- Email: [email protected]
Feasibility Study for Markerless Tracking of Lung Tumors in Stereotactic Body Radiotherapy
Int. J. Radiation Oncology Biol. Phys., Vol. 78, No. 2, pp. 618–627, 2010 Copyright Ó 2010 Elsevier Inc. Printed in the USA. All rights reserved 0360-3016/$–see front matter
doi:10.1016/j.ijrobp.2009.11.028
PHYSICS CONTRIBUTION
FEASIBILITY STUDY FOR MARKERLESS TRACKING OF LUNG TUMORS IN STEREOTACTIC BODY RADIOTHERAPY ANNE RICHTER, M.SC., JUERGEN WILBERT, PH.D., KURT BAIER, M.SC., MICHAEL FLENTJE, M.D., AND MATTHIAS GUCKENBERGER, M.D. University of Wuerzburg, Department of Radiation Oncology, Wuerzburg, Germany Purpose: To evaluate the feasibility and accuracy of a method for markerless tracking of lung tumors in electronic portal imaging device (EPID) movies and to analyze intra- and interfractional variations in tumor motion. Methods and Materials: EPID movies were acquired during stereotactic body radiotherapy (SBRT) given to 40 patients with 49 pulmonary targets and retrospectively analyzed. Tumor visibility and tracking accuracy were determined by three observers. Tumor motion of 30 targets was analyzed in detail via four-dimensional computed tomography (4DCT) and EPID in the superior-inferior direction for intra- and interfractional variations. Results: Tumor visibility was sufficient for markerless tracking in 47% of the EPID movies. Tumor size and visibility in the DRR were correlated with visibility in the EPID images. The difference between automatic and manual tracking was a maximum of 2 mm for 98.3% in the x direction and 89.4% in the y direction. Motion amplitudes in 4DCT images (range, 0.7-17.9 mm; median, 4.9 mm) were closely correlated with amplitudes in the EPID movies. Intrafractional and interfractional variability of tumor motion amplitude were of similar magnitude: 1 mm on average to a maximum of 4 mm. A change in moving average of more than ±1 mm, ±2 mm, and ±4 mm were observed in 47.1%, 17.1%, and 4.5% of treatment time for all trajectories, respectively. Mean tumor velocity was 3.4 mm/sec, to a maximum 61 mm/sec. Conclusions: Tracking of pulmonary tumors in EPID images without implanted markers was feasible in 47% of all treatment beams. 4DCT is representative of the evaluation of mean breathing motion on average, but larger deviations occurred in target motion between treatment planning and delivery effort a monitoring during delivery. Ó 2010 Elsevier Inc. Tumor motion, Markerless tracking, Interfractional, Intrafractional, EPID.
METHODS AND MATERIALS
INTRODUCTION Temporal changes in the patients’ anatomy due to breathing motion may reduce the accuracy of radiotherapy treatment, especially for hypofractionated stereotactic body radiotherapy (SBRT). Precise assessment and quantification of breathinginduced target motion and its integration into the treatment workflow are essential for adaptive treatment techniques (1). Existing strategies for tumor tracking approaches can be grouped roughly into three categories: external surrogate based (2, 3), markerless (4–6), and internal marker-based systems (7–9). The aim of this study was to investigate the feasibility of imaging and tracking lung tumors, using the Mega Voltage (MV) therapy beam and the electronic portal imaging device (EPID) without implanted markers.
Data acquisition and patient population
Reprint requests to: Anne Richter, M.Sc., Department of Radiation Oncology, University of Wuerzburg, Josef-Schneider-Str. 11, 97080 Wuerzburg, Germany. Tel: 49 (0)931 201-28881; Fax: 49 (0)931 201-28221; E-mail: [email protected] Anne Richter and Juergen Wilbert contributed equally to this work. This work was partially supported by the Bayerische Forschungsstiftung, Germany.
This work was presented in part at the Medical Physics and Biological Engineering World Congress September 7–12, 2009, Munich, Germany. Conflict of interest: none. Acknowledgment—We thank Mark Gainey for proofreading the manuscript. Received June 8, 2009, and in revised form Nov 3, 2009. Accepted for publication Nov 16, 2009.
Data are based on 40 consecutive lung cancer patients with 49 pulmonary targets treated with image-guided SBRT in our department (patient and treatment characteristics are shown in Table 1). For each patient, a four-dimensional computed tomography (4DCT) scan was acquired for treatment planning (Somatom Sensation Open, Siemens, Forchheim, Germany) without breath coaching. Abdominal compression was used for 20 of the 40 patients. All patients were treated on an Elekta Synergy S (Elekta, Crawley, UK) equipped with a Kilo Voltage cone beam system and an MV EPID. Patients were positioned in the stereotactic body frame or a BodyFix system (Elekta, Crawley, UK). Treatment position was verified via cone beam CT imaging prior to all treatment fractions. Details of the planning and treatment technique have been described
618
Markerless tumor tracking d A. RICHTER et al.
Table 1. Patient population* Characteristics
Minimum
Maximum
Median
Patient age (a) A 4DCT (mm) GTV volume (cm3) Number of fractions
30 0.7 0.3 1
86 18.8 126 8
66.5 4.4 9.5 3
* Characteristics for the initial patient population with 49 lesions. previously (10). EPID movies were acquired during delivery of the beams, resulting in 668 EPID movies within 278 fractions.
Evaluation of tumor visibility Digitally reconstructed radiographs (DRR) were generated for each beam based on the planning CT study using Pinnacle3 planning software (Philips Radiation Oncology Systems, Fitchburg, WI). Parameters (energy and brightness) were adapted to obtain image quality similar to that of EPIDs. The macroscopic tumor was delineated in 4DCT images by an MD. The contour of the macroscopic tumor was overlaid and stored with these DRR images. The visibility of the tumor in these DRRs (DRR visibility) was evaluated retrospectively using a scheme with three rating levels (0, not visible; 1, partially visible; 2, clearly visible). DRRs and EPID frames were displayed side by side to facilitate the identification of the pulmonary tumor in the EPID images. The visibility of the pulmonary target in EPID images (MV visibility) was investigated in all 668 EPID movies, separately. A scheme with three rating levels was introduced for the evaluation of tumor visibility: a rating of 2 represented complete or partial tumor shape that was identifiable in all EPID images of the beam; a rating of 1 represented a complete or partial tumor shape identifiable in some but not all images; and a rating of 0 represented a tumor that could not be clearly identified. All EPID images were evaluated retrospectively by three independent observers (physicists and radiographers). The final result for each image sequence was defined as the mean rating by the three observers. As the number of fractions was different for the tumors, the mean rating for the first fraction was used for evaluation of the overall tumor visibility. The maximum deviation between the mean rating of the first fraction and of each following fraction was calculated to evaluate the reproducibility of MV visibility. Statistica 6.0 software (Statsoft, Tulsa, OK) was utilized for statistical analysis. A Pearson c2 test was used for analyzing the rela-
619
tionship between MV visibility and tumor size, tumor location, DRR visibility, and the beam angle; results were considered significant for p values <0.05.
Tumor tracking accuracy Both manual and automatic tumor tracking in the EPID images were done retrospectively for a subset of 50 image sequences from 25 target volumes acquired during the first treatment fraction. Only images with an MV visibility rating of 2 were included in this part of the analysis because manual tumor tracking served as a reference. A reference mask specific for each tumor and each beam was defined. The shape of this mask was based on the macroscopic tumor contour in the DRR, which was delineated in 4DCT images in endinhale or end-exhale position. By browsing the images of the acquired EPID movie, an EPID image in a corresponding phase of breathing compared to the DRR was chosen as the reference frame. The reference mask was generated by retracing the tumor contour in the DRR and simultaneously drawing a mask ring in the EPID image. Thus, an annular contour of 4-mm width was produced in the EPID image. All pixel values within the mask ring were stored. Precise scaling and positioning of the DRR in the EPID reference frame were ensured to enable correct tracking. For automatic tracking, the mask was moved pixel-wise (equivalent to a resolution of 0.25 mm) within a certain region for any following image (30 pixels in each direction). The tracking algorithm minimized the mean of the sum of squared differences of pixels in the current mask position and in the reference mask to find the best matching position of the tracked object (11). For manual tracking, the reference mask contour was moved by the observer in each image according to tumor movement: coverage of the tumor by the mask according to the reference image was intended. The best matching mask position in each frame was stored and used for calculation of the tumor trajectory. As only clearly identifiable tumors were used in this analysis, the manual trajectory was defined as a reference describing the true tumor motion. All coordinates were calculated in the plane of the EPID panel, i.e., if the gantry and isocentric couch angles were 0 , the x direction in the panel corresponded to a lateral direction for the patient, the y direction correlated with the patient’s superior-inferior (SI) axis. To describe the accuracy of the automatic tracking the point-wise differences between automatic and manual trajectories were calculated. The absolute values were used and converted into millimeters (1 pixel is equivalent to 0.25 mm in the isocenter). The interobserver
Fig. 1. (a) Example of a trajectory in the SI direction with illustration of the global and local extreme positions and the moving average. (b) One breathing cycle of the trajectory with description of the parameters motion amplitude a, duration of the breathing cycle, tCYCLE.
I. J. Radiation Oncology d Biology d Physics
620
Volume 78, Number 2, 2010
Table 2. Visibility rating Category*
Rating
No. of beams
<0.5
0.5–1.5
>1.5
a
Mean rating all fractions
b
Mean rating 1st fraction
c
Maximum deviation
Number of beams Percentage Number of beams Percentage Number of beams Percentage
197 29.4 89 32.0 129 83.8
158 23.6 72 25.9 25 16.2
314 46.9 117 42.1 0 0
* a, mean rating for tumor visibility in all acquired image sequences given by the three observers; b, mean rating for sequences from the first fraction only; and c, maximum deviation between mean rating for the first fraction and the following fractions. For each group of mean rating the number of beams and the resulting percentage is given. error for manual tracking was evaluated by comparing 10 trajectories of three independent observers. The mean of the point-wise differences between the three trajectories was determined.
Detailed motion evaluation A detailed analysis of target motion by means of manual EPID tracking was performed for all EPID movies that were deemed to
Fig. 2. Analysis of tumor visibility. The different colors indicate the mean rating for the tumor visibility by three independent observers: red visibility rating, < 0.5; yellow visibility rating, 0.5 to 1.5; and green, >1.5. MV visibility and location of all 49 tumors in (a) right-left and (b) anterior-posterior planar images. (c) MV visibility dependent on GTV. (d) MV visibility dependent on DRR visibility.
Markerless tumor tracking d A. RICHTER et al.
621
Fig. 3. (a) Example of a DRR with overlaid tumor contour. (b) Corresponding portal image with template mask for the pulmonary tumor generated, using the complete DRR contour (c) Modified mask used for tracking. All pixel values inside the drawn line define the template mask. All images are scaled to the same size. Field sizes between DRR and EPID image seem to be different because of different windows and levels used for image presentation.
be of sufficient quality by all three observers (rating, 2) during the first fraction (170 EPID movies of 30 target volumes from 26 patients). A total of 64 of 170 EPID movies were excluded from evaluation because of the following reasons: (i) beams were delivered with table rotation; (ii) small segment beams not covering the complete PTV; and (iii) insufficient data were available due to short beam delivery time. This resulted in 30 targets of 26 patients and 106 EPID movies. The mean duration of trajectories was 45.8 sec with minimum and maximum durations of 11 sec and 118 sec, respectively. Multiple-fraction SBRT (median 3 fractions) had been performed for 11 of these 30 targets, allowing for analysis of interfractional uncertainties in additional 135 movies. Motion amplitude (A) in 4DCT (A4DCT) was determined by contouring the macroscopic tumor in extreme positions (end-inhalation, end-exhalation) and by calculating the differences of these positions. The EPID trajectory was separated into inspiration and expiration phases (Fig. 1a). Each breathing cycle, i, was analyzed regarding its amplitude (ai) and cycle duration (tCYCLE_i), as illustrated in Fig. 1b. For each beam trajectory, b, mean and maximum motion amplitude was calculated as aMean_b = mean(ai) and aMax_b. Breathing-induced tumor motion amplitude was small (<2 mm) in 16 of 106 trajectories of the first fraction: only the difference between minimum and maximum positions was analyzed in these cases. For each patient, mean and maximum motion amplitudes were determined: AMean_MV = mean (aMean_b) and AMax_MV = mean(aMax_b). The velocity (v) of tumor motion was determined by dividing the absolute distance between two points by the time difference. Drifts were analyzed separately for each beam by calculating the moving average (window width of two breathing cycles) as suggested in the literature (12, 13). The resulting curve was analyzed for a drift which occurred during delivery of each beam relative to its first value. The mean and maximum target motions (AMean_MV, AMax_MV, respectively) based on the EPID movies were compared with tumor motions measured in the 4DCT images (A4DCT). The correlation between target location in the lung and different parameters of target
motion was investigated for 4DCT and EPID data by the MannWhitney U test. The analysis was limited to SI direction, which is known to be the predominant direction of breathing motion (14–16).
RESULTS Tumor visibility in EPID images The ratings of the three observers for tumor visibility were averaged. Results are summarized in three groups: mean rating worse than 0.5, mean rating between 0.5 and 1.5, and mean rating better than 1.5 (Table 2a–c). Summarizing the results from the three observers for all image sequences, the tumor was completely or partially visible in all EPID images (mean rating >1.5) in 47% of all sequences, visible in some but not all EPID images (mean rating between 0.5 and 1.5) in 24% of all sequences, and not clearly identifiable (mean rating <0.5) in 29% of all analyzed beams. Full agreement among the three observers for the tumor visibility rating was seen in almost 60% of the EPID movies, disagreement of all three observers was seen in 4.3%. Ratings among the three observers differed by 2 rating levels in 5.4% (two observers rating 0 and one rating 2) and 4.3% (two observers rating 2 and one observer rating 0). Comparing the rating of the first fraction with the ratings of the following fractions, a maximum difference of less than 0.5 was seen in about 84% of all treatment beams, and a maximum difference of 1.5 or more was observed for none of the beams. This indicates good reproducibility of the MV visibility throughout the treatment fractions and justifies the restriction to the first fraction for further analysis. The final rating for a tumor is given by the average rating for all beams of the specific tumor. Figure 2a and b display the location of each tumor and the corresponding visibility ratings. No relationship between tumor location and visibility
622
I. J. Radiation Oncology d Biology d Physics
Volume 78, Number 2, 2010
Fig. 4. Summarized distribution of the point-wise differences between automatic and manual tracking in the x direction (a) differential and (b) cumulative. Summarized distribution of the point-wise differences between automatic and manual tracking in the y direction, (c) differential and (d) cumulative.
was seen (p = 0.7). No significant correlation was observed between the gantry angle and MV visibility. Tumors were classified according to their tumor size. with comparable numbers of tumors in each subgroup (Fig. 2c). A c2 test showed that MV visibility was correlated with gross tumor volume (GTV) (p = 0.005). Close correlation was found between DRR visibility and MV visibility (p < 0.05) (Fig. 2d). If DRR visibility rating and tumor size were combined to predict MV visibility, 92% of the beams would be evaluated with MV visibility >1.5 for GTV >10 cm3. Accuracy of manual and automatic EPID tracking Figure 3 shows an example of a DRR based on the planning CT with an overlaid tumor contour and the generation
of the corresponding reference mask in an EPID image. For some cases, the mask shape was manually modified to avoid overlap between the mask and the field edge (Fig. 3c). All automatic tracking was done with this modification. The modification was done when the tumor contour was closer than 5 mm to the field edge already in the DRR image; otherwise the initial mask was used. Summarized over all trajectories, the difference between manual and automatic tracking was maximum 2 mm for 98.3% of the images in the x direction and for 89.4% in the y direction; maximum differences were 5.5 mm and 11 mm for the x and y directions, respectively (Fig. 4). Tracking accuracy was 1.15 mm 0.72 mm and 0.92 mm 0.68 mm for AMV >10 mm and AMV <10 mm, respectively. The
Markerless tumor tracking d A. RICHTER et al.
623
Table 3. Patient subpopulation for detailed motion evaluation* Evaluated
Diagnosis
Patient
Target no. No. of fractions No. of beams Duration (sec) Age (a) Gender 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
3 1 3 1 3 1 1 1 1 1 1 1 1 1 3 1 3 1 1 3 1 1 1 3 1 3 1 3 3 3
7 2 4 2 2 2 3 1 4 4 3 7 4 4 5 3 5 4 3 3 3 4 3 5 4 1 1 6 4 3
30.8 38.8 98.0 16.7 51.9 50.1 41.9 49.9 34.2 68.8 40.4 32.6 40.8 22.3 35.9 14.6 38.1 86.9 48.9 78.6 40.8 78.7 50.1 41.2 38.6 43.2 30.8 38.8 98.0 16.7
59 63 66 67 68 47 79 72 69 68 33 77 63 69 75 68 87 73 72 33 47 33 72 69 72 85 88 78 68 62
F M M F M M M F M M M M M F F M M M M M M M M F M F M M F M
Tumor location
P/M
SI
RL
AP
M M P M M M M P M P M P M P M P P M M M M M M M M P P P M M
S S S S S I S S S S S I I S S S S S I I I I I I I I I I I I
R R R L R R L R R R L R R L L R L L L L R L R R L L R R L R
P A A P P P A A A A P P P P P P P A P P P P P P P P P P P P
GTV (cm3) A4DCT (mm) 19 2.5 25 1.5 41 1 23 11 2.7 25 7 90 5 8 15 30 125 40 126 8 12 3 7 20 4 21 4 27 28 31
0.7 1.6 1.6 2.1 2.5 2.5 2.5 2.6 3.0 3.2 3.4 4.3 4.5 4.6 4.8 5.0 5.6 5.7 7.5 7.6 7.6 7.7 7.9 9.0 11.9 13.6 16.1 16.7 16.7 17.9
* Demographic data and information retrieved from 4DCT investigations. A total of 30 of initially 49 lesions were selected for detailed motion evaluation. The number of evaluated beams and fractions is listed for each target. Demographic information: age in years, gender male (M)/female (F) and diagnosis primary tumor (P)/metastasis(M). Gross tumor volume (GTV) and motion amplitude (A4DCT), tumor location in superior-inferior (SI), anterior-posterior (AP), and lateral (LR) directions.
relationship between tracking accuracy and tumor amplitude was weak (Correlation coefficient [CC] 0.3). The mean interobserver error was 0.3 mm 0.3 mm and 0.4 mm 0.2 mm in the x and y directions, with maximum differences up to 1.5 mm and 1.75 mm in the x and y directions, respectively. Due to the small interobserver error, detailed motion analysis was based on manual tracking of one observer. Detailed motion analysis Motion amplitude. Motion amplitude in the planning 4DCT images was assessed for 30 target volumes (Table 3). Motion amplitudes ranged from 0.7 mm to 17.9 mm, with a median of 4.9 mm. Tumor location in the caudal and posterior parts of the lung was significantly correlated with increased motion (p < 0.001 and p = 0.008, respectively). Figure 5 shows the strong correlation (CC, 0.91) between A4DCT and AMean_MV. Motion amplitude in 4DCT images differed from AMean_MV and AMax_MV by 1.7 mm 1.4 mm and 3.0 mm 1.8 mm on average, respectively. However, larger deviations between motion in 4DCT and in EPID movies were seen, up to 5.9 mm and 8.7 mm for
AMean_MV and AMax_MV, respectively. Intrafractional variation of AMean_MV and AMax_MV within the first treatment fraction is shown in Fig. 6a. Standard deviation (SD) of AMean_MV within the first treatment fraction was 1 mm, on average, and a maximum of 3.7 mm. There was no systematic intrafractional change of the motion amplitude over the duration of the first treatment fraction: the difference in AMean_MV between the first and last radiation fields was 0.4 mm 1.4 mm, with a maximum of 3.2 mm. Interfractional variability of AMean_MV was 1.5 mm on average, with a maximum of 3.8 mm; variability of the AMax_MV was of a magnitude similar to 2.2 mm, with a maximum of 4.9 mm (Fig. 6b). There was no systematic change to increased or decreased motion amplitude between the first and last treatment fraction. The relationship between SD(AMean_MV) and absolute motion amplitude showed a weak correlation (CC, 0.66), indicating increased variation in motion amplitude for mobile targets compared to less mobile targets. The relative variation SD(AMean_MV)/AMean_MV was 15 to 20% for mobile targets (AMean_MV > 5 mm). Larger variability (up to 60%) was observed for small motion amplitudes (AMean_MV < 5 mm).
624
I. J. Radiation Oncology d Biology d Physics
Fig. 5. Correlation between motion amplitudes measured in the 4DCT (A4DCT) and average motion amplitude in the EPID movies during treatment (AMean_MV) in the SI direction.
Tumor location (right-left [RL], SI, or anterior-posterior [AP]) was correlated with intrafractional variability of AMean_MV in the SI direction: variability increased for tumor locations in caudal parts of the lung (p = 0.04).
Volume 78, Number 2, 2010
Breathing rate and tumor velocity. Irregularities of breathing rate are described by variations of the cycle duration (tCYCLE): intra- and interfractional variability of tCYCLE is shown in Fig. 7a and b, respectively. Mean breathing cycle duration for all targets was 3.3 0.7 sec. Intra- and interfractional variations of the breathing cycle duration were 0.2 sec and 0.2 sec, on average, with a corresponding maximum variation of 2.9 sec and 3.3 sec, respectively. For all targets, no systematic intrafractional or interfractional change of the cycle duration was observed. The velocity of tumor motion was analyzed for the first fraction of each target. In Fig. 8a, maximum and mean values of tumor velocity are shown for each treatment beam. Mean velocity during the first fraction was 3.4 mm/sec 3.3 mm/sec, on average, to a maximum of 61 mm/sec. Velocity values of all trajectories were summarized in a histogram plot (Fig. 8b). Velocity exceeded 10 mm/sec and 25 mm/sec in 12.4% and 2.4% of the total treatment time. As expected, larger tumor velocity was observed for targets with large motion amplitudes. Target location in the lung in AP and SI directions was associated with mean tumor velocity (p < 0.001, p = 0.008). Moving average drifts. Each trajectory was evaluated separately for drifts in the SI direction. The accumulated histogram for all trajectories shows the change in moving averages of more than 1 mm, 2 mm, and 4 mm in 47.1%, 17.1%, and 4.5% of treatment time, respectively (Fig. 9). Frequent and large drifts were found for targets 28, 29, and 30, which had the largest motion amplitudes.
Fig. 6. (a) The intrafractional variation of the motion amplitude in the SI direction is shown for 30 targets. Maximum and mean motion amplitudes are marked with open and filled circles, respectively. Each data point represents the results of one beam of the first fraction. (b) For 11 targets, interfractional variability of the mean and maximum amplitudes is plotted; each data point represents the results of one beam and treatment fractions are displayed in chronological order.
Markerless tumor tracking d A. RICHTER et al.
625
Fig. 7. (a) Intrafractional variation of the mean cycle duration for 28 targets. Each data point represents the results of one beam of the first fraction. (b) Interfractional variability for the first 3 fractions of 11 targets. Each data point represents the results of one beam. Treatment fractions are displayed in chronological order.
DISCUSSION Reliable identification of the tumor in any image of the acquired EPID movie was not possible in 29% of the analyzed beams. Continuous target monitoring from all gantry angles of the treatment plan was possible for only 8 of 49 tumors (42% for the first fraction). Tumor size and DRR visibility were shown to be the strongest predictors for visibility of the pulmonary targets in the EPID images. If the DRR visibility was 2, and the GTV was at least 10 cm3, the MV visibility was over 90% of the treatment beams. These findings will help in the selection of patients with tumors suitable for our method of markerless tumor tracking in the EPID images. The rather small percentage of completely visible targets certainly limits application in clinical practice. However,
the approach of markerless tumor tracking in EPID movies could be improved with the help of external motion signals (5), with further image processing of the EPID movies to improve the signal-to-noise-contrast, or based on machine learning algorithms (4, 17). Tang et al. recently published a method to decide if the tumor is either inside or outside the beam aperture based on neural networks (17). Furthermore, the approach of markerless EPID tracking can be seen as an additional method for quality assurance compared to current clinical practice, where intrafractional monitoring of the target itself is not routinely performed. This is in agreement with the American Association of Physicists in Medicine report no. 91 (18), which recommends monitoring and (if possible) motion management if the tumor amplitude
Fig 8. (a) The intrafractional variation of the velocity is shown for the first fraction of each target number. Maximum and mean motion amplitudes are marked with open and filled circles, respectively. (b) Probability distribution of tumor velocity for all trajectories.
626
I. J. Radiation Oncology d Biology d Physics
Fig. 9. Accumulated histogram of variations in the moving average. Intervals were defined to quantify the incidence when the drift exceeded the limits of 1 mm, 2 mm, and 4 mm.
exceeds 5 mm. The evaluation of the EPID movies showed that 49% and 25% of investigated target volumes in our study were moving more than 5 mm and 10 mm, respectively. For the subset of patients with visible targets in the EPID images, the comparison between manual and automatic target tracking was accurate and reproducible. Comparison of manual and automatic tracking resulted in mean differences of 0.6 mm 0.6 mm and 1.0 mm 1.1 mm in x and y directions, and maximum errors were 2 mm in 98.3% and 89.4% of all EPID images, respectively. An additional uncertainty should be added due to interobserver variability of manual EPID tracking, which was 0.3 mm 0.3 mm and 0.4 mm 0.2 mm in x and y directions, respectively. Arimura et al. (4) recently presented a method for markerless tumor tracking in EPID images with similar results for tracking accuracy. They focused on the development of a computerized method to find the tumor position and not on quantitative evaluation of trajectories due to the low frame rate of 0.5 frames per sec (fps) (4). The issue of visibility was not addressed in their work. We investigated whether 4DCT imaging is reliable for estimating the motion pattern and range of pulmonary tumors. We observed a strong correlation between target motion amplitude at treatment planning (4DCT studies) and during actual treatment (EPID movies); the difference between AMean_MV and A4DCT was 1.7 mm 1.4 mm on average, up to a 6-mm maximum. The results show that 4DCT describes the mean motion amplitude in EPID movies for the majority of patients, which is relevant for gated, tracking, and free-breathing techniques. However, larger deviations between motion analysis in 4DCT and EPID movies were observed (up to 5.9 mm and 8.7 mm for AMean_MV and AMax_MV, respectively) which again necessitates a monitoring of tumor motion during delivery. Velocity of target motion was investigated with regard to the inherent limits of breathing-synchronized treatment tech-
Volume 78, Number 2, 2010
niques (6, 19–21). An average maximum leaf velocity in the isocenter plane of 33 mm/sec is reported in the literature (9, 22). Regarding the total treatment time, this value was exceeded for less than 1% of the treatment time. Another approach is compensation of breathing motion by the treatment couch. In 16.9% of treatment times, the tumor velocity exceeded the speed limits of the HexaPOD table (8 mm/sec), which could be used for breathing-synchronized counter steering of the treatment couch (6). The trajectories were analyzed for drifts which occurred during beam delivery. Variations of the moving average of more than 2 mm and 4 mm occurred in 17% and 5% of the monitored treatment times, respectively. For more mobile targets, located mainly in the lower lobe, we observed a larger intrafractional variation in mean tumor position, which is an important issue for SBRT and its prolonged delivery times. Shirato et al. (9) showed that readjustments were needed as a result of a baseline drift of the tumor position to improve efficiency during gated irradiation. Limitations The EPID movies provided information regarding the motion in the plane perpendicular to the beam direction only. Image acquisition from multiple gantry angles or imaging with two stereoscopic devices would be necessary to obtain full information of the 3D motion signal. However, this inline imaging plane is considered sufficient for target monitoring in photon radiotherapy (18, 23). The system presented herein offers a frame rate of 2 fps, which is in conformity with the sampling theorem (Nyquist-Shannon). The rate of 2 fps may affect the accuracy of detecting the end-inhalation point. The end-exhalation is less sensitive, while the exhalation phase is usually longer than the inhalation phase. The time needed by the tracking algorithm to find the best match of the mask in the actual MV picture ranged from 0.09 sec to 0.60 sec (mean, 0.29 sec 0.11 sec), depending on the size of the mask (approximately 0.1 sec per 1.000 pixels on a 2.2-GHz Core2 Duo central processing unit). This time could be decreased by a factor of 4 if tracking resolution is reduced to 0.5 mm (2 pixels). As the tracking process and the resulting trajectories are independent for each beam, the evaluation of the EPID movies presented here is currently limited to detecting drifts in moving average for each beam separately. Interbeam drifts were not considered and are expected to be larger than intrabeam drifts. We are currently working on this issue: absolute mask positions in each EPID movie will allow a concatenation of intrabeam baselines for the whole treatment fraction. Motion trajectories were based on manual tracking, which is error prone due to inter- and intraobserver variability, which was determined as small. CONCLUSIONS Continuous monitoring of the tumor position was possible for 47% of the analyzed treatment beams (42% for the first
Markerless tumor tracking d A. RICHTER et al.
fraction). Continuous target monitoring from all gantry angles was possible for only 8 of 49 tumors. Tumor size and target visibility in the planning DRR images can be used to predict target visibility in the EPID images. The comparison between manual and automatic target tracking was accurate
627
and reproducible in the subset of patients with visible targets. The 4DCT describes the mean motion amplitude in EPID movies for the majority of the patients well. However, larger deviations in target motion were observed which necessitate a monitoring during delivery.
REFERENCES 1. Rietzel E, Chen GT, Choi NC, et al. Four-dimensional imagebased treatment planning: Target volume segmentation and dose calculation in the presence of respiratory motion. Int J Radiat Oncol Biol Phys 2005;61:1535–1550. 2. Mageras GS, Pevsner A, Yorke ED, et al. Measurement of lung tumor motion using respiration-correlated CT. Int J Radiat Oncol Biol Phys 2004;60:933–941. 3. Vedam SS, Keall PJ, Kini VR, et al. Determining parameters for respiration-gated radiotherapy. Med Phys 2001;28:2139–2146. 4. Arimura H, Egashira Y, Shioyama Y, et al. Computerized method for estimation of the location of a lung tumor on EPID cine images without implanted markers in stereotactic body radiotherapy. Phys Med Biol 2009;54:665–677. 5. Lin T, Cervino LI, Tang X, et al. Fluoroscopic tumor tracking for image-guided lung cancer radiotherapy. Phys Med Biol 2009;54:981–992. 6. Wilbert J, Meyer J, Baier K, et al. Tumor tracking and motion compensation with an adaptive tumor tracking system (ATTS): System description and prototype testing. Med Phys 2008;35:3911–3921. 7. Berbeco RI, Hacker F, Ionascu D, et al. Clinical feasibility of using an EPID in CINE mode for image-guided verification of stereotactic body radiotherapy. Int J Radiat Oncol Biol Phys 2007;69:258–266. 8. Korreman SS, Juhler-Nottrup T, Boyer AL. Respiratory gated beam delivery cannot facilitate margin reduction, unless combined with respiratory correlated image guidance. Radiother Oncol 2008;86:61–68. 9. Shirato H, Suzuki K, Sharp GC, et al. Speed and amplitude of lung tumor motion precisely detected in four-dimensional setup and in real-time tumor-tracking radiotherapy. Int J Radiat Oncol Biol Phys 2006;64:1229–1236. 10. Guckenberger M, Wulf J, Mueller G, et al. Dose-response relationship for image-guided stereotactic body radiotherapy of pulmonary tumors: Relevance of 4d dose calculation. Int J Radiat Oncol Biol Phys 2009;74:47–54. 11. Meyer J, Richter A, Baier K, et al. Tracking moving objects with megavoltage portal imaging: A feasibility study. Med Phys 2006;33:1275–1280.
12. Lu W, Nystrom MM, Parikh PJ, et al. A semi-automatic method for peak and valley detection in free-breathing respiratory waveforms. Med Phys 2006;33:3634–3636. 13. Suh Y, Dieterich S, Cho B, et al. An analysis of thoracic and abdominal tumour motion for stereotactic body radiotherapy patients. Phys Med Biol 2008;53:3623–3640. 14. Liu HH, Balter P, Tutt T, et al. Assessing respirationinduced tumor motion and internal target volume using four-dimensional computed tomography for radiotherapy of lung cancer. Int J Radiat Oncol Biol Phys 2007;68: 531–540. 15. Seppenwoolde Y, Shirato H, Kitamura K, et al. Precise and realtime measurement of 3D tumor motion in lung due to breathing and heartbeat, measured during radiotherapy. Int J Radiat Oncol Biol Phys 2002;53:822–834. 16. Sonke JJ, Lebesque J, van Herk M. Variability of four-dimensional computed tomography patient models. Int J Radiat Oncol Biol Phys 2008;70:590–598. 17. Tang X, Lin T, Jiang S. A feasibility study of treatment verification using EPID cine images for hypofractionated lung radiotherapy. Phys Med Biol 2009;54:S1–S8. 18. Keall PJ, Mageras GS, Balter JM, et al. The management of respiratory motion in radiation oncology report of AAPM Task Group 76. Med Phys 2006;33:3874–3900. 19. D’Souza WD, Naqvi SA, Yu CX. Real-time intra-fractionmotion tracking using the treatment couch: A feasibility study. Phys Med Biol 2005;50:4021–4033. 20. Keall PJ, Kini VR, Vedam SS, et al. Motion adaptive x-ray therapy: A feasibility study. Phys Med Biol 2001;46:1–10. 21. McQuaid D, Webb S. Target-tracking deliveries using conventional multileaf collimators planned with 4D direct-aperture optimization. Phys Med Biol 2008;53:4013–4029. 22. Wijesooriya K, Bartee C, Siebers JV, et al. Determination of maximum leaf velocity and acceleration of a dynamic multileaf collimator: implications for 4D radiotherapy. Med Phys 2005; 32:932–941. 23. Suh Y, Dieterich S, Keall PJ. Geometric uncertainty of 2D projection imaging in monitoring 3D tumor motion. Phys Med Biol 2007;52:3439–3454.
doi:10.1016/j.ijrobp.2009.11.028
PHYSICS CONTRIBUTION
FEASIBILITY STUDY FOR MARKERLESS TRACKING OF LUNG TUMORS IN STEREOTACTIC BODY RADIOTHERAPY ANNE RICHTER, M.SC., JUERGEN WILBERT, PH.D., KURT BAIER, M.SC., MICHAEL FLENTJE, M.D., AND MATTHIAS GUCKENBERGER, M.D. University of Wuerzburg, Department of Radiation Oncology, Wuerzburg, Germany Purpose: To evaluate the feasibility and accuracy of a method for markerless tracking of lung tumors in electronic portal imaging device (EPID) movies and to analyze intra- and interfractional variations in tumor motion. Methods and Materials: EPID movies were acquired during stereotactic body radiotherapy (SBRT) given to 40 patients with 49 pulmonary targets and retrospectively analyzed. Tumor visibility and tracking accuracy were determined by three observers. Tumor motion of 30 targets was analyzed in detail via four-dimensional computed tomography (4DCT) and EPID in the superior-inferior direction for intra- and interfractional variations. Results: Tumor visibility was sufficient for markerless tracking in 47% of the EPID movies. Tumor size and visibility in the DRR were correlated with visibility in the EPID images. The difference between automatic and manual tracking was a maximum of 2 mm for 98.3% in the x direction and 89.4% in the y direction. Motion amplitudes in 4DCT images (range, 0.7-17.9 mm; median, 4.9 mm) were closely correlated with amplitudes in the EPID movies. Intrafractional and interfractional variability of tumor motion amplitude were of similar magnitude: 1 mm on average to a maximum of 4 mm. A change in moving average of more than ±1 mm, ±2 mm, and ±4 mm were observed in 47.1%, 17.1%, and 4.5% of treatment time for all trajectories, respectively. Mean tumor velocity was 3.4 mm/sec, to a maximum 61 mm/sec. Conclusions: Tracking of pulmonary tumors in EPID images without implanted markers was feasible in 47% of all treatment beams. 4DCT is representative of the evaluation of mean breathing motion on average, but larger deviations occurred in target motion between treatment planning and delivery effort a monitoring during delivery. Ó 2010 Elsevier Inc. Tumor motion, Markerless tracking, Interfractional, Intrafractional, EPID.
METHODS AND MATERIALS
INTRODUCTION Temporal changes in the patients’ anatomy due to breathing motion may reduce the accuracy of radiotherapy treatment, especially for hypofractionated stereotactic body radiotherapy (SBRT). Precise assessment and quantification of breathinginduced target motion and its integration into the treatment workflow are essential for adaptive treatment techniques (1). Existing strategies for tumor tracking approaches can be grouped roughly into three categories: external surrogate based (2, 3), markerless (4–6), and internal marker-based systems (7–9). The aim of this study was to investigate the feasibility of imaging and tracking lung tumors, using the Mega Voltage (MV) therapy beam and the electronic portal imaging device (EPID) without implanted markers.
Data acquisition and patient population
Reprint requests to: Anne Richter, M.Sc., Department of Radiation Oncology, University of Wuerzburg, Josef-Schneider-Str. 11, 97080 Wuerzburg, Germany. Tel: 49 (0)931 201-28881; Fax: 49 (0)931 201-28221; E-mail: [email protected] Anne Richter and Juergen Wilbert contributed equally to this work. This work was partially supported by the Bayerische Forschungsstiftung, Germany.
This work was presented in part at the Medical Physics and Biological Engineering World Congress September 7–12, 2009, Munich, Germany. Conflict of interest: none. Acknowledgment—We thank Mark Gainey for proofreading the manuscript. Received June 8, 2009, and in revised form Nov 3, 2009. Accepted for publication Nov 16, 2009.
Data are based on 40 consecutive lung cancer patients with 49 pulmonary targets treated with image-guided SBRT in our department (patient and treatment characteristics are shown in Table 1). For each patient, a four-dimensional computed tomography (4DCT) scan was acquired for treatment planning (Somatom Sensation Open, Siemens, Forchheim, Germany) without breath coaching. Abdominal compression was used for 20 of the 40 patients. All patients were treated on an Elekta Synergy S (Elekta, Crawley, UK) equipped with a Kilo Voltage cone beam system and an MV EPID. Patients were positioned in the stereotactic body frame or a BodyFix system (Elekta, Crawley, UK). Treatment position was verified via cone beam CT imaging prior to all treatment fractions. Details of the planning and treatment technique have been described
618
Markerless tumor tracking d A. RICHTER et al.
Table 1. Patient population* Characteristics
Minimum
Maximum
Median
Patient age (a) A 4DCT (mm) GTV volume (cm3) Number of fractions
30 0.7 0.3 1
86 18.8 126 8
66.5 4.4 9.5 3
* Characteristics for the initial patient population with 49 lesions. previously (10). EPID movies were acquired during delivery of the beams, resulting in 668 EPID movies within 278 fractions.
Evaluation of tumor visibility Digitally reconstructed radiographs (DRR) were generated for each beam based on the planning CT study using Pinnacle3 planning software (Philips Radiation Oncology Systems, Fitchburg, WI). Parameters (energy and brightness) were adapted to obtain image quality similar to that of EPIDs. The macroscopic tumor was delineated in 4DCT images by an MD. The contour of the macroscopic tumor was overlaid and stored with these DRR images. The visibility of the tumor in these DRRs (DRR visibility) was evaluated retrospectively using a scheme with three rating levels (0, not visible; 1, partially visible; 2, clearly visible). DRRs and EPID frames were displayed side by side to facilitate the identification of the pulmonary tumor in the EPID images. The visibility of the pulmonary target in EPID images (MV visibility) was investigated in all 668 EPID movies, separately. A scheme with three rating levels was introduced for the evaluation of tumor visibility: a rating of 2 represented complete or partial tumor shape that was identifiable in all EPID images of the beam; a rating of 1 represented a complete or partial tumor shape identifiable in some but not all images; and a rating of 0 represented a tumor that could not be clearly identified. All EPID images were evaluated retrospectively by three independent observers (physicists and radiographers). The final result for each image sequence was defined as the mean rating by the three observers. As the number of fractions was different for the tumors, the mean rating for the first fraction was used for evaluation of the overall tumor visibility. The maximum deviation between the mean rating of the first fraction and of each following fraction was calculated to evaluate the reproducibility of MV visibility. Statistica 6.0 software (Statsoft, Tulsa, OK) was utilized for statistical analysis. A Pearson c2 test was used for analyzing the rela-
619
tionship between MV visibility and tumor size, tumor location, DRR visibility, and the beam angle; results were considered significant for p values <0.05.
Tumor tracking accuracy Both manual and automatic tumor tracking in the EPID images were done retrospectively for a subset of 50 image sequences from 25 target volumes acquired during the first treatment fraction. Only images with an MV visibility rating of 2 were included in this part of the analysis because manual tumor tracking served as a reference. A reference mask specific for each tumor and each beam was defined. The shape of this mask was based on the macroscopic tumor contour in the DRR, which was delineated in 4DCT images in endinhale or end-exhale position. By browsing the images of the acquired EPID movie, an EPID image in a corresponding phase of breathing compared to the DRR was chosen as the reference frame. The reference mask was generated by retracing the tumor contour in the DRR and simultaneously drawing a mask ring in the EPID image. Thus, an annular contour of 4-mm width was produced in the EPID image. All pixel values within the mask ring were stored. Precise scaling and positioning of the DRR in the EPID reference frame were ensured to enable correct tracking. For automatic tracking, the mask was moved pixel-wise (equivalent to a resolution of 0.25 mm) within a certain region for any following image (30 pixels in each direction). The tracking algorithm minimized the mean of the sum of squared differences of pixels in the current mask position and in the reference mask to find the best matching position of the tracked object (11). For manual tracking, the reference mask contour was moved by the observer in each image according to tumor movement: coverage of the tumor by the mask according to the reference image was intended. The best matching mask position in each frame was stored and used for calculation of the tumor trajectory. As only clearly identifiable tumors were used in this analysis, the manual trajectory was defined as a reference describing the true tumor motion. All coordinates were calculated in the plane of the EPID panel, i.e., if the gantry and isocentric couch angles were 0 , the x direction in the panel corresponded to a lateral direction for the patient, the y direction correlated with the patient’s superior-inferior (SI) axis. To describe the accuracy of the automatic tracking the point-wise differences between automatic and manual trajectories were calculated. The absolute values were used and converted into millimeters (1 pixel is equivalent to 0.25 mm in the isocenter). The interobserver
Fig. 1. (a) Example of a trajectory in the SI direction with illustration of the global and local extreme positions and the moving average. (b) One breathing cycle of the trajectory with description of the parameters motion amplitude a, duration of the breathing cycle, tCYCLE.
I. J. Radiation Oncology d Biology d Physics
620
Volume 78, Number 2, 2010
Table 2. Visibility rating Category*
Rating
No. of beams
<0.5
0.5–1.5
>1.5
a
Mean rating all fractions
b
Mean rating 1st fraction
c
Maximum deviation
Number of beams Percentage Number of beams Percentage Number of beams Percentage
197 29.4 89 32.0 129 83.8
158 23.6 72 25.9 25 16.2
314 46.9 117 42.1 0 0
* a, mean rating for tumor visibility in all acquired image sequences given by the three observers; b, mean rating for sequences from the first fraction only; and c, maximum deviation between mean rating for the first fraction and the following fractions. For each group of mean rating the number of beams and the resulting percentage is given. error for manual tracking was evaluated by comparing 10 trajectories of three independent observers. The mean of the point-wise differences between the three trajectories was determined.
Detailed motion evaluation A detailed analysis of target motion by means of manual EPID tracking was performed for all EPID movies that were deemed to
Fig. 2. Analysis of tumor visibility. The different colors indicate the mean rating for the tumor visibility by three independent observers: red visibility rating, < 0.5; yellow visibility rating, 0.5 to 1.5; and green, >1.5. MV visibility and location of all 49 tumors in (a) right-left and (b) anterior-posterior planar images. (c) MV visibility dependent on GTV. (d) MV visibility dependent on DRR visibility.
Markerless tumor tracking d A. RICHTER et al.
621
Fig. 3. (a) Example of a DRR with overlaid tumor contour. (b) Corresponding portal image with template mask for the pulmonary tumor generated, using the complete DRR contour (c) Modified mask used for tracking. All pixel values inside the drawn line define the template mask. All images are scaled to the same size. Field sizes between DRR and EPID image seem to be different because of different windows and levels used for image presentation.
be of sufficient quality by all three observers (rating, 2) during the first fraction (170 EPID movies of 30 target volumes from 26 patients). A total of 64 of 170 EPID movies were excluded from evaluation because of the following reasons: (i) beams were delivered with table rotation; (ii) small segment beams not covering the complete PTV; and (iii) insufficient data were available due to short beam delivery time. This resulted in 30 targets of 26 patients and 106 EPID movies. The mean duration of trajectories was 45.8 sec with minimum and maximum durations of 11 sec and 118 sec, respectively. Multiple-fraction SBRT (median 3 fractions) had been performed for 11 of these 30 targets, allowing for analysis of interfractional uncertainties in additional 135 movies. Motion amplitude (A) in 4DCT (A4DCT) was determined by contouring the macroscopic tumor in extreme positions (end-inhalation, end-exhalation) and by calculating the differences of these positions. The EPID trajectory was separated into inspiration and expiration phases (Fig. 1a). Each breathing cycle, i, was analyzed regarding its amplitude (ai) and cycle duration (tCYCLE_i), as illustrated in Fig. 1b. For each beam trajectory, b, mean and maximum motion amplitude was calculated as aMean_b = mean(ai) and aMax_b. Breathing-induced tumor motion amplitude was small (<2 mm) in 16 of 106 trajectories of the first fraction: only the difference between minimum and maximum positions was analyzed in these cases. For each patient, mean and maximum motion amplitudes were determined: AMean_MV = mean (aMean_b) and AMax_MV = mean(aMax_b). The velocity (v) of tumor motion was determined by dividing the absolute distance between two points by the time difference. Drifts were analyzed separately for each beam by calculating the moving average (window width of two breathing cycles) as suggested in the literature (12, 13). The resulting curve was analyzed for a drift which occurred during delivery of each beam relative to its first value. The mean and maximum target motions (AMean_MV, AMax_MV, respectively) based on the EPID movies were compared with tumor motions measured in the 4DCT images (A4DCT). The correlation between target location in the lung and different parameters of target
motion was investigated for 4DCT and EPID data by the MannWhitney U test. The analysis was limited to SI direction, which is known to be the predominant direction of breathing motion (14–16).
RESULTS Tumor visibility in EPID images The ratings of the three observers for tumor visibility were averaged. Results are summarized in three groups: mean rating worse than 0.5, mean rating between 0.5 and 1.5, and mean rating better than 1.5 (Table 2a–c). Summarizing the results from the three observers for all image sequences, the tumor was completely or partially visible in all EPID images (mean rating >1.5) in 47% of all sequences, visible in some but not all EPID images (mean rating between 0.5 and 1.5) in 24% of all sequences, and not clearly identifiable (mean rating <0.5) in 29% of all analyzed beams. Full agreement among the three observers for the tumor visibility rating was seen in almost 60% of the EPID movies, disagreement of all three observers was seen in 4.3%. Ratings among the three observers differed by 2 rating levels in 5.4% (two observers rating 0 and one rating 2) and 4.3% (two observers rating 2 and one observer rating 0). Comparing the rating of the first fraction with the ratings of the following fractions, a maximum difference of less than 0.5 was seen in about 84% of all treatment beams, and a maximum difference of 1.5 or more was observed for none of the beams. This indicates good reproducibility of the MV visibility throughout the treatment fractions and justifies the restriction to the first fraction for further analysis. The final rating for a tumor is given by the average rating for all beams of the specific tumor. Figure 2a and b display the location of each tumor and the corresponding visibility ratings. No relationship between tumor location and visibility
622
I. J. Radiation Oncology d Biology d Physics
Volume 78, Number 2, 2010
Fig. 4. Summarized distribution of the point-wise differences between automatic and manual tracking in the x direction (a) differential and (b) cumulative. Summarized distribution of the point-wise differences between automatic and manual tracking in the y direction, (c) differential and (d) cumulative.
was seen (p = 0.7). No significant correlation was observed between the gantry angle and MV visibility. Tumors were classified according to their tumor size. with comparable numbers of tumors in each subgroup (Fig. 2c). A c2 test showed that MV visibility was correlated with gross tumor volume (GTV) (p = 0.005). Close correlation was found between DRR visibility and MV visibility (p < 0.05) (Fig. 2d). If DRR visibility rating and tumor size were combined to predict MV visibility, 92% of the beams would be evaluated with MV visibility >1.5 for GTV >10 cm3. Accuracy of manual and automatic EPID tracking Figure 3 shows an example of a DRR based on the planning CT with an overlaid tumor contour and the generation
of the corresponding reference mask in an EPID image. For some cases, the mask shape was manually modified to avoid overlap between the mask and the field edge (Fig. 3c). All automatic tracking was done with this modification. The modification was done when the tumor contour was closer than 5 mm to the field edge already in the DRR image; otherwise the initial mask was used. Summarized over all trajectories, the difference between manual and automatic tracking was maximum 2 mm for 98.3% of the images in the x direction and for 89.4% in the y direction; maximum differences were 5.5 mm and 11 mm for the x and y directions, respectively (Fig. 4). Tracking accuracy was 1.15 mm 0.72 mm and 0.92 mm 0.68 mm for AMV >10 mm and AMV <10 mm, respectively. The
Markerless tumor tracking d A. RICHTER et al.
623
Table 3. Patient subpopulation for detailed motion evaluation* Evaluated
Diagnosis
Patient
Target no. No. of fractions No. of beams Duration (sec) Age (a) Gender 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
3 1 3 1 3 1 1 1 1 1 1 1 1 1 3 1 3 1 1 3 1 1 1 3 1 3 1 3 3 3
7 2 4 2 2 2 3 1 4 4 3 7 4 4 5 3 5 4 3 3 3 4 3 5 4 1 1 6 4 3
30.8 38.8 98.0 16.7 51.9 50.1 41.9 49.9 34.2 68.8 40.4 32.6 40.8 22.3 35.9 14.6 38.1 86.9 48.9 78.6 40.8 78.7 50.1 41.2 38.6 43.2 30.8 38.8 98.0 16.7
59 63 66 67 68 47 79 72 69 68 33 77 63 69 75 68 87 73 72 33 47 33 72 69 72 85 88 78 68 62
F M M F M M M F M M M M M F F M M M M M M M M F M F M M F M
Tumor location
P/M
SI
RL
AP
M M P M M M M P M P M P M P M P P M M M M M M M M P P P M M
S S S S S I S S S S S I I S S S S S I I I I I I I I I I I I
R R R L R R L R R R L R R L L R L L L L R L R R L L R R L R
P A A P P P A A A A P P P P P P P A P P P P P P P P P P P P
GTV (cm3) A4DCT (mm) 19 2.5 25 1.5 41 1 23 11 2.7 25 7 90 5 8 15 30 125 40 126 8 12 3 7 20 4 21 4 27 28 31
0.7 1.6 1.6 2.1 2.5 2.5 2.5 2.6 3.0 3.2 3.4 4.3 4.5 4.6 4.8 5.0 5.6 5.7 7.5 7.6 7.6 7.7 7.9 9.0 11.9 13.6 16.1 16.7 16.7 17.9
* Demographic data and information retrieved from 4DCT investigations. A total of 30 of initially 49 lesions were selected for detailed motion evaluation. The number of evaluated beams and fractions is listed for each target. Demographic information: age in years, gender male (M)/female (F) and diagnosis primary tumor (P)/metastasis(M). Gross tumor volume (GTV) and motion amplitude (A4DCT), tumor location in superior-inferior (SI), anterior-posterior (AP), and lateral (LR) directions.
relationship between tracking accuracy and tumor amplitude was weak (Correlation coefficient [CC] 0.3). The mean interobserver error was 0.3 mm 0.3 mm and 0.4 mm 0.2 mm in the x and y directions, with maximum differences up to 1.5 mm and 1.75 mm in the x and y directions, respectively. Due to the small interobserver error, detailed motion analysis was based on manual tracking of one observer. Detailed motion analysis Motion amplitude. Motion amplitude in the planning 4DCT images was assessed for 30 target volumes (Table 3). Motion amplitudes ranged from 0.7 mm to 17.9 mm, with a median of 4.9 mm. Tumor location in the caudal and posterior parts of the lung was significantly correlated with increased motion (p < 0.001 and p = 0.008, respectively). Figure 5 shows the strong correlation (CC, 0.91) between A4DCT and AMean_MV. Motion amplitude in 4DCT images differed from AMean_MV and AMax_MV by 1.7 mm 1.4 mm and 3.0 mm 1.8 mm on average, respectively. However, larger deviations between motion in 4DCT and in EPID movies were seen, up to 5.9 mm and 8.7 mm for
AMean_MV and AMax_MV, respectively. Intrafractional variation of AMean_MV and AMax_MV within the first treatment fraction is shown in Fig. 6a. Standard deviation (SD) of AMean_MV within the first treatment fraction was 1 mm, on average, and a maximum of 3.7 mm. There was no systematic intrafractional change of the motion amplitude over the duration of the first treatment fraction: the difference in AMean_MV between the first and last radiation fields was 0.4 mm 1.4 mm, with a maximum of 3.2 mm. Interfractional variability of AMean_MV was 1.5 mm on average, with a maximum of 3.8 mm; variability of the AMax_MV was of a magnitude similar to 2.2 mm, with a maximum of 4.9 mm (Fig. 6b). There was no systematic change to increased or decreased motion amplitude between the first and last treatment fraction. The relationship between SD(AMean_MV) and absolute motion amplitude showed a weak correlation (CC, 0.66), indicating increased variation in motion amplitude for mobile targets compared to less mobile targets. The relative variation SD(AMean_MV)/AMean_MV was 15 to 20% for mobile targets (AMean_MV > 5 mm). Larger variability (up to 60%) was observed for small motion amplitudes (AMean_MV < 5 mm).
624
I. J. Radiation Oncology d Biology d Physics
Fig. 5. Correlation between motion amplitudes measured in the 4DCT (A4DCT) and average motion amplitude in the EPID movies during treatment (AMean_MV) in the SI direction.
Tumor location (right-left [RL], SI, or anterior-posterior [AP]) was correlated with intrafractional variability of AMean_MV in the SI direction: variability increased for tumor locations in caudal parts of the lung (p = 0.04).
Volume 78, Number 2, 2010
Breathing rate and tumor velocity. Irregularities of breathing rate are described by variations of the cycle duration (tCYCLE): intra- and interfractional variability of tCYCLE is shown in Fig. 7a and b, respectively. Mean breathing cycle duration for all targets was 3.3 0.7 sec. Intra- and interfractional variations of the breathing cycle duration were 0.2 sec and 0.2 sec, on average, with a corresponding maximum variation of 2.9 sec and 3.3 sec, respectively. For all targets, no systematic intrafractional or interfractional change of the cycle duration was observed. The velocity of tumor motion was analyzed for the first fraction of each target. In Fig. 8a, maximum and mean values of tumor velocity are shown for each treatment beam. Mean velocity during the first fraction was 3.4 mm/sec 3.3 mm/sec, on average, to a maximum of 61 mm/sec. Velocity values of all trajectories were summarized in a histogram plot (Fig. 8b). Velocity exceeded 10 mm/sec and 25 mm/sec in 12.4% and 2.4% of the total treatment time. As expected, larger tumor velocity was observed for targets with large motion amplitudes. Target location in the lung in AP and SI directions was associated with mean tumor velocity (p < 0.001, p = 0.008). Moving average drifts. Each trajectory was evaluated separately for drifts in the SI direction. The accumulated histogram for all trajectories shows the change in moving averages of more than 1 mm, 2 mm, and 4 mm in 47.1%, 17.1%, and 4.5% of treatment time, respectively (Fig. 9). Frequent and large drifts were found for targets 28, 29, and 30, which had the largest motion amplitudes.
Fig. 6. (a) The intrafractional variation of the motion amplitude in the SI direction is shown for 30 targets. Maximum and mean motion amplitudes are marked with open and filled circles, respectively. Each data point represents the results of one beam of the first fraction. (b) For 11 targets, interfractional variability of the mean and maximum amplitudes is plotted; each data point represents the results of one beam and treatment fractions are displayed in chronological order.
Markerless tumor tracking d A. RICHTER et al.
625
Fig. 7. (a) Intrafractional variation of the mean cycle duration for 28 targets. Each data point represents the results of one beam of the first fraction. (b) Interfractional variability for the first 3 fractions of 11 targets. Each data point represents the results of one beam. Treatment fractions are displayed in chronological order.
DISCUSSION Reliable identification of the tumor in any image of the acquired EPID movie was not possible in 29% of the analyzed beams. Continuous target monitoring from all gantry angles of the treatment plan was possible for only 8 of 49 tumors (42% for the first fraction). Tumor size and DRR visibility were shown to be the strongest predictors for visibility of the pulmonary targets in the EPID images. If the DRR visibility was 2, and the GTV was at least 10 cm3, the MV visibility was over 90% of the treatment beams. These findings will help in the selection of patients with tumors suitable for our method of markerless tumor tracking in the EPID images. The rather small percentage of completely visible targets certainly limits application in clinical practice. However,
the approach of markerless tumor tracking in EPID movies could be improved with the help of external motion signals (5), with further image processing of the EPID movies to improve the signal-to-noise-contrast, or based on machine learning algorithms (4, 17). Tang et al. recently published a method to decide if the tumor is either inside or outside the beam aperture based on neural networks (17). Furthermore, the approach of markerless EPID tracking can be seen as an additional method for quality assurance compared to current clinical practice, where intrafractional monitoring of the target itself is not routinely performed. This is in agreement with the American Association of Physicists in Medicine report no. 91 (18), which recommends monitoring and (if possible) motion management if the tumor amplitude
Fig 8. (a) The intrafractional variation of the velocity is shown for the first fraction of each target number. Maximum and mean motion amplitudes are marked with open and filled circles, respectively. (b) Probability distribution of tumor velocity for all trajectories.
626
I. J. Radiation Oncology d Biology d Physics
Fig. 9. Accumulated histogram of variations in the moving average. Intervals were defined to quantify the incidence when the drift exceeded the limits of 1 mm, 2 mm, and 4 mm.
exceeds 5 mm. The evaluation of the EPID movies showed that 49% and 25% of investigated target volumes in our study were moving more than 5 mm and 10 mm, respectively. For the subset of patients with visible targets in the EPID images, the comparison between manual and automatic target tracking was accurate and reproducible. Comparison of manual and automatic tracking resulted in mean differences of 0.6 mm 0.6 mm and 1.0 mm 1.1 mm in x and y directions, and maximum errors were 2 mm in 98.3% and 89.4% of all EPID images, respectively. An additional uncertainty should be added due to interobserver variability of manual EPID tracking, which was 0.3 mm 0.3 mm and 0.4 mm 0.2 mm in x and y directions, respectively. Arimura et al. (4) recently presented a method for markerless tumor tracking in EPID images with similar results for tracking accuracy. They focused on the development of a computerized method to find the tumor position and not on quantitative evaluation of trajectories due to the low frame rate of 0.5 frames per sec (fps) (4). The issue of visibility was not addressed in their work. We investigated whether 4DCT imaging is reliable for estimating the motion pattern and range of pulmonary tumors. We observed a strong correlation between target motion amplitude at treatment planning (4DCT studies) and during actual treatment (EPID movies); the difference between AMean_MV and A4DCT was 1.7 mm 1.4 mm on average, up to a 6-mm maximum. The results show that 4DCT describes the mean motion amplitude in EPID movies for the majority of patients, which is relevant for gated, tracking, and free-breathing techniques. However, larger deviations between motion analysis in 4DCT and EPID movies were observed (up to 5.9 mm and 8.7 mm for AMean_MV and AMax_MV, respectively) which again necessitates a monitoring of tumor motion during delivery. Velocity of target motion was investigated with regard to the inherent limits of breathing-synchronized treatment tech-
Volume 78, Number 2, 2010
niques (6, 19–21). An average maximum leaf velocity in the isocenter plane of 33 mm/sec is reported in the literature (9, 22). Regarding the total treatment time, this value was exceeded for less than 1% of the treatment time. Another approach is compensation of breathing motion by the treatment couch. In 16.9% of treatment times, the tumor velocity exceeded the speed limits of the HexaPOD table (8 mm/sec), which could be used for breathing-synchronized counter steering of the treatment couch (6). The trajectories were analyzed for drifts which occurred during beam delivery. Variations of the moving average of more than 2 mm and 4 mm occurred in 17% and 5% of the monitored treatment times, respectively. For more mobile targets, located mainly in the lower lobe, we observed a larger intrafractional variation in mean tumor position, which is an important issue for SBRT and its prolonged delivery times. Shirato et al. (9) showed that readjustments were needed as a result of a baseline drift of the tumor position to improve efficiency during gated irradiation. Limitations The EPID movies provided information regarding the motion in the plane perpendicular to the beam direction only. Image acquisition from multiple gantry angles or imaging with two stereoscopic devices would be necessary to obtain full information of the 3D motion signal. However, this inline imaging plane is considered sufficient for target monitoring in photon radiotherapy (18, 23). The system presented herein offers a frame rate of 2 fps, which is in conformity with the sampling theorem (Nyquist-Shannon). The rate of 2 fps may affect the accuracy of detecting the end-inhalation point. The end-exhalation is less sensitive, while the exhalation phase is usually longer than the inhalation phase. The time needed by the tracking algorithm to find the best match of the mask in the actual MV picture ranged from 0.09 sec to 0.60 sec (mean, 0.29 sec 0.11 sec), depending on the size of the mask (approximately 0.1 sec per 1.000 pixels on a 2.2-GHz Core2 Duo central processing unit). This time could be decreased by a factor of 4 if tracking resolution is reduced to 0.5 mm (2 pixels). As the tracking process and the resulting trajectories are independent for each beam, the evaluation of the EPID movies presented here is currently limited to detecting drifts in moving average for each beam separately. Interbeam drifts were not considered and are expected to be larger than intrabeam drifts. We are currently working on this issue: absolute mask positions in each EPID movie will allow a concatenation of intrabeam baselines for the whole treatment fraction. Motion trajectories were based on manual tracking, which is error prone due to inter- and intraobserver variability, which was determined as small. CONCLUSIONS Continuous monitoring of the tumor position was possible for 47% of the analyzed treatment beams (42% for the first
Markerless tumor tracking d A. RICHTER et al.
fraction). Continuous target monitoring from all gantry angles was possible for only 8 of 49 tumors. Tumor size and target visibility in the planning DRR images can be used to predict target visibility in the EPID images. The comparison between manual and automatic target tracking was accurate
627
and reproducible in the subset of patients with visible targets. The 4DCT describes the mean motion amplitude in EPID movies for the majority of the patients well. However, larger deviations in target motion were observed which necessitate a monitoring during delivery.
REFERENCES 1. Rietzel E, Chen GT, Choi NC, et al. Four-dimensional imagebased treatment planning: Target volume segmentation and dose calculation in the presence of respiratory motion. Int J Radiat Oncol Biol Phys 2005;61:1535–1550. 2. Mageras GS, Pevsner A, Yorke ED, et al. Measurement of lung tumor motion using respiration-correlated CT. Int J Radiat Oncol Biol Phys 2004;60:933–941. 3. Vedam SS, Keall PJ, Kini VR, et al. Determining parameters for respiration-gated radiotherapy. Med Phys 2001;28:2139–2146. 4. Arimura H, Egashira Y, Shioyama Y, et al. Computerized method for estimation of the location of a lung tumor on EPID cine images without implanted markers in stereotactic body radiotherapy. Phys Med Biol 2009;54:665–677. 5. Lin T, Cervino LI, Tang X, et al. Fluoroscopic tumor tracking for image-guided lung cancer radiotherapy. Phys Med Biol 2009;54:981–992. 6. Wilbert J, Meyer J, Baier K, et al. Tumor tracking and motion compensation with an adaptive tumor tracking system (ATTS): System description and prototype testing. Med Phys 2008;35:3911–3921. 7. Berbeco RI, Hacker F, Ionascu D, et al. Clinical feasibility of using an EPID in CINE mode for image-guided verification of stereotactic body radiotherapy. Int J Radiat Oncol Biol Phys 2007;69:258–266. 8. Korreman SS, Juhler-Nottrup T, Boyer AL. Respiratory gated beam delivery cannot facilitate margin reduction, unless combined with respiratory correlated image guidance. Radiother Oncol 2008;86:61–68. 9. Shirato H, Suzuki K, Sharp GC, et al. Speed and amplitude of lung tumor motion precisely detected in four-dimensional setup and in real-time tumor-tracking radiotherapy. Int J Radiat Oncol Biol Phys 2006;64:1229–1236. 10. Guckenberger M, Wulf J, Mueller G, et al. Dose-response relationship for image-guided stereotactic body radiotherapy of pulmonary tumors: Relevance of 4d dose calculation. Int J Radiat Oncol Biol Phys 2009;74:47–54. 11. Meyer J, Richter A, Baier K, et al. Tracking moving objects with megavoltage portal imaging: A feasibility study. Med Phys 2006;33:1275–1280.
12. Lu W, Nystrom MM, Parikh PJ, et al. A semi-automatic method for peak and valley detection in free-breathing respiratory waveforms. Med Phys 2006;33:3634–3636. 13. Suh Y, Dieterich S, Cho B, et al. An analysis of thoracic and abdominal tumour motion for stereotactic body radiotherapy patients. Phys Med Biol 2008;53:3623–3640. 14. Liu HH, Balter P, Tutt T, et al. Assessing respirationinduced tumor motion and internal target volume using four-dimensional computed tomography for radiotherapy of lung cancer. Int J Radiat Oncol Biol Phys 2007;68: 531–540. 15. Seppenwoolde Y, Shirato H, Kitamura K, et al. Precise and realtime measurement of 3D tumor motion in lung due to breathing and heartbeat, measured during radiotherapy. Int J Radiat Oncol Biol Phys 2002;53:822–834. 16. Sonke JJ, Lebesque J, van Herk M. Variability of four-dimensional computed tomography patient models. Int J Radiat Oncol Biol Phys 2008;70:590–598. 17. Tang X, Lin T, Jiang S. A feasibility study of treatment verification using EPID cine images for hypofractionated lung radiotherapy. Phys Med Biol 2009;54:S1–S8. 18. Keall PJ, Mageras GS, Balter JM, et al. The management of respiratory motion in radiation oncology report of AAPM Task Group 76. Med Phys 2006;33:3874–3900. 19. D’Souza WD, Naqvi SA, Yu CX. Real-time intra-fractionmotion tracking using the treatment couch: A feasibility study. Phys Med Biol 2005;50:4021–4033. 20. Keall PJ, Kini VR, Vedam SS, et al. Motion adaptive x-ray therapy: A feasibility study. Phys Med Biol 2001;46:1–10. 21. McQuaid D, Webb S. Target-tracking deliveries using conventional multileaf collimators planned with 4D direct-aperture optimization. Phys Med Biol 2008;53:4013–4029. 22. Wijesooriya K, Bartee C, Siebers JV, et al. Determination of maximum leaf velocity and acceleration of a dynamic multileaf collimator: implications for 4D radiotherapy. Med Phys 2005; 32:932–941. 23. Suh Y, Dieterich S, Keall PJ. Geometric uncertainty of 2D projection imaging in monitoring 3D tumor motion. Phys Med Biol 2007;52:3439–3454.