 and within 5 mm generally. In both software programs, the use of ROIs to guide image deformation had the possibility to reduce any unrealistic image deformation and might be effective to keep the DIR physically reasonable. The pCT/CBCT DIR for the prostate cancer did not reduce the DIR accuracy because of the use of ROIs to guide the image deformation.

algorithms in RayStation and MIM Maestro were evaluated. Contours of the prostate, bladder, rectum, and seminal vesicles (SVs) were used as region-of-interest (ROIs) to guide image deformation in the hybrid DIR and to compare the DIR accuracy. To evaluate robustness of the hybrid DIR for prostate cancer patients with organs with volume that vary on a daily basis, such as the bladder and rectum, the DIR algorithms were performed on ten pairs of CT volumes from ten patients who underwent prostate intensity-modulated radiation therapy or volumetric modulated arc therapy. In a visual evaluation, MIM caused unrealistic image deformation in soft tissues, organs, and pelvic bones. The mean dice similarity coefficient (DSC) ranged from 0.46 to 0.90 for the prostate, bladder, rectum, and SVs; the SVs had the lowest DSC. Target registration error (TRE) at the centroid of the ROIs was about 2 mm for the prostate and bladder, and about 6 mm for the rectum and SVs. RayStation did not cause unrealistic image deformation, and could maintain the shape of pelvic bones in most cases. The mean DSC and TRE at the centroid of the ROIs were about 0.9 and within 5 mm generally. In both software programs, the use of ROIs to guide image deformation had the possibility to reduce any unrealistic image deformation and might be effective to keep the DIR physically reasonable. The pCT/ CBCT DIR for the prostate cancer did not reduce the DIR accuracy because of the use of ROIs to guide the image deformation.

| INTRODUCTION
The prostate is known to undergo motion and shape variations [1][2][3][4][5][6][7][8][9] during a course of radiation therapy, caused by physiological changes, such as bladder volume changes and rectum filling. In image-guided radiotherapy, the prostate is imaged using kilo-voltage (kV) or mega-voltage cone-beam computed tomography (CBCT) to take into account the variety of the prostate position before the treatment.
To evaluate the daily shape and positional variations of the prostate during radiation therapy, deformable image registration (DIR), which is a non-rigid image registration process to find corresponding Deformable image registration is essential for the recently developed adaptive radiation therapy; exported deformation vector fields (DVFs), which represent the shift value and direction for a particular voxel to match to a corresponding voxel, are used for dose accumulation 10 and automatic contour propagation. 11 The DVFs exported from the DIR between planning CT (pCT) and daily CBCT are useful for evaluation of the prostate motion and deformation during a treatment. Moreover, the impact of organ motion and deformation, caused by variable volumes such as those of the bladder and rectum, on the prostate can be estimated using the DVFs.
However, it has been reported that DIR using the CBCT for the prostate region has poor accuracy because of unfavorable conditions such as noise, poor low-contrast resolution, and abdominal motion artifacts. 12 An intensity-based DIR algorithm evaluates image similarity between CT volumes by using image intensities in an optimization problem. To measure the image intensity, mathematical methods such as the sum-of-squared differences, correlation coefficients, and mutual information are employed. Similar image intensities are found at many points in homogeneous regions and may cause unrealistic image transformation. Murphy 13 investigated the influence of noise of the CBCT on the intensity-based DIR accuracy by simulating the fan-beam CT/CBCT registration in the presence of added target image noise. They found that DIR accuracy with a noisy CBCT was within a level consistent with interobserver variability for the purpose of automatic contouring. Furthermore, when CT volumes are registered deformably, every point must have a corresponding point in the other. The variable presence of bowel gas during and between treatment causes errors in the correspondence. Foskey 14 processed each image exhibiting the problem to shrink the gassy region to a point using their "deflation" algorithm in the DIR.
A feature-based DIR algorithm needs to define relevant landmarks such as points or contours in both the reference and target volumes. Because, landmarks in a target volume are transferred to match to the corresponding landmarks in the reference volume, human anatomy can be taken into account. The feature-based method is useful for noisy CBCT images. However, the selection of landmarks is usually time-consuming and similarity of the landmarks hides unrealistic image transformations in regions that are not of interest. 15,16 A hybrid DIR algorithm, which is based on the intensity-based DIR algorithm combined with the feature-based technique, has been reported by Kim et al. 17 They employed their piecewise rigid registration method for femoral and pelvic bone registrations to preserve their shape. Mean contours and point features were then incorporated as constraints into a B-spline-based DIR algorithm.
They reported high dice similarity coefficients (DSCs), which represent a spatial overlap of region-of-interest (ROIs) between CT volumes, of 0.9 for the prostate, rectum, and bladder, and 0. RayStation. A pair of volumes of pCT and CBCT from two prostate patients was tested, for whom they reported high DSCs of more than 0.9 for the prostate, rectum, and bladder. Thus, the hybrid DIR algorithm may be useful for pCT/CBCT DIR in the prostate region.
There have been few reports regarding the hybrid DIR algorithm for use in pCT/CBCT of the prostate region. In this study, we investigated the feasibility of a commercial hybrid DIR algorithm in pCT/ CBCT in the prostate region. Volumes of pCT/CBCT from ten patients who underwent radiation therapy were used to evaluate robustness of the hybrid DIR for prostate cancer patients, using the DIR algorithms in RayStation (V. 4.7.4) and MIM Maestro (V. 6.6.8) software.

2.A | Patients
To evaluate the accuracy of the commercial hybrid DIR algorithms, ten prostate cancer patients who underwent intensity-modulated radiation therapy or volumetric modulated arc therapy were selected. This retrospective study was reviewed by our institutional review board. All patients were prescribed a dose of 76 Gy in 38 fractions. They were instructed to empty their rectum and bladder, and subsequently drink 500 ml of water 30 min before the pCT and treatment.

2.B | Datasets
A pair of the volumes of the pCT and the CBCT was respectively obtained by a CT-simulation and a routine image acquisition before the treatment. The pCT images were acquired using an Aquilion One  Weistrand and Svensson and in other reports. [18][19][20] The non-linear optimization problem in ANACONDA is performed using the combination of four terms. Similarity between images is measured by a correlation coefficient in the first term of the problem. The second term of the problem is a regularization, which penalizes large shape deviations of ROIs. When an ROI type is set to "Avoidance" or "Organ", the ROI is considered in the term. When a user includes controlling ROIs and points-of-interest to guide image deformation using key organs and points, the contour regularization term and a contour matching term are added in the third and fourth term of the problem, aimed at deforming the selected structure in the target image to the corresponding structure in the reference image. 19 The terms allow fast convergence even for ROIs with large differences in size. This situation often occurs for bladders with different filling conditions. 18 When FOCUS ROIs are selected, the ROIs are defined as focus regions of the DIR. Gaussian smoothing is applied in ANA-CONDA to prevent noisy deformation grids.
MIM employs the VoxAlign Deformation Engine for the DIR, which is a constrained, intensity-based, FFD algorithm. [19][20][21] Image similarity is measured using the sum-of-squared differences technique. Multiple types of regularization are used to keep the transformation reasonable. The algorithm actively attempts to match bone and avoid tears/folds in the deformation field. 22

2.D | Deformable image registration instructions
A pair of pCT and CBCT was respectively set as the reference and target volumes for the DIR. Since a rigid image registration must exist before initiating DIR in RayStation and MIM, pCT and CBCT were rigidly registered before the DIR. No mask algorithms to account for rectal and bowel gas were applied.
In the hybrid DIR for RayStation, the contours of the prostate, bladder, rectum, and SVs were used as the controlling ROIs to guide the image deformation. Although it is possible to ignore image information with controlling ROIs, the setting was not applied for the hybrid DIR using image intensity and geometrical information. Since

2.E | Evaluation
The DIR accuracy was evaluated visually from the aspects of the unrealistic image deformation of soft tissue, organs, and pelvis bones.
Quantitative evaluation of the DIR accuracy was performed using DSC for the ROIs, target registration error (TRE) at the centroid of the ROIs and multiple evaluation points, and Jacobian determinants (JD) of the ROIs according to the American Association of Physicists in Medicine (AAPM) publication Task Group No 132 (TG-132). 23 Volume agreement between the pCT and deformed CBCT was expressed using the DSC: in eq. (1) represents the volume. A DICE score of 1 refers to two organs that overlap perfectly whilst 0 is of two ROIs that do not overlap at all.
The spatial discrepancy at the centroid position of the ROIs between pCT and deformed CBCT was expressed by the TRE and calculated by: and "r" and "d" in eqs. and may indicate an error in the registration or a limitation in the algorithm to handle complex deformation. 23 To assess the statistical significance of the mean DSC, TRE, and JD between the intensity-based DIR and hybrid DIR, P-values by two-tailed paired t-test were evaluated. The t-tests were performed between the DIRs of each software packages. In the DIR for RayStation, expansion of soft tissue at the image edges was observed in Fig. 1(c) and 1(d). The bladder for the patient had a volume change about 30 cc from the pCT to CBCT. When using the hybrid DIR, the agreement of the bladder improved as seen in Fig. 1(o). In the intensity-based DIR, two of the cases had unrealistic image deformation of the external body shape. However, this was corrected in the hybrid DIR.

3.A | Visual evaluation
In the DIR for MIM, when using the intensity-based DIR, soft tissue was expanded at the image edges as seen in Fig. 1(f) and eight cases had physically unrealistic deformation of bone structure. The expansion of the image edge was caused by the small FOV for the CBCT. The intensity-based DIR tended to move the rectum and prostate toward their buttocks. When using the hybrid DIR, four cases had less physically unrealistic deformation of bone structure and seven cases reduced the soft tissue expansion at the image edges as seen in Fig. 1(e). Physically unrealistic deformation did not improve in two cases with the hybrid DIR.  In MIM, the intensity-based DIR had the mean TRE for each ROI ranged about 6-17 mm. The maximum TRE of 35 mm was found.

3.B.2 | Target registration error
When using the hybrid DIR, MIM could reduce the TRE by 7 mm on average. In the t-test of the TREs, the statistical differences were found between the DIRs, except for the SVs. For multiple evaluation points, the average TREs for both DIRs were more than 6 mm and average SD were more than 5 mm. No significant difference was found between the DIRs. Figure 5 shows the mean JD for the ROIs with the error bars showing one SD between the pCT and the deformed CBCT. Since the SD F I G . 2. Comparison of mean dice similarity coefficient for the ROIs between the hybrid and intensity-based DIR. Error bars show one standard deviation for the data from ten patients. When the significant differences were detected between the hybrid and intensity-based DIR by t-test, P-values were indicated in the figure.

3.B.3 | Local volume change
F I G . 3. Comparison of mean Target registration error (TRE) at the region-ofinterests (ROIs) between the hybrid and intensity-based DIR. Error bars show one standard deviation for the data from ten patients. When the significant differences were detected between the hybrid and intensity-based DIR by t-test, P-values were indicated in the figure. for MIM was not obtained using the statistical tool, no error bars were added.
In RayStation, the mean JD for the bladder, prostate, and rectum were reduced from about 1.0 to 0.9 when using the hybrid DIR. The F I G . 4. Comparison of mean TRE for the multiple evaluation points between the hybrid and intensity-based DIR. Error bars show one standard deviation for the data from ten evaluation pointes for ten patients. When the significant differences were detected between the hybrid and intensity-based DIR by t-test, P-values were indicated in the figure.
F I G . 5. Comparison of mean Jacobian determinants for the ROIs between the hybrid and intensity-based DIR. Error bars show one standard deviation for the data from ten patients. When the significant differences were detected between the hybrid and intensity-based DIR by t-test, Pvalues were indicated in the figure.
In AAPM TG-132, the TRE tolerance should be no more than | 235 mm in RayStation. In order to investigate the influence of the grid resolution on the DIR accuracy in RayStation, comparisons in different grid resolution were performed. As shown in Table 1, the DIR accuracy with the largest grid resolution of 5 mm × 5 mm × 5 mm had the least TRE. No significant difference in the comparison of DSC and JD was detected. Therefore, there is little possibility that the grid resolution in RayStation affects the result of the comparison of the two software in this study.

| CONCLUSIONS
In this study, hybrid DIR algorithms in the RayStation and MIM Maestro were evaluated for the pCT/CBCT DIR in the prostate region when using non-rigid image registration. The use of ROIs of prostate, bladder, rectum, and SVs to guide the image deformation had the possibility to reduce nonphysical image deformation and the DIR accuracy improved. However, the accuracy of the hybrid DIR algorithms had large variation between the two software packages. The hybrid DIR in RayStation was robust for the chosen prostate cancer patients.