Cancer Translational Medicine

: 2018  |  Volume : 4  |  Issue : 6  |  Page : 143--152

Assessing the feasibility of using deformable registration for onboard multimodality-based target localization in radiation therapy

Ge Ren1, Yawei Zhang2, Lei Ren2,  
1 Department of Radiation Oncology, Duke University Medical Center; Department of Radiation Oncology, Duke University, Durham, North Carolina, USA; Department of Medical Physics, Duke Kunshan University, Kunshan, Jiangsu, China
2 Department of Radiation Oncology, Duke University Medical Center; Department of Radiation Oncology, Duke University, Durham, North Carolina, USA

Correspondence Address:
Dr. Lei Ren
Department of Radiation Oncology, Duke University Medical Center, DUMC Box 3295, Durham, North Carolina 27710


Aim: This study aims to explore the feasibility of using prior images and deformable image registration to generate onboard multimodality images to improve the soft tissue contrast of cone beam computed tomography (CBCT), for target localization in radiation therapy. Methods: B-spline-based deformable registration is used to register magnetic resonance/computed tomography (MR/CT) images with CBCT images to generate synthetic onboard MR/CT images for onboard target localization. Liver, prostate, and breast patient data were used in the study to investigate the feasibility of the method. Results: Most of the tumor volume defined by the onboard synthetic images was covered by the planning target volume (PTV) based on the shifts applied in clinical practice. For 6 liver cases, 5 prostate cases, and all the breast cases, the synthetic images allowed the reduction of the PTV margin to 1.5–7 mm, 1–4 mm, and 1–1.5 mm, respectively. The dose to the normal tissue can be reduced based on the optimized margin. Conclusion: This study demonstrated the feasibility of using onboard synthetic multimodality imaging to improve the soft tissue contrast for target localization in low contrast regions. This new technique holds great promises to optimize the PTV margin and improve the treatment accuracy in radiation therapy.

How to cite this article:
Ren G, Zhang Y, Ren L. Assessing the feasibility of using deformable registration for onboard multimodality-based target localization in radiation therapy.Cancer Transl Med 2018;4:143-152

How to cite this URL:
Ren G, Zhang Y, Ren L. Assessing the feasibility of using deformable registration for onboard multimodality-based target localization in radiation therapy. Cancer Transl Med [serial online] 2018 [cited 2020 Jul 14 ];4:143-152
Available from:

Full Text


Cone beam computed tomography (CBCT) is the most commonly used onboard 3D imaging technique to generate 3D/4D images of the patient tissues, for target localization in radiation therapy.[1],[2] However, CBCT has poorer image quality than the computed tomography (CT) images in terms of image contrast and artifacts, partially due to the large amount of scatter caused by its large imaging volume.[3],[4] As a result, the visualization of the target volume in the soft tissue, such as liver tumor, is rather limited in the CBCT.[5] This poor visualization of the target in CBCT images would severely limit the accuracy of the target localization before each fraction of treatment.[6]

Because of the limited visualization of the low contrast tumors in CBCT images, the rigid registration between the planning CT (P-CT) images and CBCT images is mainly based on the landmarks or the boundary of the organs. However, large deformations can exist from CT to CBCT images due to patient respiratory motion variations or anatomical changes. As a result, the rigid registration based on organ boundary may lead to large localization errors of the tumor volume inside the organ and thus requires large planning treatment volume (PTV) margins to account for the localization error.[7],[8],[9] Enhancing the tumor volume contrast in onboard imaging will be vital for improving the precision of localization and reducing the PTV margins to minimize excessive radiation dose to the surrounding healthy tissues. This is especially critical for hypofractionated stereotactic body radiotherapy (SBRT) treatments with high fractional dose and low number of fractions.[10],[11]

Onboard magnetic resonance imaging (MRI) radiotherapy systems have been developed to provide superb soft tissue contrast in onboard MRI images for target localization.[12],[13] However, due to its high cost, MRI radiotherapy systems are available in only a very limited number of hospitals worldwide. This severely limits its applicability for routine patient treatments. In the foreseeable future, CBCT remains the standard onboard 3D imaging technique for most of the treatments in radiation therapy.[14] Therefore, it is important to enhance the image contrast in CBCT images to improve the accuracy of localization. One potential solution is to use prior high contrast images of patients, acquired during simulation, such as the contrast-enhanced CT (CE-CT) images and MR images.[15] While CE-CT and MR images have been routinely used for target delineation in the treatment planning process, their usage for onboard target localization has not been fully explored.[16],[17]

This study evaluated the feasibility of using prior high contrast images and deformable image registration (DIR) to generate onboard multimodality images to improve the target localization accuracy in radiation therapy. The research includes three aims: (1) Evaluate whether the registration and margin design in the current clinical practice is sufficient to ensure the coverage of the deformed tumor. (2) Evaluate the potential for margin optimization based on the synthetic multimodality imaging technique. (3) Evaluate the effects of hypothetical DIR uncertainties on the optimized PTV margin. This feasibility study paves the way to generate high contrast onboard images on a conventional linear accelerator with X-ray imaging systems to improve the localization, precision, and optimization of PTV margin for low contrast soft tissue tumors.


Generation of the onboard multimodality images

Image acquisition

MRI and CT images were acquired during the simulation stage for treatment planning. For the 8 liver SBRT patients, the MRI_T1 FSPGR (Fast Spoiled Gradient Recalled Echo) or the CE-CT images were used as the prior images. The tumor contour on the MR/CT was delineated by the oncologist. Patients labeled as L-1, L-3, L-4, and L-7 were scanned with breath-hold method, with tumor contoured as the clinical target volume (CTV) in the simulation images. Patients labeled as L-2, L-5, L-6, and L-8 were scanned with free breathing method, with tumor contoured as the internal target volume (ITV) in the images. For the 7 prostate patients and 7 breast patients, the P-CT images were used as the prior images, with tumor contoured as CTV in the images. For all patients, daily CBCT images were acquired onboard. These patient image sets were exported from Eclipse (Varian Medical Systems) and anonymized for the DIR.

Deformable image registration

The anonymized images were imported to the velocity AI (Varian Medical Systems) software for image registration. The MR/CT images were set as the source images and registered with the CBCT images to generate synthetic onboard images in the velocity AI [Figure 1]. This registration includes two parts: (1) A rigid registration between the images to translate and rotate the two images closer before starting a deformable registration and (2) A B-spline algorithm-based deformable registration. The CBCT/MR-corrected single-pass deformable registration tools were applied for this process in velocity AI. The average computational time was about 3 min. Then, the deformation vector field obtained from the DIR was used to propagate the tumor contour, drawn by the physician, to the synthetic image coordinate to generate onboard tumor volume. The regions of interest used in the registration were the whole images, the prostate and surrounding organs, and the breast for the liver, prostate, and breast lesions, respectively.{Figure 1}

Assessment of the quality of deformable image registration

The most intuitive approach to validate the DIR algorithm is to check the anatomical landmarks on the merged images.[18] To determine a better modality of prior image for DIR, the MR and CT images of the same patient were registered to the CBCT images. The quality of the DIR was checked in the registered images based on the alignment of the landmarks, such as the boundary of the target organ and the surrounding organs. For the prostate patients, the quality of deformable registration was checked through the shape of the bladder and rectum. To minimize the observer bias, the DIR quality was checked by two researchers independently. The image registration accuracy was qualitatively evaluated for different prior images, and the images with better registration accuracy were used as prior images in the following assessments. Then, the images and contours were exported from velocity AI for further onboard target localization and evaluation.

Evaluation of the tumor coverage based on the current clinical practice

The images and contours were imported into Eclipse patient data sets for evaluation. To evaluate the onboard tumor volume coverage in the clinical practice, the synthetic onboard images were translated based on the shifts determined by the onboard CT-CBCT rigid registration in clinical practice [Figure 2]. The CTV/ITV volume in the shifted onboard synthetic images was then mapped onto the P-CT images to evaluate its coverage in the original clinical plan.{Figure 2}

Planning target volume margin optimization and dosimetric effects

Equivalent volume margin

In clinical practice, the tumor volume may be expanded anisotropically to protect some critical organs. To simplify the representation of the original planned margin, the concept of equivalent volume margin (EVM) was used to quantify the original planned margins. EVM is defined as the isotropic margin that would result in the same PTV as the planned PTV with anisotropic margins.

Optimization of the clinical target volume/internal target volume to planning target volume margin

A minimum ITV to PTV margin was determined to ensure the target coverage based on the synthetic onboard images [Figure 3]. To obtain the minimum margin, the tumor volume on the P-CT was first expanded by 1 mm to generate a PTV. The P-CT images were translationally shifted to register with the onboard synthetic image to verify if the PTV can cover the tumor volume in the synthetic onboard images. The tumor coverage was checked slice by slice on three sectional views. If the generated PTV was not large enough to cover the deformed tumor volume in the synthetic images, the PTV margin would be expanded further until it was sufficient to cover the onboard tumor volume. The final margin determined is regarded as the optimized PTV margin based on the onboard synthetic images. The optimized margin was also compared with the EVM to determine the reduction or increase of the CTV/ITV to PTV margin.{Figure 3}

Evaluation of the dosimetric effects

The dosimetric effect of using the optimized PTV margin was evaluated in Eclipse. The clinical PTV was replaced by the optimized PTV. The constraints and other settings of the optimized plan were controlled in the same way as the clinical plan. The optimization of the beam fluence started with the fluence in the original clinical treatment plan. The dose distribution and dose-volume histogram (DVH) to specific organs at risk (OAR) were compared between the reoptimized plan and the original clinical plan.

Evaluation of the hypothetical margin to account for uncertainties in deformable image registration

The effects of uncertainties in DIR were evaluated by adding hypothetical margins to the tumor volume in the onboard synthetic images that accounted for the uncertainties [Figure 4]. We defined the benefit margin as the largest DIR uncertainty margin that will lead to reduced PTV volume, compared to the original planning PTV, based on the optimization of the CTV/ITV to PTV margin. Specifically, the deformed tumor contour in the onboard synthetic images was first expanded by 1 mm DIR margin and then was registered to the P-CT using a translational shift. If the original planning PTV was large enough to cover the expanded onboard tumor volume, the DIR margin was increased by another 1 mm until the planning PTV was not sufficient to cover the tumor volume. The final DIR margin determined is the benefit margin defined in the study.{Figure 4}


Deformable image registration quality comparison of using computed tomography/magnetic resonance as the prior images

In liver cases, before DIR, the boundary of the liver does not match in the axial view of the MR/CBCT merged images and the CE-CT/CBCT merged images. In the sagittal view of the merged CE-CT/CBCT images, the altitude of the diaphragm on CBCT images was higher than on the CT images. After DIR, the MR liver boundary extended out of the CBCT liver boundary [Figure 5], red arrow in the second line], while the CT liver boundary aligns well with the CBCT liver boundary [[Figure 5], fourth line, red arrow].{Figure 5}

For the prostate cases, the alignment of the boundary of the landmarks on the MR and CBCT images does not change significantly after DIR. The bladder level is quite different in the MR/CBCT merged images, while the alignment in the CT/CBCT merged images was better. However, the rectum boundary does not align well with the CBCT images for both modalities [Figure 6].{Figure 6}

In breast cases, after DIR, the breast boundary matched well for the two modalities. However, the MR/CBCT-merged images showed unrealistically deformed mammary ducts after DIR. The changes in the breast were large, and the internal anatomical structures were strongly distorted on the deformed MR images, which strongly influenced the location of the tumor and making this modality unreasonable for target localization. As compared with this, the change of internal anatomical structures in CT to CBCT DIR was more realistic [Figure 7].{Figure 7}

Coverage of the tumor volume based on clinical shifts

For the liver cases, the deformed CTV/ITV coverage of 6 cases (8 cases in total) is above 90%. The lowest coverage is 76% in the case of L-8 [Table 1]. For the prostate cases, the coverage of all the cases is above 94% [Table 2]. For the breast cases, the coverage is above 90% except patient B-4, which showed a lowest coverage of 87.9% [Table 3]. The average coverage for liver, prostate, and breast are 97.4%, 99.3%, and 96.4%, respectively.{Table 1}{Table 2}{Table 3}

The cases with coverage < 90% are shown in [Figure 8], [Figure 9], [Figure 10]. For case L-4, generally, the tumor coverage decreased as treatment day increased due to the increasing day-to-day anatomical and respiratory variations. The CBCT images on day 10 showed that the volume of the lung and the height of the liver changed substantially from those in the P-CT, which led to large localization errors and undercoverage of the tumor volume [Table 1]. For patient L-8, as shown in the CBCT images in [Figure 9], the PTV (14.9 cc) was relatively small as compared with the PTV of other liver patients (45–414 cc). As a result, even small localization errors led to significant undercoverage of the percentage volume for the tumor. For case B-4, as shown in [Figure 10], there was no margin between the CTV and planning PTV in the areas close to the skin or chest wall. As a result, localization errors based on the clinical shifts led to undercoverage of the tumor volume. For these cases, the margin design and CBCT-CT-based rigid registration may not be sufficient to ensure the target coverage.{Figure 8}{Figure 9}{Figure 10}

Planning target volume margin optimization and dosimetric effects

For most of the liver, prostate, and breast cases, the synthetic images allowed the reduction of PTV margins up to 7 mm, 4 mm, and 1.5 mm, respectively [Figure 11], [Figure 12], [Figure 13]. For patient L-4, L-7, P-4, and P-7, the margin needed was increased by 4 mm, 0.5 mm, 0.5 mm, and 1.5 mm, respectively, to ensure adequate coverage of the tumor [Table 4] and [Table 5].{Figure 11}{Figure 12}{Figure 13}{Table 4}{Table 5}

For the prostate cases, 5 cases had reduced margin while 2 cases had increased margin. For cases with reduced margins, the dose to the bladder was reduced in the optimized plan, compared to the original plan [Table 6]. For the 2 cases with the increased margin, the dose to the rectum was increased in the optimized plan [Table 7]. For breast cases, dose to the normal breast issue was reduced in the optimized plan with reduced PTV margins and was increased in the optimized plan with increased PTV margins [Table 8].{Table 6}{Table 7}{Table 8}

Results of the evaluation of the benefit margins for hypothetical deformable image registration uncertainty

For the liver cases, 7 patients had benefit margins of 2–4 mm. For the prostate cases, the benefit margins were 1–5 mm with no benefit margins for two cases. For the breast cases, the benefit margins were 2–3 mm [Table 9].{Table 9}


Comparison of the magnetic resonance imaging-based and computed tomography images-based deformable image registration

Qualitative evaluations showed that the DIR between prior CT and CBCT achieved better alignment of the organ boundaries for liver, breast, and prostate cases than the DIR using prior MRI and CBCT. This is consistent with the previous study results, which reported that the MR to CT DIR has larger average absolute error than CT to CT DIR.[19] This may be due to the different image intensity values and contrast between CT/CBCT and MR images making it challenging for intensity-based registration algorithms in the velocity.[17],[20] Other DIR algorithms, such as surface-based finite element matching, could be more suitable for MR-CBCT registration.

Evaluation of the coverage based on the clinical practice

In this study, for most cases, over 90% of the tumor volume defined by the onboard synthetic images was covered by the PTV, based on the shifts applied in clinical practice. The average dose coverage for liver, prostate, and breast were 97.4%, 99.3%, and 96.4%, respectively. For liver cases, L-4 patient had the largest treatment fractions and one of the lowest tumor coverages (88.2%) among the 9 patients. This is mainly because the patient interfraction variations tend to increase with the fraction number and lead to more errors in localizing the tumor volume. Therefore, accurate target localization is especially critical for cases with large fraction numbers to minimize the effects of interfraction variations. The liver case L-8 had the lowest coverage among all liver cases. This is mainly because the PTV for L-8 (14.9 cc) was much smaller than other liver cases (45–414 cc), making tumor volume percentage coverage degradation very susceptible to any localization errors. Under this scenario, accurate localization and adequate margin design were essential for ensuring the coverage of the tumor. For breast patients, patient B-4 had tumor coverage below 90%. This is mainly because no PTV margin was added in the areas close to the skin or chest wall.[21] As a result, any localization errors caused by the limited CBCT image quality led to degradation of the tumor coverage.

Optimization of the margin and the dosimetric effect

For most of the liver and prostate cases, the synthetic images allowed the reduction of the PTV margin by up to 6 mm and 4 mm, respectively. For the breast cases, the PTV margin was reduced by 1–1.5 mm. For patients L-4, L-7, P-4, and P-7, the margin needed was increased by 4 mm, 0.5 mm, 0.5 mm, and 1.5 mm, respectively, showing that the original margin deigns were not sufficient for target coverage. Even though the optimized margin for the four cases was larger than the original margin, it does not indicate that the DIR-based optimized plan is worse than the original plan. Instead, the optimized plan uses slightly larger PTV margin to ensure adequate coverage of the tumor, while the original plan uses inadequate smaller PTV margin without ensuring the coverage of the tumor.

For cases with reduced margins for the optimized PTV, the dose to the OARs was reduced after the optimization. For the case L-1 using the optimized margin plan, the chest wall volume receiving larger than 30 Gy was decreased from 24 to 14 cc, which was also reflected in the dose map and DVH [Figure 14]. For cases with increased margins, the doses to the OARs surrounding the target were modestly increased as compared to the original plan.{Figure 14}

Margin optimization based on hypothetical deformable image registration uncertainty

To consider the effects of DIR uncertainty on the target localization, using onboard synthetic images, margins of hypothetical DIR error were incorporated in the workflow. Results of benefit margins for different sites were included in IIID. If the uncertainties of the DIR were less than the benefit margin, the final optimized PTV using synthetic images would still be smaller than the original PTV, with potential of spare doses to the OAR. On the other hand, if the uncertainties of the DIR were larger than the benefit margin, the final optimized PTV using synthetic images would be larger than the original PTV, potentially leading to increased dose to the OAR. However, the optimized PTV with synthetic contrast-enhanced images ensures adequate coverage of the tumor during the treatment, while the original PTV with CBCT cannot guarantee the coverage of the tumor.

Evaluating the specific uncertainties of the DIR registration is a challenging task. Generally, DIR can be evaluated through the averages of the contour similarity metrics and landmark distances. However, these cannot assure sufficient accuracy within homogeneous tissue due to the low contrast of the soft tissue.[17] Kristy reported that the algorithm accuracy can be determined by comparing the computer-predicted displacement at each bifurcation point with the displacement computed from the oncologists' annotations. However, the test is limited to a narrow range of clinical situations.[19] The DIR algorithm can also be analyzed by evaluating some features. For the lung cases, (1) the landmarks inside the organ, (2) several metrics of propagated contours, and (3) the inconsistency of two opposed vector fields were used to evaluate the DIR quality.[17] In future, these methods will be used to quantitatively evaluate our DIR uncertainties and will be incorporated into the synthetic image workflow in our studies. The robustness and accuracy of the proposed approach will also be evaluated through a large cohort of patients in the future.


This study developed a workflow to use prior images and DIR, to generate onboard synthetic images with enhanced soft tissue contrast, for target localization. Preliminary results demonstrated the feasibility of this approach and evaluated its clinical impact including target coverage improvements, margin optimization, and dosimetric effects. Future studies are warranted to further incorporate registration uncertainties in the workflow and evaluate the robustness of the method through a large cohort of patients.

Financial support and sponsorship

This work was supported by the National Institutes of Health Grant No. R01-CA184173.

Conflicts of interest

There are no conflicts of interest.


1Harris W, Zhang Y, Yin FF, Ren L. Estimating 4D-CBCT from prior information and extremely limited angle projections using structural PCA and weighted free-form deformation for lung radiotherapy. Med Phys 2017; 44 (3): 1089–104.
2Chen Y, Yin FF, Zhang Y, Zhang Y, Ren L. Low dose CBCT reconstruction via prior contour based total variation (PCTV) regularization: a feasibility study. Phys Med Biol 2018; 63 (8): 085014.
3Saati S, Kaveh F, Yarmohammadi S. Comparison of cone beam computed tomography and multi slice computed tomography image quality of human dried mandible using 10 anatomical landmarks. J Clin Diagn Res 2017; 11 (2): Zc13–6.
4Tang X, Krupinski EA, Xie H, Stillman AE. On the data acquisition, image reconstruction, cone beam artifacts, and their suppression in axial MDCT and CBCT – A review. Med Phys 2018; 45 (9): e761–82.
5Weiss E, Wu J, Sleeman W, Bryant J, Mitra P, Myers M, Ivanova T, Mukhopadhyay N, Ramakrishnan V, Murphy M, Williamson J. Clinical evaluation of soft tissue organ boundary visualization on cone-beam computed tomographic imaging. Int J Radiat Oncol Biol Phys 2010; 78 (3): 929–36.
6Lechuga L, Weidlich GA. Cone Beam CT vs. Fan Beam CT: a comparison of image quality and dose delivered between two differing CT imaging modalities. Cureus 2016; 8 (9): e778.
7Brock KK. Image Processing in Radiation Therapy. 9th ed. Boca Raton, FL 33487-2742: CRC Press; 2013. p. 153–6.
8van Herk M, Bruce A, Kroes AP, Shouman T, Touw A, Lebesque JV. Quantification of organ motion during conformal radiotherapy of the prostate by three dimensional image registration. Int J Radiat Oncol Biol Phys 1995; 33 (5): 1311–20.
9Deurloo KE, Steenbakkers RJ, Zijp LJ, de Bois JA, NowakPJ, Rasch CR, van Herk M. Quantification of shape variation of prostate and seminal vesicles during external beam radiotherapy. Int J Radiat Oncol Biol Phys 2005; 61 (1): 228–38.
10Giraud P, Yorke E, Jiang S, Simon L, Rosenzweig K, Mageras G. Reduction of organ motion effects in IMRT and conformal 3D radiation delivery by using gating and tracking techniques. Cancer Radiother 2006; 10 (5): 269–82.
11Goitein M. Organ and tumor motion: an overview. Semin Radiat Oncol 2004; 14 (1): 2–9.
12Kang KM, Choi HS, Jeong BK, Song JH, Ha IB, Lee YH, Kim CH, Jeong H. MRI-based radiotherapy planning method using rigid image registration technique combined with outer body correction scheme: a feasibility study. Oncotarget 2017; 8 (33): 54497–505.
13Schmidt MA, Payne GS. Radiotherapy planning using MRI. Phys Med Biol 2015; 60 (22): R323–61.
14Dhanrajani P, Rynberg T, Ho C. Cone beam CT scan: importance of CBCT in treatment plan. Br Dent J 2018; 225 (6): 464.
15Ragel M, Nedumaran A, Makowska-Webb J. Prospective comparison of use of contrast-enhanced ultrasound and contrast-enhanced computed tomography in the Bosniak classification of complex renal cysts. Ultrasound 2016; 24 (1): 6–16.
16Yin FF, Wang Z, Yoo S, Wu QJ, Kirkpatrick J, Larrier N, Meyer J, Willett CG, Marks LB. Integration of cone-beam CT in stereotactic body radiation therapy. Technol Cancer Res Treat 2008; 7 (2): 133–9.
17Stutzer K, Haase R, Lohaus F, Barczyk S, Exner F, Lock S, Ruhaak J, Lassen-Schmidt B, Corr D, Richter C. Evaluation of a deformable registration algorithm for subsequent lung computed tomography imaging during radiochemotherapy. Med Phys 2016; 43 (9): 5028.
18Oh S, Kim S. Deformable image registration in radiation therapy. Radiat Oncol J 2017; 35 (2): 101–11.
19Brock KK. Deformable registration accuracy consortium. Results of a multi-institution deformable registration accuracy study (MIDRAS). Int J Radiat Oncol Biol Phys 2010; 76 (2): 583–96.
20Lin H, Ayan A, Zhai H, Zhu T, Both S. A quantitative evaluation of velocity AI deformable image registration. Med Phys 2010; 37 (6): 3126.
21Jones S, Fitzgerald R, Owen R, Ramsay J. Quantifying intra- and inter-fractional motion in breast radiotherapy. J Med Radiat Sci 2015; 62 (1): 40–6.