Comparison of 2D Radiographic Images and 3D Cone Beam Computed Tomography for Positioning Head-and-Neck Radiotherapy Patients

Int. J. Radiation Oncology Biol. Phys., Vol. 71, No. 3, pp. 916–925, 2008 Copyright  2008 Elsevier Inc. Printed in the USA. All rights reserved 0360-3016/08/$–see front matter

doi:10.1016/j.ijrobp.2008.01.008

PHYSICS CONTRIBUTION

COMPARISON OF 2D RADIOGRAPHIC IMAGES AND 3D CONE BEAM COMPUTED TOMOGRAPHY FOR POSITIONING HEAD-AND-NECK RADIOTHERAPY PATIENTS HENG LI, PH.D.,* X. RONALD ZHU, PH.D.,* LIFEI ZHANG, PH.D.,* LEI DONG, PH.D.,* SAM TUNG, M.S.,* ANESA AHAMAD, M.D.,y K. S. CLIFFORD CHAO, M.D.,y WILLIAM H. MORRISON, M.D.,y DAVID I. ROSENTHAL, M.D.,y DAVID L. SCHWARTZ, M.D.,y RADHE MOHAN, PH.D.,* y AND ADAM S. GARDEN, M.D. Departments of *Radiation Physics and y Radiation Oncology, The University of Texas M. D. Anderson Cancer Center, Houston, TX Purpose: To assess the positioning accuracy using two-dimensional kilovoltage (2DkV) imaging and threedimensional cone beam CT (CBCT) in patients with head and neck (H&N) cancer receiving radiation therapy. To assess the benefit of patient-specific headrest. Materials and Methods: All 21 patients studied were immobilized using thermoplastic masks with either a patientspecific vacuum bag (11 of 21, IMA) or standard clear plastic (10 of 21, IMB) headrests. Each patient was imaged with a pair of orthogonal 2DkV images in treatment position using onboard imaging before the CBCT procedure. The 2DkV and CBCT images were acquired weekly during the same session. The 2DkV images were reviewed by oncologists and also analyzed by a software tool based on mutual information (MI). Results: Ninety-eight pairs of assessable 2DkV-CBCT alignment sets were obtained. Systematic and random errors were <1.6 mm for both 2DkV and CBCT alignments. When we compared shifts determined by CBCT and 2DkV for the same patient setup, statistically significant correlations were observed in all three major directions. Among all CBCT couch shifts, 4.1% $ 0.5 cm and 18.7% $ 0.3 cm, whereas among all 2DkV (MI) shifts, 1.7% $ 0.5 cm and 11.2% $ 0.3 cm. Statistically significant difference was found on anteroposterior direction between IMA and IMB with the CBCT alignment only. Conclusions: The differences between 2D and 3D alignments were mainly caused by the relative flexibility of certain H&N structures and possibly by rotation. Better immobilization of the flexible neck is required to further reduce the setup errors for H&N patients receiving radiotherapy.  2008 Elsevier Inc. Cone beam CT, Head-and-neck cancer, Positioning accuracy.

necessity of daily imaging in additional to reliable immobilization. Thermoplastic face masks are routinely used to immobilize H&N patients (4) in combination with other devices such as bite blocks and vacuum bags. In the most common setup and alignment procedure, the lasers are aligned to external marks on the face mask and followed by portal imaging to assess the positioning accuracy. Most portal imaging studies in H&N cancers are performed with megavolt (MV) X-rays (5–14); and increasingly with kilovolt (kV) X-rays (15). The drawback of MV portal imaging is that it is a two-dimensional (2D) projection technique suffering from poor contrast resolution (16). Indeed, in a study conducted by Pisani et al. (15), who compared kV and MV 2D radiographs with regard to the

INTRODUCTION Intensity-modulated radiation therapy (IMRT) delivers highly conformal radiation to the targets while minimizing doses to normal tissues and critical organs (1, 2), better immobilization, and more precise positioning of the patients are needed at every treatment session. Positioning accuracy is even more critical for patients with head and neck (H&N) cancer undergoing IMRT because of the complexity of the anatomy and the proximity of tumors to many critical and radiation-sensitive tissues. A recent study by Zeidan et al. (3) revealed that even if every other treatment is image guided, about 11% of all treatments still are subject to threedimensional (3D) setup errors of at least 5 mm for H&N treatment from random setup error, which demonstrated the

Supported in part by an industrial grant from Varian Medical Systems. Acknowledgments—Beth Notzon of the Department of Scientific Publications of M. D. Anderson Cancer Center is greatly appreciated for her editorial review of this manuscript. Received June 21, 2007, and in revised form Jan 8, 2008. Accepted for publication Jan 13, 2008.

Reprint requests to: X. Ronald Zhu, Ph.D., Department of Radiation Physics, Unit 1150, The University of Texas M. D. Anderson Cancer Center, 1515 Holcombe Blvd., Houston, TX 77030. Tel: (713) 563-2553; Fax: (713) 563-2479; E-mail: xrzhu@ mdanderson.org Presented at the 48th Annual Meeting of the American Society for Therapeutic Radiology and Oncology (ASTRO), November 5–9, 2006, Philadelphia, PA. 916

Comparison of 2D radiographic images and 3D CBCT d H. LI et al.

online correction of setup errors, the authors concluded that kV-based images were qualitatively more effective than MV-based images. Conventional computed tomography (CT) scanners mounted on a rail system in the treatment room, often called ‘‘CT-on-rail,’’ are available in some radiation oncology clinics (17). Recently, in a study of setup uncertainties in H&N patients using 3D CT images from the CT-on-rail system, Zhang et al. (16) found variability in setup corrections for different regions of the H&N anatomy. These findings confirm the fact that the head and neck is not a rigid body. New generation, in-room systems capable of both 2DkV radiographic imaging and 3DkV cone beam CT (CBCT) imaging are now available in radiation therapy community. However, in current implementation, 2DkV images are easier to acquire than CBCT and gives much less dose to the patient. It is common to use bony structures to assess the setup deviations in 2D portal images because soft tissues are difficult to visualize in the planar projection X-ray images (16). On the other hand, volumetric CT images are capable of identifying both bony structures and soft tissues. Therefore it is not surprising that volumetric CT registration performs better for IGRT for prostate and lung cancer than does 2D bony structure registration based on portal images (18–21). In general, positioning errors could result from one or a combination of three main sources (16, 22): (1) the systematic difference between the immobilization device at the time of simulation and treatment; (2) the random setup errors in daily positioning and residual errors after corrections are executed, including execution errors in the correction process, the rotation of anatomy, and nonrigid movement of patient; and (3) trends caused by anatomic variation, such as tumor shrinkage or weight loss during radiotherapy. In this study, we only considered the first two sources of uncertainty. Assuming that rotations are small and can be approximately corrected by translational shifts (16), the main purpose of this work is to determine whether 2DkV radiographs are equivalent to 3DkV CBCT images, using bony structure alignment techniques, for determining the positional deviations of H&N patients receiving radiotherapy. In addition, we would also assess the effectiveness of patient-specific headrests using a vacuum bag. MATERIALS AND METHODS Onboard imaging system description A linear accelerator equipped with onboard imaging (OBI) system (Trilogy; Varian Medical Systems, Palo Alto, CA) was used for this work. The OBI system consists of an X-ray source and an amorphous silicon flat-panel detector mounted on two robotic arms perpendicular to the gantry of the linear accelerator. The X-ray source generates pulsed X-rays with a typical energy of 125 kVp. The flat-panel detector is 40  30 cm. The OBI system is capable of acquiring both 2D projection X-ray (kV) images and CBCT images. All 2DkV images were acquired in the anteroposterior (AP) and right lateral directions, with field of view of 26.7  20 cm. All CBCT images were acquired with the ‘‘bow-tie’’ filter, and the half-fan data acquisition geometry had a reconstruction field of view of 35  35  14 cm.

917

Fig. 1. Head phantom in a vacuum bag used to validate computed tomography–assisted (for cone beam computed tomography) and mutual information (for 2DkV) registration software.

Alignment software and verification An in-house software tool, CT-assisted targeting (CAT), was used for aligning CBCT images to the planning CT (16, 23). Another software tool based on mutual information (MI) theory (24) was implemented to align the 2DkV images with digital reconstructed radiographies (DRRs). The CAT software uses a 3D image–based structure alignment algorithm to calculate the setup shifts. The algorithm uses the region of interest (ROI) contoured in the planning CT (expanded by 2–3 mm) as an image template to match for the same image feature in the weekly CBCT, and the best match is where the maximum similarity is achieved. More detail about the registration algorithm and the software can be found elsewhere (16, 23). In using the MI software tool for 2D registration, rectangular ROIs were manually selected on the AP and right lateral 2DkV images and matched to the same feature in the DRRs. The selected ROIs mimic the ROIs normally used by radiation oncologists (ROs), which usually include isocenter, C2, and some other bony structures near the gross target volume. AP and superoinferior (SI) shifts can be calculated from the lateral 2DkV/DRR images, whereas SI and right-left (RL) shifts can be calculated from the AP 2DkV/DRR images. The three shifts are determined where the maximum combined MI for the two pairs of images is achieved. For both CAT and MI software tools, only translational shifts were considered in this work. We used a head phantom, as shown in Fig. 1, to test the reliability of both CAT and MI software in performing 3D and 2D registration. In this assessment, we performed the end-to-end procedure just as if we would do for patients, including CT simulation, treatment planning to the final phantom setup, and imaging on the treatment couch. A custom vacuum bag was made for the head phantom to improve repositioning precision throughout the study. The ‘‘planning CT’’ images were acquired on a CT-LINAC system (EXaCT, Varian Oncology Systems) (17). The spatial resolution for each CT image was 0.94 mm per pixel in the transverse plane, and the slice thickness was 0.3 cm. Fourteen sets of couch position shifts of the head phantom were performed by shifting the couch along the AP, SI, and RL directions for up to 2.5 cm in each direction, and 2DkV and CBCT images were acquired for each set of couch positions. Shifts calculated by the CAT and MI software were compared with the actual shifts.

918

I. J. Radiation Oncology d Biology d Physics

Volume 71, Number 3, 2008

Patients and imaging procedure

Statistical methods

Twenty-one consecutive H&N patients scheduled to receive IMRT were evaluated. All patients were immobilized using thermoplastic masks with either a patient-specific vacuum bag (11 of 21, IMA) or standard clear plastic (10 of 21, IMB) headrests (both from Medtec, Orange City, IA). A typical patient immobilized with thermoplastic masks and patient-specific vacuum bag is shown in Fig. 2. A foot brace device was used for all patients. Three radiopaque markers were permanently attached to the face masks (one anterior and two lateral) guided by the CT lasers to mark a reference point. The planning CT of each patient was acquired on an AcQSim PQ5000 CT simulator (Philips Medical Systems, Cleveland, OH). The treatment planning CT scan began above the top of the skull and continued to below the clavicle with a slice thickness of 2.5 mm. CT images were imported into the treatment planning system (Pinnacle; Philips Medical Systems, Bothell, WA) to generate a treatment plan. Before each imaging and treatment session, three radiopaque markers on the patient’s thermoplastic mask were aligned with room lasers. CBCT procedures were performed for each patient immediately after taking a pair of 2DkV images in his or her treatment position (25). Only 2DkV and CBCT images acquired during the same treatment session were included, yielding a total of 98 pairs of assessable 2DkV-CBCT alignment sets (a median of 5 image sets; range, 2–7 per patient). Both CBCT and 2DkV images were normally to be acquired weekly. However, for a particular week, CBCT might not be acquired or might be acquired on the different treatment section from 2DkV images, the assessable numbers of image sets would vary. This imaging schedule was used because it minimally disrupted the existing workflow of patient treatment management. The treating RO reviewed 2DkV images through visual comparisons with reference DRR and determined the treatment shifts. The CAT and MI software tools were only used for this project (not for patient treatments). The C2 vertebral body (C2) was used as the surrogate by the CAT software tool to align CBCT images to the planning CT. C2 was selected because it represented the central region of the H&N area and was also close to the anatomic pivot for head turns (16). Before the first day of treatment, planning CT images as well as DICOM RT objects containing the isocenter and contours for target volumes and normal tissues (including C2) were imported into the workstation with CAT software; DRR images were exported to a recording and verifying system (IMPAC Medical Systems, Sunnyvale, CA) and to a computer with MI software.

We followed the definition and notations of setup errors described elsewhere (16). If we denote patients as p ˛{p1; p2; .pM}, and image sections for each patient as f ˛{1; 2; .; N}, then the setup AP SI RL error Epf ¼ ½Epf ; Epf ; Epf can be written as Epf ¼ Sp þ Rpf , where Sp is the systematic error for each patient, whereas Rpf is the random error introduced in each image section. The patient (p)-specific random error sp is the standard deviation (SD) of Rpf over f. The population systematic and random variations are then calculated by the following equations (16),

Sp ¼

N X

Epf =N

f ¼1

m¼

pM X

Sp =M

p¼p1

X

pM 2 1 X Sp  m M p¼p1

¼

sp ¼

!1=2

N  2 1X Rpf  Rpf N f ¼1

1X 2 s M p¼p1 p pM

s¼

!1=2

!1=2

P where m is the population mean of systematic variations, is the standard deviation of the systematic variations, and s is the average root-mean-square of individual random variations. Margins resulting from setup uncertainty alone could then be estimated from the systematic and random variations. Van Herk has provided a review of margin formulae (26), and the commonly used form is M ¼ 2:5S þ 0:7s. Margins determined by this equation assume that the minimum dose to CTV is 95% for 90% of the patients (27). The 90% confidence range, which is lower bounded by the 5th percentile and upper bounded by the 95th percentile, could also indicate the setup margin (16). Pearson’s correlations were used to assess the similarity of results between 2D and 3D alignments. We also used t test to determine if there was difference between the population mean shifts of two different immobilization schemes.

RESULTS

Fig. 2. A patient was immobilized with a face mask and patient-specific vacuum bag on the computed tomography simulator couch.

Verification of MI and CAT software with the head phantom The effectiveness of the registration software was first verified using a head phantom, shown in Fig. 1, with known translational couch shifts. This represents the ideal situation; that is, a rigid body with no curvature change or rotation. The results of 2DkV registration using the MI software tool for 14 sets of couch shifts of the head phantom are summarized in Table 1. The ranges of couch shift errors (0.12, 0.10),

Comparison of 2D radiographic images and 3D CBCT d H. LI et al.

Table 1. Validation of MI software for 2DkV image registration in AP, SI, and RL directions AP (cm)

SI (cm)

RL (cm)

Range of real couch (2.5, 2.5) (2.5, 2.5) (1.5, 1.5) shifts Range of couch shift (0.12, 0.10) (0.08, 0.11) (0.14, 0.12) error determined by MI software Mean couch shift 0.03 0.01 0.02 error SD of couch shift 0.06 0.05 0.08 error Abbreviations: MI = mutual information; 2DkV = twodimensional kilovoltage; SD = standard deviation; AP = anteroposterior; SI = superoinferior; RL = right-left.

(0.08, 0.11), and (0.14, 0.12) cm and the mean and standard deviation errors were –0.03  0.06, –0.01  0.05, and 0.02  0.08 cm in the AP, SI, and RL directions, respectively. Similar results were obtained for CBCT image registration using CAT software for the same sets of couch shifts, –0.06  0.04, –0.02  0.07, and 0.00  0.03 cm in the AP, SI, and RL directions, respectively (25). The phantom results provide the limit for what we could do for the patients. The maximum difference between couch shifts calculated by the MI software and the actual couch shifts was less than 1.5 mm. The mean couch shift error and the systematic error introduced by the MI software were both less than 1 mm. The CAT and MI software tools are accurate when only translational positioning errors are introduced. However, that the accuracy of the validations is limited by the resolution of couch digital readouts, which is on the order of 1 mm. Distributions of setup shifts The results from the study of setup shifts in patients are summarized in Table 2, including the population meanP setup error m (cm) and their ranges, the standard deviations ( ) of the corresponding systematic shift, and the random component (s) in all three directions for the entire patient population. Both the systematic and random errors were small (<1.6 mm) for all modalities. Figure 3 shows the distribution of couch shifts in all three directions. Among all CBCT couch shifts, 4.1% were $0.5 cm and 18.7% were $0.3 cm; among

919

all MI shifts, 1.7% was $0.5 cm, 11.2% were $0.3 cm; whereas among all RO shifts, 1.6% were $0.5 cm, and 10.5% were $0.3 cm. Margins The margins calculated with Van Herk’s margin formula using systematic and random variations are tabulated in Table 3. The 90% confidence ranges listed in Table 4 are defined by the position of the 5th and 95th percentiles and have been calculated for CBCT, MI, and RO couch shifts in the AP, SI, and RL directions in all cases. These values could be used for understanding the residual setup uncertainties in the treatment process and also as a guideline for designing margins (16). In general, because the CBCT images represent the 3D anatomy and more closely depict the actual anatomy than do the 2DkV image pairs; we could regard the 3D CBCT couch shifts as the ‘‘reference shifts’’ and attempt to determine the residual errors after shifting the couch, as directed by the 2DkV-guided techniques, and assess the difference P between 2D and 3D alignment techniques. The m, , and s of the residual error with RO or MI correction are listed in Table 5. With couch positions corrected by the 2DkV images, margins needed to account for residual errors are tabulated in Table 6, and the 90% confidence ranges for residual errors are shown in Table 7. Correlations of 2D and 3D setup shifts Figure 4 demonstrates the correlation between couch shifts predicted by CBCT and 2DkV image registrations. The horizontal axis represents the couch shifts specified by the CAT software, whereas the vertical axis represents 2DkV image couch shifts specified by radiation oncologists (open circles) or by MI software for each image session. The diagonal line represents equal couch shift output predicted by the 3D and 2D registration techniques. There were strong correlations between CAT registration for CBCT and MI software registration for 2DkV images, but the correlation between CAT and RO alignments were much weaker, though still statistically significant, as listed in Table 8. The poor correlation between RO shifts and CBCT/MI shifts is mainly due to the fact that there are many zeros or no shifts in the RO data. The correlation among CBCT, MI, and RO shifts for data points where RO shifts were not zero (17 AP, 11 RL,

Table 2. Population mean setup error (range), systematic shift SDs, and average random shifts in AP, SI, and RL directions for all patients Population mean setup error m (cm) (range)

CBCT 2DkV RO 2DkV MI

Systematic shift SDs S (cm)

Average random shift sRMS (cm)

mAP

mSI

mRL

SAP

SSI

SRL

sAP

sSI

sRL

0.11 (0.84, 0.40) 0.02 (0.50, 0.50) 0.02 (0.90, 0.60)

0.01 (0.50, 0.57) 0.01 (0.40, 0.57) 0.02 (0.58, 0.57)

0.02 (0.41, 0.52) 0.01 (0.30, -0.50) 0 (0.37, 0.40)

0.16 0.05 0.11

0.15 0.06 0.10

0.13 0.06 0.12

0.16 0.12 0.16

0.16 0.12 0.15

0.12 0.07 0.12

Abbreviations: CBCT = cone beam CT; RO = radiation oncologist; other abbreviations as in Table 1.

I. J. Radiation Oncology d Biology d Physics

Volume 71, Number 3, 2008

100%

100%

75%

75%

75%

50%

Percent

100%

Percent

AP

Percent

920

50%

25%

25%

25%

0%

0% -0.75

-0.50

-0.25

0.00

-0.50

0.25

75%

75%

75%

50%

Percent

100%

50%

0.25

0.50

-0.25

0.00

0.25

-0.20

0.50

100%

75%

75%

75%

Percent

100%

Percent

100%

50%

0.00

0.25

0.50

0% -0.20

RL_CBCT (cm)

0.00

0.20

0.40

-0.20

0.00

0.20

0.40

RL_RO (cm)

RL_MI (cm)

(b) Shifts from MI software

(a) Shifts from CAT software

0.40

50%

0% -0.25

0.20

25%

25%

0%

0.00

SI_RO (cm)

SI_MI (cm)

SI_CBCT (cm)

25%

0.50

0% -0.50

50%

0.25

25%

0% 0.00

0.00

50%

25%

-0.25

-0.25

AP_RO (cm)

100%

0% -0.50

Percent

0% -0.50

0.50

100%

25%

RL

0.00

AP_MI (cm)

Percent

Percent

AP_CBCT (cm)

SI

50%

(c) Shifts from radiation oncologist

Fig. 3. Distribution of setup shifts for all 21 patients (98 image sections) in anteroposterior (AP), superoinferior (SI), and right-left (RL) directions.

and 26 SI shifts) are tabulated in Table 9. The correlations in this case are much stronger among all three techniques. Comparison of immobilization techniques Figure 5 shows histogram plots for the two groups of patients using two different immobilization techniques, IMA and IMB, and Table 10 summarizes the setup shifts. For the CBCT and 2DkV registration performed by MI software, the systematic variations S for IMA (0.06–0.11 cm) are less than that for IMB (0.15–0.22 cm) in all directions. However, there was a slight increase in the random shift error for

IMA of from 0.01 to 0.06 cm. The population mean setup error m in the AP direction from 3D CBCT registration shows a statistically significant difference between IMA and IMB, which has been confirmed by the t test (t = 3.180, p = 0.003), but no statistical difference can be identified for the 2D registration methods. The setup margins calculated with Van Herk’s margin formula for IMA and IMB using 2D and 3D alignment techniques are shown in Table 11. Table 4. The 90% confidence range for shifts of different alignment techniques

Table 3. Margins for different alignment techniques

AP SI RL

CBCT (cm)

2DkV RO (cm)

2DkV MI (cm)

0.51 0.49 0.41

0.21 0.23 0.20

0.39 0.36 0.38

Abbreviations as in Table 1 and 2.

CBCT (cm) AP I90 SI I90 RL I90

(0.51, 0.26) (0.44, 0.32) (0.27, 0.34)

2DkV RO (cm)

2DkV MI (cm)

(0.31, 0.21) (0.30, 0.30) (0.20, 0.30)

(0.38, 0.30) (0.32, 0.31) (0.30, 0.25)

Abbreviations: I90 = the 90% confidence range; other abbreviations as in Table 1 and 2.

Comparison of 2D radiographic images and 3D CBCT d H. LI et al.

921

Table 5. Distributions of residual errors after MI/RO correction Population mean setup error m (cm) (range) mAP

mSI

Systematic shift SDs S (cm) mRL

2DkV RO 0.10 (0.84, 0.32) 0.01 (0.77, 0.78) 0.02 (0.30, –0.50) 2DkV MI 0.09 (0.77, 0.78) 0.01 (0.58, 0.57) 0.03 (0.37, 0.40)

Average random shift sRMS (cm)

SAP

SSI

SRL

sAP

sSI

sRL

0.17 0.13

0.14 0.12

0.10 0.08

0.15 0.15

0.17 0.12

0.11 0.10

Abbreviations as in Table 2.

DISCUSSION 2D and 3D alignment results In the setup uncertainty analysis, both the systematic and random errors were small (<1.6 mm) for all modalities (Table 2), indicating that there were no significant systematic errors between treatment simulation and the actual treatment delivery and that the positioning techniques were consistent throughout the treatment period. The stability of the treatment delivery was confirmed by the histogram plots in Fig. 3. The RO 2DkV shifts yielded the smallest systematic shifts. This is probably because ROs tend to accept the couch position for treatment, or no shift, if the setup error is less than 2–3 mm in all directions (28). This practice yields many zeros shifts, as illustrated in Fig. 4. We also found the statistical findings for C2 setup shifts using CBCT are consistent with our previous study for a different group of patients using CT-on-rails (16), indicating the consistency of immobilization, localization, and registration approaches between current and previous studies. The correlation between CBCT and MI shifts is strong (Table 8) as expected because both methods use similar bony structure registration techniques. On the other hand, the correlation between ROs and the other two techniques are much weaker. This is again because ROs tend to accept small setup errors. If the correlations were calculated only for data points where RO shifts were non-zero (17 AP, 11 RL, and 26 SI shifts), the correlations between RO and the other two methods become much stronger, especially between RO and MI (the same image sets are used). The correlation between 2D and 3D registration techniques is stronger on RL and AP direction than SI direction, which might be due to the fact that planning CT has lower resolution on SI direction and therefore higher uncertainty. The head phantom study confirms that 2DkV and CBCT images are equivalent for determining the setup deviations if the subject is a rigid body without deformation and rotation

Table 6. Margins for residual errors after 2DkV MI/RO correction

AP SI RL

2DkV RO (cm)

2DkV MI (cm)

0.53 0.47 0.33

0.43 0.38 0.27

Abbreviations as in Table 1 and 2.

(Table 1). For patients, however, although there indeed was a strong correlation between the shifts predicted from CBCT and 2DkV images, we found that there were cases where setup errors (up to 0.8 cm) appeared in the CBCT alignment but were not recognized in the 2DkV images. We believe that such large differences in 3D and 2D registration were mainly due to the presence of nonrigid or rotational movements of H&N anatomy, which are hard to track in 2DkV images. Shown in Fig. 6 is an example of a change in patient neck curvature between the simulation and one of the image/treatment sessions. All contours were generated on the planning CT in the treatment planning system and contours overlaid on CBCT were based on no shift (RO shift) and C2 alignment using CAT software. When no shift is applied to the CBCT image, the C2 and the spinal cord contours nearby are clearly off in the AP direction, whereas the spinal cord contour in the lower section neck (e.g., C6) matches well with the anatomy. After C2 alignment, the spinal cord contour is partially overlapped by vertebras in the low neck region on the CBCT image. For this particular case, the C2 alignment using CAT software predicted a 0.6 cm couch shift up, whereas the treating RO did request the patient be shifted, and MI software predicted a less than 0.1 cm shift up. Margins With systematic and random errors listed in Table 2, the maximum margins from setup uncertainty for 2D and 3D alignment techniques are: 0.51 cm for CBCT, 0.23 cm for RO, and 0.39 cm for MI, respectively. These estimated margins are consistent with the 90% confidence ranges as shown in Table 4. With CBCT shifts as the ‘‘reference,’’ we can also calculate the residual errors after MI or RO shifts, the margins needed to account for the residual error, and the 90% confidence ranges of the residual errors, as tabulated in Tables 5–7. The results suggested that, with CBCT as reference, Table 7. 90% confidence range of residual errors after 2DkV MI/RO correction

IRAP 90 IRSI 90 IRRL 90

2DkV RO (cm)

2DkV MI (cm)

(0.51, 0.26) (0.44, 0.32) (0.22, 0.27)

(0.40, 0.20) (0.41, 0.28) (0.20, 0.25)

Abbreviations: IR90 = the 90% confidence range of residual errors; other abbreviations as in Table 1 and 2.

I. J. Radiation Oncology d Biology d Physics

922

1.0

Volume 71, Number 3, 2008

AP_CBCT AP_RO AP_CBCT AP_MI

AP

1.0

0.5

SI OBI shifts (cm)

AP OBI shifts (cm)

0.5

SI_CBCT SI_RO SI_CBCT SI_MI

SI

0.0

0.0

-0.5

-0.5

-1.0

-1.0 -1.0

-0.5

0.0

0.5

1.0

-1.0

-0.5

AP CBCT shifts (cm)

0.0

0.5

1.0

SI CBCT shifts (cm) RL_CBCT RL_RO RL_CBCT RL_MI

1.0

RL RL OBI shifts (cm)

0.5

0.0

-0.5

-1.0 -1.0

-0.5

0.0

0.5

1.0

RL CBCT shifts (cm)

Fig. 4. Correlation of setup shifts between cone beam computed tomography, 2DkV mutual information, and 2DkV radiation oncologist registration for all image sections in anteroposterior (AP), superoinferior (SI), and right-left (RL) directions.

0.4–0.5 cm setup uncertainty margin is still necessary even with MI or RO shifts being administrated. These results again indicate that the difference between 3D CBCT and 2DkV alignments are significant. Discrepancy between 2D and 3D techniques Factors that can lead to treatment setup uncertainties have been discussed elsewhere (16, 22): difference between simulation and treatment, the random setup shifts in daily positioning and residual errors after corrections are executed, and anatomic variation. In this study, we focused on the uncertainties that can lead to differences between 2D and

3D image alignment, mainly the positioning and correction errors. CBCT used C2 vertebral body consistently because it is well defined in 3D, whereas 2DkV used a region that will have the most visible bony landmark for each projection radiographs, which may not be the same bony landmark on two radiographs and may be different from the ROI used for CBCT. In a previous study of setup uncertainties in H&N patients using 3D CT images from the CT-on-rails system, variations were found for setup corrections using different regions of the H&N bony structures (C2, C6, and the palatine process of the maxilla) (16). For 2DkV images, additional uncertainty was introduced by the limitation of

Table 8. Pearson’s correlation of couch shifts resulting from different registration techniques (p < 0.001 for all cases)

Table 9. Pearson’s correlation of couch shifts resulting from different registration techniques, for cases with RO shifts >0 only (p < 0.05 for all cases)

Correlation AP Correlation SI Correlation RL

CBCT and 2DkV MI

CBCT and 2DkV RO

2DkV RO and MI

0.589 0.691 0.738

0.361 0.398 0.426

0.511 0.433 0.343

Abbreviations as in Table 1 and 2.

Correlation AP Correlation SI Correlation RL

CBCT and 2DkV MI

CBCT and 2DkV RO

2DkV RO and MI

0.718 0.692 0.907

0.692 0.571 0.857

0.915 0.659 0.875

Abbreviations as in Table 1 and 2.

Comparison of 2D radiographic images and 3D CBCT d H. LI et al.

AP

SI

40

40 IMA IMB

IMA IMB

35

30

30

25

25

Counts

Counts

35

20

20

15

15

10

10

5

5

0

923

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

0

1

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

SI CBCT shifts (cm)

AP CBCT shifts (cm)

RL 40 IMA IMB

35 30

Counts

25 20 15 10 5 0

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

RL CBCT shifts (cm)

Fig. 5. Two immobilization techniques are compared. The distribution of shifts indicate that IMA (patient-specific vacuum bag, red) has a slightly smaller range compared with IMB (standard clear plastic, blue).

projection images: out-of-plane rotations were not identifiable from 2D projections in general and even bony anatomic structures could be difficult to be distinguished in 2D projection images if they were obscured by other anatomic features, for example, overlapping bony structures. CBCT alignment using C2 as the alignment target could be more reliable than 2D techniques for tumors near C2 such as posterior pharyngeal wall cancers or chordomas. In addition to targeting the tumor volumes, protecting spinal cord is also

critically important in H&N treatment, particular for complex dose distributions delivered by IMRT. Using vertebra such as C2 as the alignment target ensures spinal cord near the alignment vertebra to be properly spared. Immobilization techniques In the comparison of two immobilization techniques, we found no statistical difference between the two techniques with 2DkV alignment, but reduction in population mean

Table 10. Comparison of different immobilization techniques population mean setup error m (cm) (range)

IMA (CBCT) IMB (CBCT) IMA (2DkV RO) IMB (2DkV RO) IMA (2DkV MI) IMB (2DkV MI)

Systematic shift SDs S (cm)

Average random shift sRMS (cm)

mAP

mSI

mRL

SAP

SSI

SRL

sAP

sSI

sRL

0.04 (0.50, 0.40) 0.17 (0.84, 0.26) 0.03 (0.50, 0.50) 0.01 (0.50, 0.30) 0.01 (0.90, 0.60) 0.03 (0.57, 0.42)

0.03 (0.50, 0.36) 0.00 (0.45, 0.57) 0.02 (0.27, 0.38) 0 (0.41, 0.52) 0 (0.45, -0.46) 0.04 (0.58, 0.57)

0.06 (0.27, –0.38) 0.00 (0.41, 0.52) 0 (0.30, 0.30) 0 (0.20, 0.50) 0.02 (0.35, 0.30) 0.01 (0.37, 0.40)

0.08 0.18 0.06 0.05 0.11 0.22

0.06 0.19 0.04 0.08 0.09 0.18

0.07 0.15 0.04 0.07 0.07 0.18

0.18 0.14 0.16 0.08 0.19 0.13

0.19 0.14 0.12 0.13 0.16 0.14

0.12 0.11 0.07 0.07 0.13 0.10

Abbreviations: IMA = immobilization technique A (patient-specific vacuum bag); IMB = immobilization technique B (standard clear plastic); other abbreviations as in Table 1 and 2.

I. J. Radiation Oncology d Biology d Physics

924

Volume 71, Number 3, 2008

Table 11. Margins for different immobilization techniques

AP SI RL

IMA/CBCT (cm)

IMB/CBCT (cm)

IMA/2DkV RO (cm)

IMB/2DkV RO (cm)

IMA/2DkV MI (cm)

IMB/2DkV MI (cm)

0.33 0.28 0.26

0.55 0.57 0.45

0.26 0.18 0.15

0.18 0.29 0.22

0.41 0.34 0.27

0.64 0.55 0.52

Abbreviations as in Tables 1, 2, and 10.

setup error in AP direction with the patient-specific headrest assessed by CBCT. Our hypothesis is that patient-specific headrests reduce the patient specific variation in neck curvature (on AP direction), thereby reducing the errors in the AP direction. This result indicates that CBCT alignment being more sensitive in detecting setup errors than 2DkV image alignment techniques. The smaller values of systematic variation (S) for IMA than IMB, determined by CBCT and 2DkV MI, indicate that the patient-specific vacuum bag is a more stable immobilization. This is confirmed by the histogram plots in Fig. 5 that show a comparison of the distribution of shifts of between two immobilizations. In addition, because IMA has a smaller S on all directions than IMB, it could be expected that the setup uncertainty margin needed for IMA is smaller than IMB, as listed in Table 11. The maximum margins calculated with Van Herk’s margin formula for IMA are 0.33 cm for CBCT, 0.26 cm for 2DkV RO, and 0.41 cm for 2DkV MI; the margins for IMB are 0.57 cm for CBCT, 0.29 cm for 2DkV RO, and 0.64 cm for 2DkV MI. Therefore, regardless of imaging modality, to further reduce the setup uncertainties, it is essential to improve immobilization method, which appears more important than the methods of image alignment. A good immobilization device could potentially reduce the flexibility of the H&N region and the magnitude of rotations. As a result, 3D and 2D alignment

could be equivalent as in the case of head phantom. Radiopaque fiducials, in additional to immobilization, could be potentially useful localization method and will be studied in the future (29, 30).

CONCLUSIONS In conclusion, we have found that 2DkV and 3D CBCT alignments are highly correlated for H&N patients, but not equivalent. The discrepancy is most likely caused by that the H&N region is not a rigid body and 2D and 3D alignments use somewhat different ROIs. Rotations could also contribute to the differences, but was not specifically considered in this study. The current results are consistent with our previous findings of variability in setup corrections for different regions of H&N anatomy using CT-on-rails (16). It is clear better localization (2D or 3D alignment) alone would not completely solve the problem for H&N patients. Better immobilization that would reduce the variation of the curvature of the neck and rotation would be logically the next step to further reduce the discrepancy between 2D and 3D alignments and the setup errors for H&N patients. Patientspecific vacuum bag based headrests that slightly reduce shifts in the AP direction is promising and require further investigation.

Fig. 6. Example of patient neck curve change between simulation and treatment illustrated by sagittal images. Contours were created on the planning computed tomography (CT) in the treatment planning system. (A) Planning CT images with selected contours, (B) contours overlaid on cone beam CT with no shift, and (C) with shifts based on C2 alignment using CT-assisted targeting software.

Comparison of 2D radiographic images and 3D CBCT d H. LI et al.

925

REFERENCES 1. Webb S. Intensity-modulated radiation therapy. Bristol (UK): Institute of Physics Publishing; 2001. 2. Palta JR, Mackie TR, editors. Intensity-modulated radiation therapy—the state of the art. Madison (WI): Medical Physics Publishing; 2003. 3. Zeidan OA, Langen KM, Meeks SL, et al. Evaluation of imageguidance protocols in the treatment of head and neck cancers. Int J Radiat Oncol Biol Phys 2007;67:670–677. 4. Boda-Heggemann J, Walter C, Rahn A, et al. Repositioning accuracy of two different mask systems-3D revisited: Comparison using true 3D/3D matching with cone-beam CT. Int J Radiat Oncol Biol Phys 2006;66:1568–1575. 5. Hong TS, Tome WA, Chappell RJ, et al. The impact of daily setup variations on head-and neck intensity modulated radiation therapy. Int J Radiat Oncol Biol Phys 2005;61:779–788. 6. Prisciandaro JI, Frechette CM, Herman MG, et al. A methodology to determine margins by EPID measurements of patient setup variation and motion as applied to immobilization devices. Med Phys 2004;31:2978–2988. 7. Manning MA, Wu Q, Cardinale RM, et al. The effect of setup uncertainty on normal tissue sparing with IMRT for headand-neck cancer. Int J Radiat Oncol Biol Phys 2001;51: 1400–1409. 8. Hatherly KE, Smylie JC, Rodger A, et al. A double exposed portal image comparison between electronic portal imaging hard copies and port films in radiation therapy treatment setup confirmation to determine its clinical application in radiotherapy center. Int J Radiat Oncol Biol Phys 2001;49:191–198. 9. Bel A, Keus R, Vijlbrief RE, et al. Setup deviations in wedged pair irradiation of parotid gland and tonsillar tumors, measured with an electronic portal imaging device. Radiother Oncol 1995;37:153–159. 10. Willner J, Hadinger U, Neumann M, et al. Three dimensional variability in patient positioning using bite block immobilization in 3D-conformal radiation treatment for ENT-tumors. Radiother Oncol 1997;43:315–321. 11. Karger CP, Jakel O, Debus J, et al. Three-dimensional accuracy and interfractional reproducibility of patient fixation and positioning using stereotactic head mask system. Int J Radiat Oncol Biol Phys 2001;49:1493–1504. 12. Brock KK, McShan DL, Balter JM. A comparison of computer controlled versus manual on-line patient setup adjustment. J Appl Med Phys 2002;3:241–247. 13. de Boer HC, van Sornsen de Koste JR, et al. Electronic portal image assisted reduction of systematic set-up errors in head and neck irradiation. Radiother Oncol 2001;61:299–308. 14. Bel A, Petrascu O, Van de Vonde I, et al. A computerized remote table control for fast on-line patient repositioning: implementation and clinical feasibility. Med Phys 2000;27: 354–358. 15. Pisani L, Lockman D, Jaffray D, et al. Setup error in radiotherapy: On-line correction using electronic kilovoltage and mega-

16. 17. 18. 19.

20. 21. 22. 23.

24. 25. 26. 27.

28.

29.

30.

voltage radiographs. Int J Radiat Oncol Biol Phys 2000;47: 825–839. Zhang L, Garden AS, Lo J, et al. Multiple regions-of-interest analysis of setup uncertainties for head-and-neck cancer radiotherapy. Int J Radiat Oncol Biol Phys 2006;64:1559–1569. Court L, Rosen I, Mohan R, et al. Evaluation of mechanical precision and alignment uncertainties for an integrated CT/ LINAC system. Med Phys 2003;30:1198–1210. O’Daniel JC, Dong L, Zhang L, et al. Dosimetric comparison of four target alignment methods for prostate cancer radiotherapy. Int J Radiat Oncol Biol Phys 2006;66:883–891. Borst GR, Sonke JJ, Betgen A, et al. Kilo-voltage cone-beam computed tomography setup measurements for lung cancer patients; first clinical results and comparison with electronic portal-imaging device. Int J Radiat Oncol Biol Phys 2007;68:555–561. Court LE, Dong L. Automatic registration of the prostate for computed-tomography-guided radiotherapy. Med Phys 2003; 30:2750–2757. Bijhold J, Lebesque JV, Hart AA, et al. Maximizing setup accuracy using portal images as applied to a conformal boost technique for prostatic cancer. Radiother Oncol 1992;24:261–271. Mubata CD, Bidmead AM, Ellingham LM, et al. Portal imaging protocol for radical dose-escalated radiotherapy treatment of prostate cancer. Int J Radiat Oncol Biol Phys 1998;40:221–231. Zhang L, Dong L, Court L, et al. Validation of CT-assisted targeting (CAT) software for soft tissue and bony target localization. The American Association of Physicists in Medicine (AAPM) 47th Annual Meeting. Seattle, WA; 2005. Pluim JPW, Maintz JBA, Viergever MA. Mutual information based registration of medical images: a survey. IEEE Trans Med Imaging 2003;22:986–1004. Zhu XR, Zhang L, Wu R, et al. Clinical Implementation of cone beam computed tomography for imaging guided radiation therapy. Int J Radiat Oncol Biol Phys. In press. van Herk M. Errors and margins in radiotherapy. Semin Radiat Oncol 2004;14:52–64. van Herk M, Remeijer P, Rasch C, et al. The probability of correct target dosage: Dose-population histograms for deriving treatment margins in radiotherapy. Int J Radiat Oncol Biol Phys 2000;47:1121–1135. Lam KL, Balter JM, Haken RKT. Effect of daily localization and correction on the setup uncertainty: Dependences on the measurement uncertainty, re-positioning uncertainty and action level. Phys Med Biol 2007;52:6575–6587. Asselen B, Dehnad H, Raaijmakers C, et al. Implanted gold markers for position verification during irradiation of headand-neck cancers: a feasibility study. Int J Radiat Oncol Biol Phys 2004;59:1011–1017. Oita M, Ohmori K, Obinata K, et al. Uncertainty in treatment of head-and-neck tumors by use of intraoral mouthpiece and embedded fiducials. Int J Radiat Oncol Biol Phys 2006;64: 1581–1588.