Inferential modeling and predictive feedback control in real-time motion compensation using the treatment couch during radiotherapy

Technical Note: Comprehensive performance tests of the first clinical real-time motion tracking and compensation system using MLC and jaws

Purpose

To evaluate the performance of the first clinical real‐time motion tracking and compensation system using multileaf collimator (MLC) and jaws during helical tomotherapy delivery.

Methods

Appropriate mechanical and dosimetry tests were performed on the first clinical real‐time motion tracking system (Synchrony on Radixact, Accuray Inc) recently installed in our institution. kV radiography dose was measured by CTDIw using a pencil chamber. Changes of beam characteristics with jaw offset and MLC leaf shift were evaluated. Various dosimeters and phantoms including A1SL ion chamber (Standard Imaging), Gafchromic EBT3 films (Ashland), TomoPhantom (Med Cal), ArcCheck (Sun Nuclear), Delta4 (ScandiDos), with fiducial or high contrast inserts, placed on two dynamical motion platforms (CIRS dynamic motion‐CIRS, Hexamotion‐ScandiDos), were used to assess the dosimetric accuracy of the available Synchrony modalities: fiducial tracking with nonrespiratory motion (FNR), fiducial tracking with respiratory modeling (FR), and fiducial free (e.g., lung tumor tracking) with respiratory modeling (FFR). Motion detection accuracy of a tracking target, defined as the difference between the predicted and instructed target positions, was evaluated with the root mean square (RMS). The dose accuracy of motion compensation was evaluated by verifying the dose output constancy and by comparing measured and planned (predicted) three‐dimensional (3D) dose distributions based on gamma analysis.

Results

The measured CTDIw for a single radiograph with a 120 kVp and 1.6 mAs protocol was 0.084 mGy, implying a low imaging dose of 8.4 mGy for a typical Synchrony motion tracking fraction with 100 radiographs. The dosimetric effect of the jaw swing or MLC leaf shift was minimal on depth dose (<0.5%) and was <2% on both beam profile width and output for typical motions. The motion detection accuracies, that is, RMS, were 0.84, 1.13, and 0.48 mm for FNR, FR, and FFR, respectively, well within the 1.5 mm recommended tolerance. Dose constancy with Synchrony was found to be within 2%. The gamma passing rates of 3D dose measurements for a variety of Synchrony plans were well within the acceptable level.

Conclusions

The motion tracking and compensation using kV radiography, MLC shifting, and jaw swing during helical tomotherapy delivery was tested to be mechanically and dosimetrically accurate for clinical use.

Patient specific methods for room-mounted x-ray imagers for monoscopic/stereoscopic prostate motion monitoring

PURPOSE

To investigate the improvement of combined monoscopic/stereoscopic prostate motion monitoring with room-mounted dual x-ray systems by adopting patient specific methods.

METHODS

The linac couch was used as a motion stage to simulate 40 highly dynamic real patient prostate trajectories. For each trajectory, 40 s pretreatment and 120 s treatment periods were extracted to represent a typical treatment fraction. Motion was monitored via continuous stereoscopic x-ray imaging of a single gold fiducial and images were retrospectively divided into periods of stereoscopic and monoscopic imaging to simulate periodic blocking of the room-mounted system by the gantry during arc-based therapy. The accuracy of the combined motion monitoring was assessed by comparison with the linac couch log files. To estimate 3-D marker position during monoscopic imaging, the use of population statistics was compared to both maximum likelihood estimation and stereoscopic localization based estimation of individualized prostate probability density functions (PDFs) from the pretreatment period. The inclusion of intrafraction updating was compared to pretreatment initialization alone.

RESULTS

Combined mono/stereoscopic localization was successfully implemented. During the transitions from stereoscopic to monoscopic imaging, fiducial localization exhibits sharp discontinuities when population PDFs were employed. Patient specific PDFs successfully reduced the localization error when estimated from stereoscopic localizations, whereas maximum likelihood estimation (MLE) was too unstable in the room-mounted geometry. Intrafraction stereoscopic updating provided further increases in accuracy. Residual error tended to decrease throughout the treatment fraction, as the patient-specific PDFs became more refined.

CONCLUSIONS

This is the first demonstration of toggled monoscopic/stereoscopic localization using room-mounted dual x-ray imagers, enabling continuous intrafraction motion monitoring for these systems. We showed that both pretreatment individualization and intrafraction updating should be used to provide the most accurate motion monitoring.

Continuous monitoring of prostate position using stereoscopic and monoscopic kV image guidance

PURPOSE

To demonstrate continuous kV x-ray monitoring of prostate motion using both stereoscopic and monoscopic localizations, assess the spatial accuracy of these techniques, and evaluate the dose delivered from the added image guidance.

METHODS

The authors implemented both stereoscopic and monoscopic fiducial localizations using a room-mounted dual oblique x-ray system. Recently developed monoscopic 3D position estimation techniques potentially overcome the issue of treatment head interference with stereoscopic imaging at certain gantry angles. To demonstrate continuous position monitoring, a gold fiducial marker was placed in an anthropomorphic phantom and placed on the Linac couch. The couch was used as a programmable translation stage. The couch was programmed with a series of patient prostate motion trajectories exemplifying five distinct categories: stable prostate, slow drift, persistent excursion, transient excursion, and high frequency excursions. The phantom and fiducial were imaged using 140 kVp, 0.63 mAs per image at 1 Hz for a 60 s monitoring period. Both stereoscopic and monoscopic 3D localization accuracies were assessed by comparison to the ground-truth obtained from the Linac log file. Imaging dose was also assessed, using optically stimulated luminescence dosimeter inserts in the phantom.

RESULTS

Stereoscopic localization accuracy varied between 0.13 ± 0.05 and 0.33 ± 0.30 mm, depending on the motion trajectory. Monoscopic localization accuracy varied from 0.2 ± 0.1 to 1.1 ± 0.7 mm. The largest localization errors were typically observed in the left-right direction. There were significant differences in accuracy between the two monoscopic views, but which view was better varied from trajectory to trajectory. The imaging dose was measured to be between 2 and 15 μGy/mAs, depending on location in the phantom.

CONCLUSIONS

The authors have demonstrated the first use of monoscopic localization for a room-mounted dual x-ray system. Three-dimensional position estimation from monoscopic imaging permits continuous, uninterrupted intrafraction motion monitoring even in the presence of gantry rotation, which may block kV sources or imagers. This potentially allows for more accurate treatment delivery, by ensuring that the prostate does not deviate substantially from the initial setup position.

Toward correcting drift in target position during radiotherapy via computer-controlled couch adjustments on a programmable Linac

PURPOSE

Real-time tracking of respiratory target motion during radiation therapy is technically challenging, owing to rapid and possibly irregular breathing variations. The authors report on a method to predict and correct respiration-averaged drift in target position by means of couch adjustments on an accelerator equipped with such capability.

METHODS

Dose delivery is broken up into a sequence of 10 s field segments, each followed by a couch adjustment based on analysis of breathing motion from an external monitor as a surrogate of internal target motion. Signal averaging over three respiratory cycles yields a baseline representing target drift. A Kalman filter predicts the baseline position 5 s in advance, for determination of the couch correction. The method's feasibility is tested with a motion phantom programmed according to previously recorded patient signals. Computed couch corrections are preprogrammed into a research mode of an accelerator capable of computer-controlled couch translations synchronized with the motion phantom. The method's performance is evaluated with five cases recorded during hypofractionated treatment and five from respiration-correlated CT simulation, using a root-mean-squared deviation (RMSD) of the baseline from the treatment planned position.

RESULTS

RMSD is reduced in all 10 cases, from a mean of 4.9 mm (range 2.7-9.4 mm) before correction to 1.7 mm (range 0.7-2.3 mm) after correction. Treatment time is increased ∼5% relative to that for no corrections.

CONCLUSIONS

This work illustrates the potential for reduction in baseline respiratory drift with periodic adjustments in couch position during treatment. Future treatment machine capabilities will enable the use of "on-the-fly" couch adjustments during treatment.

A novel four-dimensional radiotherapy planning strategy from a tumor-tracking beam's eye view

To investigate the feasibility of four-dimensional radiotherapy (4DRT) planning from a tumor-tracking beam's eye view (ttBEV) with reliable gross tumor volume (GTV) delineation, realistic normal tissue representation, high planning accuracy and low clinical workload, we propose and validate a novel 4D conformal planning strategy based on a synthesized 3.5D computed tomographic (3.5DCT) image with a motion-compensated tumor. To recreate patient anatomy from a ttBEV in the moving tumor coordinate system for 4DRT planning (or 4D planning), the centers of delineated GTVs in all phase CT images of 4DCT were aligned, and then the aligned CTs were averaged to produce a new 3.5DCT image. This GTV-motion-compensated CT contains a motionless target (with motion artifacts minimized) and motion-blurred normal tissues (with a realistic temporal density average). Semi-automatic threshold-based segmentation of the tumor, lung and body was applied, while manual delineation was used for other organs at risk (OARs). To validate this 3.5DCT-based 4D planning strategy, five patients with peripheral lung lesions of small size (<5 cm(3)) and large motion range (1.2-3.5 cm) were retrospectively studied for stereotactic body radiotherapy (SBRT) using 3D conformal radiotherapy planning tools. The 3.5DCT-based 4D plan (3.5DCT plan) with 9-10 conformal beams was compared with the 4DCT-based 4D plan (4DCT plan). The 4DCT plan was derived from multiple 3D plans based on all phase CT images, each of which used the same conformal beam configuration but with an isocenter shift to aim at the moving tumor and a minor beam aperture and weighting adjustment to maintain plan conformality. The dose-volume histogram (DVH) of the 4DCT plan was created with two methods: one is an integrated DVH (iDVH(4D)), which is defined as the temporal average of all 3D-phase-plan DVHs, and the other (DVH(4D)) is based on the dose distribution in a reference phase CT image by dose warping from all phase plans using the displacement vector field (DVF) from a free-form deformable image registration (DIR). The DVH(3.5D) (for the 3.5DCT plan) was compared with both iDVH(4D) and DVH(4D). To quantify the DVH difference between the 3.5DCT plan and the 4DCT plan, two methods were used: relative difference (%) of the areas underneath the DVH curves and the volumes receiving more than 20% (V20) and 50% (V50) of prescribed dose of these 4D plans. The volume of the delineated GTV from different phase CTs varied dramatically from 24% to 112% among the five patients, whereas the GTV from 3.5DCT deviated from the averaged GTV in 4DCT by only -6%±6%. For planning tumor volume (PTV) coverage, the difference between the DVH(3.5D) and iDVH(4D) was negligible (<1% area), whereas the DVH(3.5D) and DVH(4D) were quite different, due to DIR uncertainty (∼2 mm), which propagates to PTV dose coverage with a pronounced uncertainty for small tumors (0.3-4.0 cm(3)) in stereotactic plans with sharp dose falloff around PTV. For OARs, such as the lung, heart, cord and esophagus, the three DVH curves (DVH(3.5D), DVH(4D) and iDVH(4D)) were found to be almost identical for the same patients, especially in high-dose regions. For the tumor-containing lung, the relative difference of the areas underneath the DVH curves was found to be small (5.3% area on average), of which 65% resulted from the low-dose region (D < 20%). The averaged V20 difference between the two 4D plans was 1.2% ± 0.8%. For the mean lung dose (MLD), the 3.5DCT plan differed from the 4DCT plan by -1.1%±1.3%. GTV-motion-compensated CT (3.5DCT) produces an accurate and reliable GTV delineation, which is close to the mean GTV from 4DCT. The 3.5DCT plan is equivalent to the 4DCT plan with <1% dose difference to the PTV and negligible dose difference in OARs. The 3.5DCT approach simplifies 4D planning and provides accurate dose calculation without a substantial increase of clinical workload for motion-tracking delivery to treat small peripheral lung tumors with large motion.

Online monitoring and error detection of real-time tumor displacement prediction accuracy using control limits on respiratory surrogate statistics

PURPOSE

To evaluate Hotelling's T(2) statistic and the input variable squared prediction error (Q((X))) for detecting large respiratory surrogate-based tumor displacement prediction errors without directly measuring the tumor's position.

METHODS

Tumor and external marker positions from a database of 188 Cyberknife Synchrony™ lung, liver, and pancreas treatment fractions were analyzed. The first ten measurements of tumor position in each fraction were used to create fraction-specific models of tumor displacement using external surrogates as input; the models were used to predict tumor position from subsequent external marker measurements. A partial least squares (PLS) model with four scores was developed for each fraction to determine T(2) and Q((X)) confidence limits based on the first ten measurements in a fraction. The T(2) and Q((X)) statistics were then calculated for every set of external marker measurements. Correlations between model error and both T(2) and Q((X)) were determined. Receiver operating characteristic analysis was applied to evaluate sensitivities and specificities of T(2), Q((X)), and T(2)∪Q((X)) for predicting real-time tumor localization errors >3 mm over a range of T(2) and Q((X)) confidence limits.

RESULTS

Sensitivity and specificity of detecting errors >3 mm varied with confidence limit selection. At 95% sensitivity, T(2)∪Q((X)) specificity was 15%, 2% higher than either T(2) or Q((X)) alone. The mean time to alarm for T(2)∪Q((X)) at 95% sensitivity was 5.3 min but varied with a standard deviation of 8.2 min. Results did not differ significantly by tumor site.

CONCLUSIONS

The results of this study establish the feasibility of respiratory surrogate-based online monitoring of real-time respiration-induced tumor motion model accuracy for lung, liver, and pancreas tumors. The T(2) and Q((X)) statistics were able to indicate whether inferential model errors exceeded 3 mm with high sensitivity. Modest improvements in specificity were achieved by combining T(2) and Q((X)) results.

Mitigating errors in external respiratory surrogate-based models of tumor position

PURPOSE

To investigate the effect of tumor site, measurement precision, tumor-surrogate correlation, training data selection, model design, and interpatient and interfraction variations on the accuracy of external marker-based models of tumor position.

METHODS AND MATERIALS

Cyberknife Synchrony system log files comprising synchronously acquired positions of external markers and the tumor from 167 treatment fractions were analyzed. The accuracy of Synchrony, ordinary-least-squares regression, and partial-least-squares regression models for predicting the tumor position from the external markers was evaluated. The quantity and timing of the data used to build the predictive model were varied. The effects of tumor-surrogate correlation and the precision in both the tumor and the external surrogate position measurements were explored by adding noise to the data.

RESULTS

The tumor position prediction errors increased during the duration of a fraction. Increasing the training data quantities did not always lead to more accurate models. Adding uncorrelated noise to the external marker-based inputs degraded the tumor-surrogate correlation models by 16% for partial-least-squares and 57% for ordinary-least-squares. External marker and tumor position measurement errors led to tumor position prediction changes 0.3-3.6 times the magnitude of the measurement errors, varying widely with model algorithm. The tumor position prediction errors were significantly associated with the patient index but not with the fraction index or tumor site. Partial-least-squares was as accurate as Synchrony and more accurate than ordinary-least-squares.

CONCLUSIONS

The accuracy of surrogate-based inferential models of tumor position was affected by all the investigated factors, except for the tumor site and fraction index.

Incidence of changes in respiration-induced tumor motion and its relationship with respiratory surrogates during individual treatment fractions

PURPOSE

To determine how frequently (1) tumor motion and (2) the spatial relationship between tumor and respiratory surrogate markers change during a treatment fraction in lung and pancreas cancer patients.

METHODS AND MATERIALS

A Cyberknife Synchrony system radiographically localized the tumor and simultaneously tracked three respiratory surrogate markers fixed to a form-fitting vest. Data in 55 lung and 29 pancreas fractions were divided into successive 10-min blocks. Mean tumor positions and tumor position distributions were compared across 10-min blocks of data. Treatment margins were calculated from both 10 and 30 min of data. Partial least squares (PLS) regression models of tumor positions as a function of external surrogate marker positions were created from the first 10 min of data in each fraction; the incidence of significant PLS model degradation was used to assess changes in the spatial relationship between tumors and surrogate markers.

RESULTS

The absolute change in mean tumor position from first to third 10-min blocks was >5 mm in 13% and 7% of lung and pancreas cases, respectively. Superior-inferior and medial-lateral differences in mean tumor position were significantly associated with the lobe of lung. In 61% and 54% of lung and pancreas fractions, respectively, margins calculated from 30 min of data were larger than margins calculated from 10 min of data. The change in treatment margin magnitude for superior-inferior motion was >1 mm in 42% of lung and 45% of pancreas fractions. Significantly increasing tumor position prediction model error (mean ± standard deviation rates of change of 1.6 ± 2.5 mm per 10 min) over 30 min indicated tumor-surrogate relationship changes in 63% of fractions.

CONCLUSIONS

Both tumor motion and the relationship between tumor and respiratory surrogate displacements change in most treatment fractions for patient in-room time of 30 min.

Clinical development of a failure detection-based online repositioning strategy for prostate IMRT--experiments, simulation, and dosimetry study

PURPOSE

To implement and evaluate clinic-ready adaptive imaging protocols for online patient repositioning (motion tracking) during prostate IMRT using treatment beam imaging supplemented by minimal, as-needed use of on-board kV.

METHODS

The authors examine the two-step decision-making strategy: (1) Use cine-MV imaging and online-updated characterization of prostate motion to detect target motion that is potentially beyond a predefined threshold and (2) use paired MV-kV 3D localization to determine overthreshold displacement and, if needed, reposition the patient. Two levels of clinical implementation were evaluated: (1) Field-by-field based motion correction for present-day linacs and (2) instantaneous repositioning for new-generation linacs with capabilities of simultaneous MV-kV imaging and remote automatic couch control during treatment delivery. Experiments were performed on a Varian Trilogy linac in clinical mode using a 4D motion phantom programed with prostate motion trajectories taken from patient data. Dosimetric impact was examined using a 2D ion chamber array. Simulations were done for 536 trajectories from 17 patients.

RESULTS

Despite the loss of marker detection efficiency caused by the MLC leaves sometimes obscuring the field at the marker's projected position on the MV imager, the field-by-field correction halved (from 23% to 10%) the mean percentage of time that target displacement exceeded a 3 mm threshold, as compared to no intervention. This was achieved at minimal cost in additional imaging (average of one MV-kV pair per two to three treatment fractions) and with a very small number of repositionings (once every four to five fractions). Also with low kV usage (approximation 2/fraction), the instantaneous repositioning approach reduced overthreshold time by more than 75% (23% to 5%) even with severe MLC blockage as often encountered in current IMRT and could reduce the overthreshold time tenfold (to < 2%) if the MLC blockage problem were relieved. The information acquired for repositioning using combined MV-kV images was found to have submillimeter accuracy.

CONCLUSIONS

This work demonstrated with a current clinical setup that substantial reduction of adverse targeting effects of intrafraction prostate motion can be realized. The proposed adaptive imaging strategy incurs minimal imaging dose to the patient as compared to other stereoscopic imaging techniques.

Electromagnetic real-time tumor position monitoring and dynamic multileaf collimator tracking using a Siemens 160 MLC: geometric and dosimetric accuracy of an integrated system

PURPOSE

Dynamic multileaf collimator tracking represents a promising method for high-precision radiotherapy to moving tumors. In the present study, we report on the integration of electromagnetic real-time tumor position monitoring into a multileaf collimator-based tracking system.

METHODS AND MATERIALS

The integrated system was characterized in terms of its geometric and radiologic accuracy. The former was assessed from portal images acquired during radiation delivery to a phantom in tracking mode. The tracking errors were calculated from the positions of the tracking field and of the phantom as extracted from the portal images. Radiologic accuracy was evaluated from film dosimetry performed for conformal and intensity-modulated radiotherapy applied to different phantoms moving on sinusoidal trajectories. A static radiation delivery to the nonmoving target served as a reference for the delivery to the moving phantom with and without tracking applied.

RESULTS

Submillimeter tracking accuracy was observed for two-dimensional target motion despite the relatively large system latency of 500 ms. Film dosimetry yielded almost complete recovery of a circular dose distribution with tracking in two dimensions applied: 2%/2 mm gamma-failure rates could be reduced from 59.7% to 3.3%. For single-beam intensity-modulated radiotherapy delivery, accuracy was limited by the finite leaf width. A 2%/2 mm gamma-failure rate of 15.6% remained with tracking applied.

CONCLUSION

The integrated system we have presented marks a major step toward the clinical implementation of high-precision dynamic multileaf collimator tracking. However, several challenges such as irregular motion traces or a thorough quality assurance still need to be addressed.

Inferring positions of tumor and nodes in Stage III lung cancer from multiple anatomical surrogates using four-dimensional computed tomography

PURPOSE

To investigate the feasibility of modeling Stage III lung cancer tumor and node positions from anatomical surrogates.

METHODS AND MATERIALS

To localize their centroids, the primary tumor and lymph nodes from 16 Stage III lung cancer patients were contoured in 10 equal-phase planning four-dimensional (4D) computed tomography (CT) image sets. The centroids of anatomical respiratory surrogates (carina, xyphoid, nipples, mid-sternum) in each image set were also localized. The correlations between target and surrogate positions were determined, and ordinary least-squares (OLS) and partial least-squares (PLS) regression models based on a subset of respiratory phases (three to eight randomly selected) were created to predict the target positions in the remaining images. The three-phase image sets that provided the best predictive information were used to create models based on either the carina alone or all surrogates.

RESULTS

The surrogate most correlated with target motion varied widely. Depending on the number of phases used to build the models, mean OLS and PLS errors were 1.0 to 1.4 mm and 0.8 to 1.0 mm, respectively. Models trained on the 0%, 40%, and 80% respiration phases had mean (+/- standard deviation) PLS errors of 0.8 +/- 0.5 mm and 1.1 +/- 1.1 mm for models based on all surrogates and carina alone, respectively. For target coordinates with motion >5 mm, the mean three-phase PLS error based on all surrogates was 1.1 mm.

CONCLUSIONS

Our results establish the feasibility of inferring primary tumor and nodal motion from anatomical surrogates in 4D CT scans of Stage III lung cancer. Using inferential modeling to decrease the processing time of 4D CT scans may facilitate incorporation of patient-specific treatment margins.

Real-time tumor tracking: automatic compensation of target motion using the Siemens 160 MLC

PURPOSE

Advanced high quality radiation therapy techniques such as IMRT require an accurate delivery of precisely modulated radiation fields to the target volume. Interfractional and intrafractional motion of the patient's anatomy, however, may considerably deteriorate the accuracy of the delivered dose to the planned dose distributions. In order to compensate for these potential errors, a dynamic real-time capable MLC control system was designed.

METHODS

The newly developed adaptive MLC control system contains specialized algorithms which are capable of continuous optimization and correction of the aperture of the MLC according to the motion of the target volume during the dose delivery. The algorithms calculate the new leaf positions based on target information provided online to the system. The algorithms were implemented in a dynamic target tracking control system designed for a Siemens 160 MLC. To assess the quality of the new target tracking system in terms of dosimetric accuracy, experiments with various types of motion patterns using different phantom setups were performed. The phantoms were equipped with radiochromic films placed between solid water slabs. Dosimetric results of exemplary deliveries to moving targets with and without dynamic MLC tracking applied were compared in terms of the gamma criterion to the reference dose delivered to a static phantom.

RESULTS

Our measurements indicated that dose errors for clinically relevant two-dimensional target motion can be compensated by the new control system during the dose delivery of open fields. For a clinical IMRT dose distribution, the gamma success rate was increased from 19% to 77% using the new tracking system. Similar improvements were achieved for the delivery of a complete IMRT treatment fraction to a moving lung phantom. However, dosimetric accuracy was limited by the system's latency of 400 ms and the finite leaf width of 5 mm in the isocenter plane.

CONCLUSIONS

Different experimental setups representing different target tracking scenarios proved that the tracking concept, the new algorithms and the dynamic control system make it possible to effectively compensate for dose errors due to target motion in real-time. These early results indicate that the method is suited to increasing the accuracy and the quality of the treatment delivery for the irradiation of moving tumors.

Delivery of four-dimensional radiotherapy with TrackBeam for moving target using an AccuKnife dual-layer MLC: dynamic phantoms study

Respiratory motion has been considered a clinical challenge for lung tumor treatments due to target motion. In this study, we aimed to perform an experimental evaluation based on dynamic phantoms using MLC‐based beam tracking. TrackBeam, a prototype real‐time beam tracking system, has been assembled and evaluated in our clinic. TrackBeam includes an orthogonal dual‐layer micro multileaf collimator (DmMLC), an on‐board mega‐voltage (MV) portal imaging device, and an image processing workstation. With a fiducial marker implanted in a moving target, the onboard imaging device can capture the motion. The TrackBeam workstation processes the online MV fluence and detects and predicts tumor motion. The DmMLC system then dynamically repositions each leaf to form new beam apertures based on the movement of the fiducial marker. In this study, a dynamic phantom was used for the measurements. Three delivery patterns were evaluated for dosimetric verification based on radiographic films: no‐motion lung‐tumor (NMLT), three‐dimensional conformal radiotherapy (3DCRT), and four‐dimensional tracking radiotherapy (4DTRT). The displacement between the DmMLC dynamic beam isocenter and the fiducial marker was in the range of 0.5 mm to 1.5 mm. With radiographic film analysis, the planar dose histogram difference between 3DCRT and NLMT was 48.6% and 38.0% with dose difference tolerances of 10% and 20%, respectively. The planar dose histogram difference between 4DTRT and NLMT was 15.2% and 4.0%, respectively. Based on dose volume histogram analysis, 4DTRT reduces the mean dose for the surrounding tissue from 35.4 Gy to 19.5 Gy, reduces the relative volume of the total lung from 28% to 18% at V20, and reduces the amount of dose from 35.2 Gy to 15.0 Gy at D20. The experimental results show that MLC‐based real‐time beam tracking delivery provides a potential solution to respiratory motion control. Beam tracking delivers a highly conformal dose to a moving target, while sparing surrounding normal tissue.

PACS number: 87.55.de, 87.55.ne, 87.56.nk

Optimization of an adaptive neural network to predict breathing

PURPOSE

To determine the optimal configuration and performance of an adaptive feed forward neural network filter to predict breathing in respiratory motion compensation systems for external beam radiation therapy.

METHOD AND MATERIALS

A two-layer feed forward neural network was trained to predict future breathing amplitudes for 27 recorded breathing histories. The prediction intervals ranged from 100 to 500 ms. The optimal sampling frequency, number of input samples, training rate, and number of training epochs were determined for each breathing history and prediction interval. The overall optimal filter configuration was determined from this parameter survey, and its accuracy for each breathing example was compared to the individually optimal filter setups. Prediction accuracy was also compared to breathing stability as measured by the autocorrelation of the breathing signal.

RESULTS

The survey of filter configurations converged on a standard setup for all examples of breathing. For 24 of the 27 breathing histories the accuracy of the standard filter for a 300 ms prediction interval was within a few percent of the individually optimized filter setups; for the remaining three histories the standard filter was 5%-15% less accurate.

CONCLUSIONS

A standard adaptive neural network filter setup can provide approximately optimal breathing prediction for a wide range of breathing patterns. The filter accuracy has a clear correlation with the stability of breathing.