Prediction filter for real time tumor tracking using the treatment table

Abstract

Purpose A possible method to treat cancer is the irradiation of the tumor with ionizing radiation, which should be deposited as much as possible in the tumor. In cases of moving tumors their movements have to be mitigated. The method considered in this thesis measures the respiratory motion and calculates the corresponding tumor motion and moves the patient in such a way that the tumor does not move relative to the beam. Some sensors measuring the respiratory motion have large delay times, which can be compensated by prediction filters. The delay time is the time between the measurement and the actual availability of the signal for further use. The accuracy of several prediction filters has been evaluated in this work. The actuator (Treatment Table) also causes errors due to its limited dynamics, therefore, it was modeled and simulated together with the controllers and prediction filters Methods The actuator system, the Protura Treatment Couch (CIVCO Medical Solutions, Kalona, IA, USA) was modeled using methods from Mechanics and was implemented using MATLAB/Simulink (The MathWorks, Inc., Natick, MA, USA). Using measurements of the real system unknown parameters were identified. The measurements were carried out using laser triangulation systems (Micro Epsilon Messtechnik GmbH & Co. KG, Ortenburg, Germany). The six different prediction filters considered here were implemented in MATLAB and their parameters were optimized for two different sampling times and one prediction time. The performance of the prediction filters for a single respiration curve is determined by the Root Mean Square (RMS) of the differences (errors) between the estimated value and the actual value at the corresponding time instants. High performance is achieved by having small RMS errors. During the optimization for each parameter combination of the prediction filter the RMS error for 19 different respiration curves were determined and the average over the respiration curves was considered as the performance index. With the optimal parameter combination the behaviour of the RMS of the prediction filters depending on different prediction times were examined. Finally one prediction filter was built into the model of the system and the complete system was simulated. Results The model of the system has an average RMS error of 0.57 mm in longitudinal (x) direction of the actuator system and 0.5 mm in vertical (z) direction. The lateral (y) direction was not measured. The RMS errors of the prediction filters lie, with exception of the normalized Least Mean Squares (nLMS) filter, between 0.4 mm and 0.65 mm for a prediction time of 300 ms. The standard deviations of the RMS error of a prediction filter are larger than the differences between the prediction filters. Except for the Support Vector Regression (SVR) prediction filter the computing times per step are less than 1 ms, while the SVR filter is at nearly 20 ms (MATLAB implementations). The average increase of the error RMS during tumor tracking caused by a delay of 100 ms is 0.54 mm while the reduction using a prediction filter compensating the delay time is 0.53 mm.