Respiration motion management in radiation therapy typically involves the correlation between the position of the target and a measured respiration signal. In respiration synchronization techniques, the spatial position of the treatment beam is moved relative to the target. One of the limiting factors in dynamically tracking the target is the time-delay between the point in time at which the position of the tumor is measured and the point at which the treatment beam is moved. The purpose of this study was to develop a novel technique for dynamically predicting respiration motion and uncertainty up to 1.5 seconds in the future in real-time using Time-Delay Kernel Regression (TDKR) modeling.
Materials/Methods
A CCD-based respiratory gating system was used to measure 110 patient respiration cycles. These measured respiration cycles were used to construct a memory matrix for predicting the position of the target 0.5, 1.0, and 1.5 seconds into the future in real-time using Time-Delay Kernel Regression. In addition, new respiration data is continuously added to the model's memory matrix as it is acquired. The most recent 300 observations were always kept in the training set, and additional training data from the database was chosen by using an Adeli-Hung clustering algorithm. Using this clustering technique allowed the model to make accurate predictions, while limiting the number of vectors in the training set so that it still had near instantaneous run-time.
Results
The root mean squared error (RMSE) was computed 0.5, 1.0, and 1.5 seconds into the future for the measured respiration cycles. For predictions 1.5 seconds into the future, the mean RMSE was 1.4%. For predictions 1.0 second into the future, the RMSE was 1.2%, and for 0.5 seconds into the future the RMSE was only 0.7%. The mean uncertainty for the predictions at 0.5, 1.0, and 1.5 seconds into the future was 2.4%, 3.2%, and 3.4%, respectively. Figure 1 shows predictions and uncertainty for a patient's breathing pattern that becomes irregular in both phase and amplitude during the treatment session.
Conclusions
This study demonstrates that TDKR modeling can be used to learn the complex relationships present in respiration data. The results of this study indicate that TDKR modeling has similar predictive performance as previously studied parametric models for prediction up to 1.5 seconds into the future. The advantage of this technique over existing techniques comes from the TDKR ability to continuously learn new respiration cycles for each patient without the need for computationally and time intensive re-training. These advantages make the TDKR technique well suited for respiration synchronization techniques.
FULL TEXT
