Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
INSTITUTE OF PHYSICS PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY Phys. Med. Biol. 49 (2004) 425–440 PII: S0031-9155(04)69607-2 Prediction of respiratory tumour motion for real-time image-guided radiotherapy Gregory C Sharp1, Steve B Jiang1, Shinichi Shimizu2 and Hiroki Shirato2 1 Department of Radiation Oncology, Massachusetts General Hospital and Harvard Medical School, Boston, MA 02114, USA 2 Department of Radiation Medicine, Hokkaido University School of Medicine, Sapporo, Japan E-mail: [email protected] Received 26 September 2003 Published 16 January 2004 Online at stacks.iop.org/PMB/49/425 (DOI: 10.1088/0031-9155/49/3/006) Abstract Image guidance in radiotherapy and extracranial radiosurgery offers the potential for precise radiation dose delivery to a moving tumour. Recent work has demonstrated how to locate and track the position of a tumour in real-time using diagnostic x-ray imaging to find implanted radio-opaque markers. However, the delivery of a treatment plan through gating or beam tracking requires adequate consideration of treatment system latencies, including image acquisition, image processing, communication delays, control system processing, inductance within the motor, mechanical damping, etc. Furthermore, the imaging dose given over long radiosurgery procedures or multiple radiotherapy fractions may not be insignificant, which means that we must reduce the sampling rate of the imaging system. This study evaluates various predictive models for reducing tumour localization errors when a realtime tumour-tracking system targets a moving tumour at a slow imaging rate and with large system latencies. We consider 14 lung tumour cases where the peak-to-peak motion is greater than 8 mm, and compare the localization error using linear prediction, neural network prediction and Kalman filtering, against a system which uses no prediction. To evaluate prediction accuracy for use in beam tracking, we compute the root mean squared error between predicted and actual 3D motion. We found that by using prediction, root mean squared error is improved for all latencies and all imaging rates evaluated. To evaluate prediction accuracy for use in gated treatment, we present a new metric that compares a gating control signal based on predicted motion against the best possible gating control signal. We found that using prediction improves gated treatment accuracy for systems that have latencies of 200 ms or greater, and for systems that have imaging rates of 10 Hz or slower. 0031-9155/04/030425+16$30.00 © 2004 IOP Publishing Ltd Printed in the UK 425 426 G C Sharp et al 1. Introduction The goal of radiation therapy is to precisely deliver a lethal dose to the tumour while minimizing the dose to surrounding healthy tissues and critical structures. Recent technical developments such as intensity-modulated radiation therapy (IMRT) have advanced the capability of delivering highly conformal radiation dose distributions to complex three-dimensional (3D) static target volumes. However, internal organ motion and deformation during the treatment, mainly caused by respiration, may greatly degrade the effectiveness and efficiency of IMRT for the management of thoracic and abdominal lesions, especially when the treatment is done in a hypo-fraction or single-fraction manner (Jacobs et al 1996, Engelsman et al 2001, Langen and Jones 2001, Ozhasoglu and Murphy 2002). To deliver the prescribed radiation dose to the entire volume of moving tumour while avoiding high radiation to the adjacent healthy tissues, several techniques have been developed or are under development. The simplest approach is to minimize the motion by breath holding, either passively (Hanley et al 1999, Mah et al 2000, Rosenzweig et al 2000, Barnes et al 2001, Kim et al 2001, Sixel et al 2001) or actively (Wong et al 1999, Stromberg et al 2000). Alternatively, a technology called respiratory gating can be used to reduce the tumour motion during beam-on time by limiting radiation exposure to a portion of the breathing cycle. Respiratory gating can be implemented using either an external respiratory signal to infer the tumour position (Ohara et al 1989, Kubo and Hill 1996, Ramsey et al 1999a, 1999b, Kubo et al 2000, Minohara et al 2000, Vedam et al 2001, Keall et al 2002), or by directly tracking implanted radio-opaque markers fluoroscopically (Shirato et al 1999, 2000a, 2000b, Shimizu et al 2000, 2001, Harada et al 2002). Arguably, a better way to treat a moving target may be to follow the target dynamically with the radiation beam (beam tracking). This was first implemented in a robotic radiosurgery system (Adler et al 1999, Ozhasoglu et al 2000, Schweikard et al 2000, Murphy 2002, Murphy et al 2003). For linacbased radiotherapy, tumour motion can be compensated for using a dynamic multileaf collimator (MLC) (Keall et al 2001, Neicu et al 2003). For beam gating or beam tracking, the precise 3D location of a moving tumour should be available in real-time, and can be obtained through real-time x-ray imaging of fiducial markers during the treatment (Shirato et al 1999, 2000b, 2003, Shimizu et al 2001, Murphy 2002, Berbeco et al 2004). The x-ray images are generally taken at a constant frequency, with each image providing the location of the tumour at a certain instant in time. After the image is taken, but before the radiation beam can target the tumour at its new location, there is always a system latency due to the time needed for image acquisition, image processing, reaction times of the hardware (linac, MLC or Cyberknife), and so on (Litzenberg et al 2002). Let us assume images are taken at a period of t, the latency between imaging and beam reaction is t and we have just taken an image at time t (see figure 1). Because of the imaging rate and latency, decisions made for treatment between time t + t and t + t + t are based on the tumour positions derived from the image at time t, preferably using previous images too. When both the imaging period and system latency are small enough, the tumour positions between t + t and t + t + t can be approximated using its position at time t. However, this approximation can introduce significant localization errors when system latency is not negligible, or when the x-ray imaging frequency is reduced to limit imaging dose. Figure 2 shows the kind of errors that can be introduced by long system latencies (left) and reduced imaging frequency (right). In these plots, the most recent measurement is shown with a dashed line, and the true position is shown with a solid line. Treatment based solely on the most recent measurement will consistently miss the target. Prediction of respiratory tumour motion Image Capture 427 Image Capture Image Capture Image Capture Imaging Rate (∆t) Image processing, Beam reaction, etc. Image processing, Beam reaction, etc. Image processing, Beam reaction, etc. Image processing, Beam reaction, etc. System Latency (∆t') Treatment Treatment time t Treatment time t+∆t' Treatment time time t+∆t'+∆t 2 2 0 0 cranial-caudal motion (mm) cranial-caudal motion (mm) Figure 1. Because of limitations in imaging rate and system latency, the most recent image may not accurately reflect the position of a tumour at the time of treatment. For a system with an imaging period of t and a system latency of t , an image captured at time t (shown in grey) must be used for treatment between time t + t and time t + t + t. 2 4 6 8 10 65 4 6 8 10 Measurements System with latency 64 2 66 time (sec) 67 68 Measurements System with low imaging rate 64 65 66 time (sec) 67 68 Figure 2. Effect of large system latency and slow imaging rate on tumour localization accuracy. Left: a system with latency of 200 ms and imaging rate of 30 Hz. Right: a system with latency of 33 ms and imaging rate of 3 Hz. Early studies using prediction of respiratory motion to overcome system latencies suggest that improvements in targeting accuracies can be realized. Shirato et al (2000b) describe the use of linear extrapolation to estimate the future position of a tumour, and have evaluated the accuracy on phantom measurements. Murphy et al (2002) compare three methods, a tapped delay line filter, a Kalman filter and a neural network, for predicting the future position of a tumour using fluoroscopic simulation data and external markers. In this paper, we investigate the performance of standard prediction algorithms to characterize the predictability of 3D tumour motion for different imaging rates and system latencies. We investigate two kinds of linear filters, two kinds of neural networks and a Kalman filter. Only stationary forms of these predictors are considered. Prediction accuracy is evaluated for the lung tumour motion of 14 patients with peak-to-peak tumour motion greater than 8 mm, as measured during treatment (Shirato et al 2000a, Seppenwoolde et al 2002). From our results, we hope to provide a useful 428 G C Sharp et al estimate of the expected error in tumour localization during treatment for machines with different latencies, and we hope to understand how to use prediction to reduce the fluoroscopic frequency and imaging dose in a real-time image-guided radiotherapy system (Shirato et al 2000a, 2000b, Berbeco et al 2004). 2. Methods In this section we describe the prediction methods that we have evaluated for predicting the 3D marker position: linear prediction, artificial neural networks and the Kalman filter. The methods for optimization of free parameters and the methods for evaluation of predictor performance will also be described. 2.1. Prediction algorithms 2.1.1. Linear prediction. A linear predictor is a system which predicts the future output signal as a linear function of a set of inputs. We consider linear predictors of the form xt = a0 + a1 xt−1 + · · · + an xt−n (1) where xt is the 3D position of the tumour at time t. The position xt is therefore predicted as a linear combination of the known previous positions xt−1 through xt−n . From a set of training samples, the coefficients ak can be found by solving a linear equation to minimize the mean squared error of predictions on a set of training examples. However, determining the best number of histories to use is not a trivial question. We will defer discussion of parameter optimization for section 2.2. 2.1.2. Linear extrapolation. As a special case of linear prediction, we also consider linear extrapolation. Linear extrapolation uses the two most recent samples to find the signal velocity, and predicts that the signal will maintain constant velocity. For a system where images are sampled at a constant rate, we predict that xt = 2xt−1 − xt−2 . (2) Because the constant velocity assumption is only valid for a short period of time, we have only evaluated the linear extrapolation method for low latency systems. 2.1.3. Artificial neural networks. An artificial neural network (ANN) is a function approximation technique modelled on the biological processing of brain neurons (Bishop 1995). Here we consider only predictors that are multilayer perceptrons, with two feedforward stages. This means that there will be two computation stages performed sequentially. The first stage computes a vector of intermediate values (called the hidden layer) from the recent history of positions. The second stage computes the predicted positions from the hidden layer. Both these stages compute a linear transformation, followed by a possibly nonlinear transformation called the activation function. In the first stage, the input layer is a history of the ni most recent 3D positions, but the actual number of input values used is 3ni + 1. This is because in addition to the three 3D coordinates of the marker for each history used, there is an additional input with a constant value of 1. This additional input is called a bias unit, and it is used to bias the linear portion of the computation. A sigmoid function is then used as the nonlinear activation function to Prediction of respiratory tumour motion 429 generate the nh values of the hidden layer. With input values denoted xi and hidden values denoted yj , the equation for the first layer is given as yj = 1 . 3ni +1 1 + exp − i=1 vij xi (3) In this stage, the free parameters to be optimized are the weights vij . In the second layer of the neural network, predicted positions are computed from the hidden layer. To predict no future 3D positions, we generate 3no output coordinate values; nh + 1 intermediate values are used, including a bias node as before. The activation function we use in the second stage is linear. Therefore, with hidden values denoted yj and output values denoted zk , the equation for the second layer is given as zk = n h +1 wj k yj . (4) i=1 In this stage, the free parameters to be optimized are the weights wj k . 2.1.4. Predicting slope with an ANN. As a special case of the ANN prediction we consider a network which predicts not the future position of the marker, but the future velocity of the marker. In other words, the ANN inputs are a history of positions, but the output is the velocity to the next position. The future position is then predicted by extrapolating the current position along the predicted velocity. In our experiments, this method will be referred to as ‘ANN (slope)’. Because this method is also based on the constant velocity assumption, we only evaluate the ANN (slope) method for low latency systems. 2.1.5. The Kalman filter. The Kalman filter is a method for estimation of the internal state of a linear dynamic system (Kalman 1960). An excellent introduction is given by Welch and Bishop (1995). In using the Kalman filter to predict the 3D marker position, we assume that the marker motion is generated by a linear dynamic system model with no dependence on a control input. The equations for such a system are given by xt = Axt−1 + w (5) zt = Bxt + v (6) where xt is the internal state of the system at time t, A is the state transition matrix which describes how the state changes over time, w is normally distributed process noise with covariance Q, zt is the observation at time t, B is the measurement matrix which describes how the observation depends on internal state and v is normally distributed measurement noise with covariance R. The measurement zt is the 3D position of the tumour as observed by the x-ray imaging system. The Kalman filter generates an optimal estimate for the state xt by combining the previous estimate xt−1 , its covariance Pt−1 and the current measurement zt . As shown in figure 3, a prediction of the internal state, denoted x̂t , is computed from the previous estimate of the state xt−1 . From x̂t , we can predict the future value of the measurement using equation (6), and use it for treatment. Later, when the actual measurement arrives, the internal state can be updated to its optimal estimate xt . Kalman filter estimation requires complete knowledge of the dynamic system. Specifically, the transition matrix A, the measurement matrix B and the noise covariance matrices Q and R must be known. In this study, we treat the state vector x as an abstract 430 G C Sharp et al Previous Estimated State State Prediction Predicted State Measurement Prediction Predicted Tumour Location Treatment Delivery State Estimation Measured Tumour Location Imaging System Estimated State Figure 3. The Kalman filter generates an estimate of the current state of a linear dynamic system from the previous estimated state and the measurement. Before the actual measurement arrives, a predicted measurement ẑt can be computed from the predicted state and used for treatment. unknown state of fixed dimension, and we treat the dynamic system matrices as unknowns to be optimized. The matrices are optimized by estimating their maximum-likelihood values using expectation maximization (EM) on a set of training samples (Digalakis et al 1993). 2.2. Parameter tuning Each of the prediction methods described above, with the exception of linear extrapolation, has a number of parameters that must be determined. We distinguish between two groups of parameters, which we call configuration parameters and free parameters. The configuration parameters are those which are generally considered to be design criteria, while the free parameters are those generally solved by numeric optimization. For example, consider the linear predictor. The number of previous histories to use is a design decision, but the weights given to each history can be solved by linear regression. A list of the configuration parameters and free parameters for each method is given in table 1. Finding the best values for the configuration parameters is a question of model selection, which is not an easy one. Various strategies have been developed to address this problem, such as Akaike information criterion, minimum description length, cross-validation and bootstrapping (e.g. Duda et al 2001). Solving for the best configuration parameters is beyond the scope of this paper. Instead, we simply choose configuration parameters that minimize error over the entire patient population, based on running multiple off-line simulations. Our configuration parameter values are therefore tuned to perform well at a given latency and imaging rate, but may not be well tuned for a given patient. Typical values used are between two and five histories for the ANN and linear predictors, between five and ten hidden units for the ANN and between three and five dimensions for the Kalman filter state vector. Prediction of respiratory tumour motion 431 Gather training data Compute parameters Warm-up predictor Treatment with prediction 15 sec 2 sec 4 sec As required Figure 4. Training of free parameters is performed for each treatment fraction, and requires time for gathering data, computing predictor parameters and warming up the predictor. Table 1. Configuration parameters, free parameters and optimization methods used in this study. Prediction method Configuration parameters Free parameters Optimization method Linear Linear extrapolation ANN, ANN (slope) Kalman filter Number of histories None a (equation (1)) None Regression None Number of histories, number of hidden units Dimension of state vector vij , wj k (equations (3), (4)) A, B, Q, R (equations (5), (6)) Conjugate gradient Expectation maximization Once the configuration parameters are known, the free parameters are easily solved using standard numerical optimization algorithms. To account for different breathing patterns in different patients, we estimate these parameters on a fraction by fraction basis, according to the fixed schedule described in figure 4. At the beginning of each fraction, we use the first 15 s (about four to five breaths) to gather training data which will be used to optimize the free parameters. Next, we allocate exactly 2 s of computer processing time to perform the optimization. The free parameters are fixed based on this computation, and are not updated using subsequent data. Next, 4 s are allocated to ‘warm-up’ the predictors. The warming-up period includes the time needed to gather the recent history of data points, or time to converge to a stable state vector in the case of the Kalman filter. Finally, prediction accuracy is evaluated by running the predictor on the remaining data. 2.3. Interpolation Although the prediction methods described above can be used for prediction of an unevenly sampled time series, we restrict our attention to uniform imaging rates. However, from a uniform imaging rate we have no knowledge of the shape of the signal between samples. Therefore, we implement a simple linear interpolation scheme to fill in the missing values between predictions, as shown in figure 5. Depending upon the amount of latency and the sampling rate, between one and three future marker positions will be predicted. The signal is then interpolated between the most recent sample and these future positions to generate a 30 Hz prediction over the prediction interval. 2.4. Evaluating accuracy We consider two metrics for evaluating the accuracy of the predictions: root mean squared (rms) error and gating error. Root mean squared error is useful as a measure of accuracy for 432 G C Sharp et al latency interval original 30 Hz measurements prediction interval prediction interpolated to 30 Hz predicted future 3 Hz measurements 3 Hz measurements Figure 5. Low frame rate predictions are interpolated up to full frame rate over a prediction window by making multiple future predictions and interpolating between them. 6mm Gating Window True Signal Predicted Signal Ideal Gating Signal Predicted Gating Control Signal 68 70 72 74 76 78 time (sec) Figure 6. The gating error metric compares the ideal gating control signal against the gating control signal determined by the predicted motion. a treatment delivery method that tracks the tumour with the radiation beam (Ozhasoglu et al 2000, Keall et al 2001, Neicu et al 2003). To combine the error results for different patients, we compute rms error for each patient separately, and average the results over all patients. On the other hand, to provide a measure of accuracy for gated treatment (Shirato et al 1999, Shimizu et al 2000), we introduce the concept of gating error. Ideally, a gating system should enable a gating control signal and begin treatment when the radio-opaque marker enters a predetermined gating window, and disable the control signal when the marker leaves the gating window. The gating error metric quantifies the error in the control of the gating signal. It is computed by comparing a gating control signal based on the predicted motion against the best possible gating control signal (see figure 6). For the experiments shown in this study, we use a gating window which is a 6 mm diameter 3D cubic volume centred at the exhale position of the setup (cf Shimizu et al 2001). Two kinds of gating errors are possible: false negatives where the marker is in the window but the prediction is out of the window, and false positives where the marker is out of the window but the prediction is within of the window. From the false negatives, we can compute the duty cycle of a treatment as a percentage of the maximum possible duty cycle. Let FN be the number of false negatives, and let Ptot be the total number of time steps when the marker was actually within the gating volume. Then, we have Ptot − FN . (7) duty cycle (% of max) = Ptot Prediction of respiratory tumour motion 433 Table 2. Mean and standard deviation of breathing amplitude for the 14 lung tumour patients used in this study. Patient ID Mean motion amplitude (mm) Standard deviation (mm) 1 6 7 9 11 13 17 22.2 13.2 15.7 14.1 9.1 16.3 12.5 4.9 3.3 3.6 3.0 0.8 2.6 2.2 Patient ID Mean motion amplitude (mm) Standard deviation (mm) 18 26 30 31 34 38 40 11.6 11.9 27.5 14.3 14.7 13.8 31.6 1.5 1.3 5.3 1.7 3.3 1.7 2.3 Thus, the duty cycle describes the reduction in gated treatment efficiency due to prediction errors. From the false positives, we can compute the gating failure rate, which is the number of incorrect beam-on predictions as a percentage of total beam-on predictions. Let FP be the number of false positives, and let Ppred be the total number of time steps that the beam was predicted to be within the gating window. Then, we have gating failure rate = FP . Ppred (8) Thus, the gating failure rate quantifies the incorrect beam-on time during gated treatment due to prediction errors. Both duty cycle and gating failure rate should be considered when comparing prediction methods. 3. Materials We performed this study retrospectively using lung tumour movement data of patients treated using the real-time tumour-tracking (RTRT) system at Hokkaido University (Shirato et al 1999, 2000b, Shimizu et al 2001, Seppenwoolde et al 2002). In this system, a 1.5 mm diameter gold marker is inserted into or near the lung tumour mass through bronchofibrescopy prior to treatment. The marker is imaged at a rate of 30 Hz using diagnostic fluoroscopy and image processing computers to determine the three-dimensional coordinates of the marker in real-time. Lung tumour motion data of 40 patients treated between 2001 and 2002 were analysed, and the 14 patients with breathing amplitude greater than 8 mm were identified and used to evaluate motion prediction accuracy. The average and standard deviation of the breathing amplitudes of these 14 patients are shown in table 2. Evaluation of accuracy was based on predictions made after the predictor was trained and warmed up, i.e. the fourth stage of figure 4. The duration of the evaluation period varied between 48 and 342 s, with an average evaluation period of 115 s. A qualitative examination of the breathing patterns of the 14 patients shows that there are significant differences between patients. An example of the breathing signals from four different patients is shown in figure 7. Figure 7(a) shows a breathing pattern that is regular in both period and amplitude. Regular breathing represents an ideal case for prediction, because the motion pattern used for training continues during treatment. Figure 7(b) highlights the fact that the raw tracking signal is often noisy. Two kinds of noise are present: an approximately Gaussian noise of around 1–2 mm, and a spike noise of up to 10 mm. Figure 7(c) shows 434 G C Sharp et al 10 L-R L-R 20 5 S-I 10 Motion (mm) Motion (mm) 0 -5 -10 S-I 0 -10 A-P A-P -15 -20 -20 -30 80 100 Time (secs) 44 46 Time (secs) (a) (b) 120 140 160 38 180 40 42 48 50 52 15 20 L-R 10 L-R 10 S-I 0 S-I Motion (mm) Motion (mm) 5 0 -5 -20 -30 A-P -10 -10 A-P -40 -15 -50 -20 10 15 20 25 30 Time (secs) (c) 35 40 45 80 90 100 110 Time (secs) 120 130 (d) Figure 7. Examples of tumour motion and signal quality: (a) regular breathing; (b) noisy signal; (c) breathing with cardiac motion; (d) drift in position and amplitude. the motion pattern of a tumour that is influenced by both breathing and cardiac motion. Figure 7(d) demonstrates a breathing pattern where both the exhale home position and the amplitude drift over time. Prediction of respiratory tumour motion 8 Extrapolation Linear ANN ANN (slope) Kalman filter No prediction 7 RMS Error (mm) 435 6 5 4 3 2 1 0 Latency = 33 ms 30 8 1 Linear ANN Kalman filter No prediction 7 RMS Error (mm) 10 3 Imaging Rate (Hz) 6 5 4 3 2 1 0 Latency = 200 ms 30 10 3 Imaging Rate (Hz) 1 12 RMS Error (mm) 10 8 6 4 Linear ANN Kalman filter No prediction 2 0 30 Latency = 1 sec 10 3 Imaging Rate (Hz) 1 Figure 8. Root mean squared error of tumour position predictions as a function of imaging frequency for (top) 33 ms latency, (middle) 200 ms latency and (bottom) 1 s latency. 4. Experimental results and discussion Figure 8 shows the rms error averaged over all patients as a function of imaging rate for each algorithm. Plots for three different latencies are shown: 33 ms (top), 200 ms (middle) and 436 G C Sharp et al 1 s (bottom). Results for linear extrapolation and ANN (slope) methods are computed only for the case of 33 ms latency. We only consider this case for these two methods because both methods are based on extrapolation of the motion velocity, which is only useful for a short time. It is seen that most forms of prediction have lower rms error compared to no prediction. One notable exception is the case of linear extrapolation, which has a higher error at all imaging rates. Prediction using linear extrapolation has been previously tested in phantom experiments for the RTRT system at Hokkaido University (Shirato et al 2000b), but has not been used for treatment because of our awareness about these potential errors. The reason for the poor prediction accuracy of linear extrapolation is that the random signal noise such as that shown in figure 7 makes it difficult to estimate the motion velocity from two points. This noise occurs when the pattern matching software is employed on poor contrast images such as those which occur when the marker is obscured by the bony skeleton. In contrast to linear extrapolation, the other predictors use multiple history readings, and therefore they can both smooth and predict the data. We also note the relatively worse performance of the Kalman filter predictor when compared with linear and ANN predictors. This may reflect the difficulty in estimating the state transition matrix from such a small amount of data. Therefore, we expect that a handtuned Kalman filter will perform better than the Kalman filter derived from EM estimation. In terms of accuracy, the best predictor allows us to achieve an rms error of 2 mm at sampling rates as low as 3–4 Hz for the low latency (33 ms) case. A similar error rate is possible at 30 Hz with 200 ms latency. However, with a 1 s latency we were not able to achieve less than 5 mm rms error using the methods we investigated. Figure 9 shows the gating error averaged over all patients as a function of imaging rate for each algorithm. The plots on the left show the duty cycle achieved as a percentage of the maximum duty cycle, while the plots on the right show the gating failure rate (see section 3). Plots for the same three latencies are shown: 33 ms (top), 200 ms (middle) and 1 s (bottom). From the plots it is seen that at high frame rates and low latencies, gating failure rates are worse when using prediction than when not using prediction. However, when using prediction there is also a corresponding increase in duty cycle suggesting that actual performance may be similar. This represents a trade-off between duty cycle and gating failure rates that can be realized by adjusting the size of the gating window. At slower imaging rates, it is seen that prediction improves the gating failure rate with a similar duty cycle. With longer latencies, gating failure rates are still improved by using prediction, but the duty cycle becomes considerably degraded, suggesting that prediction reliability may be insufficient. One important consideration that we did not address within this study is the effect of the non-stationary nature of breathing. Because we used only the first 15 s for training data, and we did not update the predictor parameters online, it is likely that the prediction quality will degrade over time as the breathing pattern changes. We expect this effect to be greater for patients whose breathing patterns notably drift over time, and less for patients with regular breathing. Prediction failures due to changes in breathing patterns can never be eliminated, but they might be reduced by using the adaptive versions of these prediction algorithms. 5. Summary In this paper we have presented a study on the use of prediction to compensate for system latencies and reduced imaging rates in real-time image-guided radiation treatment. Five prediction methods were compared for low latency systems, and three prediction methods Prediction of respiratory tumour motion 437 14 Extrapolation Linear ANN ANN (slope) Kalman filter No prediction 90 Gating Failure Rate (%) Gating Duty Cycle (% of max) 12 80 70 60 Extrapolation Linear ANN ANN (slope) Kalman filter No prediction 50 8 6 4 2 Latency = 33 ms Latency = 33 ms 40 30 10 10 3 Imaging Rate (Hz) 0 1 30 10 3 Imaging Rate (Hz) 1 20 Linear ANN Kalman filter No prediction 16 80 Gating Failure Rate (%) Gating Duty Cycle (% of max) 90 70 60 Linear ANN Kalman filter No prediction 50 12 8 4 Latency = 200 ms Latency = 200 ms 0 30 10 3 Imaging Rate (Hz) 1 30 10 3 Imaging Rate (Hz) 1 40 Linear ANN Kalman filter No prediction Gating Failure Rate (%) Gating Duty Cycle (% of max) 60 50 40 30 Linear ANN Kalman filter No prediction 20 30 10 3 Imaging Rate (Hz) 30 20 10 Latency = 1 sec 1 Latency = 1 sec 0 30 10 3 Imaging Rate (Hz) 1 Figure 9. Gating error as a function of imaging frequency for (top) 33 ms latency, (middle) 200 ms latency and (bottom) 1 s latency. These plots show (left) the percentage of the maximum possible efficiency achieved, and (right) the percentage of time the beam would be enabled when the tumour is outside the gating window. were compared for high latency systems. Two different metrics were used to evaluate the prediction methods: an rms error metric to compare the absolute accuracy of the prediction and a gating error metric to compare the accuracy of gated treatment based on the predictions. In all cases, we found that using prediction results in a lower rms error when compared with not using prediction. This statement must be qualified, however, to exclude simple 438 G C Sharp et al extrapolation (at all imaging rates) and ANN prediction (at high imaging rates). When comparing gating error, we found that in most cases using prediction results in fewer cases where the beam is mistakenly enabled. The exception to this statement is made when both the latency is low (33 ms), and the imaging rate is high (10 Hz or faster). In these cases, there is no clear improvement in gating error when using prediction. The prediction methods used in this paper represent only simple, commonly used algorithms and are not meant to be an exhaustive study. Additional improvements will be made by incorporating data gathered from prior knowledge, physiological insight, online parameter adaptation, outlier detection, optimizations targeted towards reducing gating error and nonlinear filtering. In addition, we look forward to future developments that use nonuniform imaging to selectively acquire images according to targeting uncertainty. Acknowledgments This work is partially supported by Varian Medical Systems Inc. and the Whitaker Foundation. The authors thank Dr Ross Berbeco, Dr Toni Neicu, Dan Ruan and Dr David Castañón for valuable discussions. Artificial neural network training was accomplished using the netlab software package provided by Ian Nabney and Christopher Bishop (2003). EM fitting of Kalman filter parameters was accomplished using the the Kalman filter software package written by Kevin Murphy (2003). References Adler J R Jr, Murphy M J, Chang S D and Hancock S L 1999 Image-guided robotic radiosurgery Neurosurgery 44 1299–306 (discussion 1306–7) Barnes E A, Murray B R, Robinson D M, Underwood L J, Hanson J and Roa W H 2001 Dosimetric evaluation of lung tumor immobilization using breath hold at deep inspiration Int. J. Radiat. Oncol. Biol. Phys. 50 1091–8 Berbeco R I, Jiang S B, Sharp G C, Chen G T Y, Mostafavi H and Shirato H 2004 Integrated radiotherapy imaging system (IRIS): design considerations of tumor tracking with linac gantry-mounted kV x-ray systems Phys. Med. Biol. 49 243–55 Bishop C 1995 Neural Networks for Pattern Recognition (Oxford: Oxford University Press) Digalakis V, Rohlicek J R and Ostendorf M 1993 ML estimation of a stochastic linear system with the EM algorithm and its application to speech recognition IEEE Trans. Speech Audio Process. 1 431–42 Duda R O, Hart P E and Stork D G 2001 Pattern Classification (New York: Wiley) Engelsman M, Damen E M, De Jaeger K, van Ingen K M and Mijnheer B J 2001 The effect of breathing and set-up errors on the cumulative dose to a lung tumor Radiother. Oncol. 60 95–105 Hanley J et al 1999 Deep inspiration breath-hold technique for lung tumors: the potential value of target immobilization and reduced lung density in dose escalation Int. J. Radiat. Oncol. Biol. Phys. 45 603–11 Harada T, Shirato H, Ogura S, Oizumi S, Yamazaki K, Shimizu S, Onimaru R, Miyasaka K, Nishimura M and Dosaka-Akita11 H 2002 Real-time tumor-tracking radiation therapy for lung carcinoma by the aid of insertion of a gold marker using bronchofiberscopy Cancer 95 1720–7 Jacobs I, Vanregemorter J and Scalliet P 1996 Influence of respiration on calculation and delivery of the prescribed dose in external radiotherapy Radiother. Oncol. 39 123–8 Kalman R E 1960 A new approach to linear filtering and prediction problems Trans. ASME D 82 35–45 Keall P J, Kini V R, Vedam S S and Mohan R 2001 Motion adaptive x-ray therapy: a feasibility study Phys. Med. Biol. 46 1–10 Keall P J, Kini V R, Vedam S S and Mohan R 2002 Potential radiotherapy improvements with respiratory gating Australas. Phys. Eng. Sci. Med. 25 1–6 Kim D J, Murray B R, Halperin R and Roa W H 2001 Held-breath self-gating technique for radiotherapy of nonsmall-cell lung cancer: a feasibility study Int. J. Radiat. Oncol. Biol. Phys. 49 43–9 Kubo H D and Hill B C 1996 Respiration gated radiotherapy treatment: a technical study Phys. Med. Biol. 41 83–91 Kubo H D, Len P M, Minohara S and Mostafavi H 2000 Breathing-synchronized radiotherapy program at the University of California Davis Cancer Center Med. Phys. 27 346–53 Prediction of respiratory tumour motion 439 Langen K M and Jones D T 2001 Organ motion and its management Int. J. Radiat. Oncol. Biol. Phys. 50 265–78 Litzenberg D W, Moran J M and Fraass B A 2002 Incorporation of realistic delivery limitations into dynamic MLC treatment delivery Med. Phys. 29 810–20 Mah D, Hanley J, Rosenzweig K E, Yorke E, Braban L, Ling C C, Leibel S A and Mageras G 2000 Technical aspects of the deep inspiration breath-hold technique in the treatment of thoracic cancer Int. J. Radiat. Oncol. Biol. Phys. 48 1175–85 Minohara S, Kanai T, Endo M, Noda K and Kanazawa M 2000 Respiratory gated irradiation system for heavy-ion radiotherapy Int. J. Radiat. Oncol. Biol. Phys. 47 1097–103 Murphy M J 2002 Fiducial-based targeting accuracy for external-beam radiotherapy Med. Phys. 29 334–44 Murphy K 2003 Kalman Filter Toolbox For Matlab http://www.ai.mit.edu/˜murphyk/Software/Kalman/kalman.html Murphy M J, Chang S D, Gibbs I C, Le Q T, Hai J, Kim D, Martin D P and Adler J R Jr 2003 Patterns of patient movement during frameless image-guided radiosurgery Int. J. Radiat. Oncol. Biol. Phys. 55 1400–8 Murphy M J, Jalden J and Isaksson M 2002 Adaptive filtering to predict lung tumor breathing motion during imageguided radiation therapy Proc. 16th Int. Conf. on Computer Assisted Radiology (CARS 2002) (June 2002) ed H U Lemke, K Inamura, M W Vannier, A G Farman and K Doi Nabney I and Bishop C 2003 Netlab Neural Network Software http://www.ncrg.aston.ac.uk/netlab Neicu T, Shirato H, Seppenwoolde Y and Jiang S B 2003 Synchronized moving aperture radiation therapy (SMART): average tumor trajectory for lung patients Phys. Med. Biol. 48 587–98 Ohara K, Okumura T, Akisada M, Inada T, Mori T, Yokota H and Calaguas M J 1989 Irradiation synchronized with respiration gate Int. J. Radiat. Oncol. Biol. Phys. 17 853–7 Ozhasoglu C and Murphy M J 2002 Issues in respiratory motion compensation during external-beam radiotherapy Int. J. Radiat. Oncol. Biol. Phys. 52 1389–99 Ozhasoglu C, Murphy M J, Glosser G, Bodduluri M, Schweikard A, Forster K M, Martin D P and Adler J R 2000 Real-time tracking of the tumor volume in precision radiotherapy and body radiosurgery—a novel approach to compensate for respiratory motion Proc. 14th Int. Conf. on Computer Assisted Radiology and Surgery (CARS 2000) (San Francisco, CA, USA) ed H U Lemke, M W Vannier, K Inamura, A G Farman and K Doi pp 691–6 Ramsey C R, Cordrey I L and Oliver A L 1999a A comparison of beam characteristics for gated and nongated clinical x-ray beams Med. Phys. 26 2086–91 Ramsey C R, Scaperoth D, Arwood D and Oliver A L 1999b Clinical efficacy of respiratory gated conformal radiation therapy Med. Dosim. 24 115–9 Rosenzweig K E et al 2000 The deep inspiration breath-hold technique in the treatment of inoperable non-small-cell lung cancer Int. J. Radiat. Oncol. Biol. Phys. 48 81–7 Schweikard A, Glosser G, Bodduluri M, Murphy M J and Adler J R 2000 Robotic motion compensation for respiratory movement during radiosurgery Comput. Aided Surg. 5 263–77 Seppenwoolde Y, Shirato H, Kitamura K, Shimizu S, van Herk M, Lebesque J V and Miyasaka K 2002 Precise and real-time measurement of 3D tumor motion in lung due to breathing and heartbeat, measured during radiotherapy Int. J. Radiat. Oncol. Biol. Phys. 53 822–34 Shimizu S, Shirato H, Kitamura K, Shinohara N, Harabayashi T, Tsukamoto T, Koyanagi T and Miyasaka K 2000 Use of an implanted marker and real-time tracking of the marker for the positioning of prostate and bladder cancers Int. J. Radiat. Oncol. Biol. Phys. 48 1591–7 Shimizu S, Shirato H, Ogura S, Akita-Dosaka H, Kitamura K, Nishioka T, Kagei K, Nishimura M and Miyasaka K 2001 Detection of lung tumor movement in real-time tumor-tracking radiotherapy Int. J. Radiat. Oncol. Biol. Phys. 51 304–10 Shirato H, Shimizu S, Shimizu T, Nishioka T and Miyasaka K 1999 Real-time tumor-tracking radiotherapy Lancet 353 1331–2 Shirato H et al 2000a Four-dimensional treatment planning and fluoroscopic real-time tumor tracking radiotherapy for moving tumor Int. J. Radiat. Oncol. Biol. Phys. 48 435–42 Shirato H et al 2000b Physical aspects of a real-time tumor-tracking system for gated radiotherapy Int. J. Radiat. Oncol. Biol. Phys. 48 1187–95 Shirato H et al 2003 Feasibility of insertion/implantation of 2.0-mm-diameter gold internal fiducial markers for precise setup and real-time tumor tracking in radiotherapy Int. J. Radiat. Oncol. Biol. Phys. 56 240–7 Sixel K E, Aznar M C and Ung Y C 2001 Deep inspiration breath hold to reduce irradiated heart volume in breast cancer patients Int. J. Radiat. Oncol. Biol. Phys. 49 199–204 Stromberg J S, Sharpe M B, Kim L H, Kini V R, Jaffray D A, Martinez A A and Won J W 2000 Active breathing control (ABC) for Hodgkin’s disease: reduction in normal tissue irradiation with deep inspiration and implications for treatment Int. J. Radiat. Oncol. Biol. Phys. 48 797–806 440 G C Sharp et al Vedam S S, Keall P J, Kini V R and Mohan R 2001 Determining parameters for respiration-gated radiotherapy Med. Phys. 28 2139–46 Welch G and Bishop G 1995 An introduction to the Kalman filter Technical Report TR 95-041 University of North Carolina at Chapel Hill (Online at http://www.cs.unc.edu/˜welch/kalman) Wong J W, Sharpe M B, Jaffray D A, Kini V R, Robertson J M, Stromberg J S and Martinez A A 1999 The use of active breathing control (ABC) to reduce margin for breathing motion Int. J. Radiat. Oncol. Biol. Phys. 44 911–9