Abstract
Accurate prediction of respiratory motion is critical for motion-adaptive radiotherapy to compensate for system latency. This paper introduces OPPRED (Orthogonal Polynomials Prediction), a framework that utilizes orthogonal polynomials for signal smoothing and data compression, combined with a Variant-Best-Neighbor (VBN) search strategy. By prioritizing recent motion through a "right-alignment" approach and leveraging a patient's breathing history, the proposed method achieves high stability and accuracy across long-range prediction horizons, significantly outperforming traditional autoregressive and regression-based models.
Introduction
The primary goal of motion-adaptive radiotherapy is to deliver precise radiation doses to tumors in the lungs or abdomen while minimizing damage to surrounding healthy tissue. However, constant movement from breathing complicates this. While real-time tracking systems are available, they are hindered by system latency—delays between 115 and 420 milliseconds between position capture and beam delivery. Without accurate prediction of the tumor's future position, radiation may miss the target. Breathing patterns are inherently complex, noisy, and patient-specific, rendering generic prediction models ineffective.
Methodology: The OPPRED Framework
The researchers developed the OPPRED framework, which consists of four integrated components designed to provide personalized, real-time motion prediction.
- Personalized Monitoring Window: The system avoids fixed timeframes, instead calculating a window size based on the individual patient's median breathing cycle.
- Orthogonal Polynomial Approximation: Legendre polynomials are used to represent breathing segments. This approach smooths sensor noise and compresses data into a few coefficients, enabling rapid searching of the patient's breathing "library."
- Variant-Best-Neighbor (VBN) Search: The algorithm scans the patient's history for segments similar to the current one. It utilizes a "Right-Alignment" strategy, ensuring that historical patterns match the most recent data points most closely, as these are the strongest predictors of immediate future motion.
- Statistical Decision Maker: After identifying the "best neighbors," the system uses Median Absolute Deviation (MAD) to filter outliers. The final prediction is derived by averaging the future values of the remaining neighbors.
Experimental Evaluation
The study utilized respiratory motion traces from 27 lung and liver cancer patients, collected via the Varian Real-time Position Management (RPM) system at a 30 Hz sampling rate. Records spanned 15 to 45 minutes per patient.
OPPRED was compared against industry standards such as wavelet-based multiscale autoregression (wLMS) and Support Vector Regression (SVRpred). Key findings include:
- Superior Accuracy: OPPRED maintained an $R^2$ value above 0.82 even at a 15-step (0.5 second) prediction horizon.
- Stability: Unlike methods like wLMS, which see performance plummet as look-ahead time increases, OPPRED remained stable across longer horizons. At the 15-step mark, OPPRED achieved an average $R^2$ of 0.824, nearly doubling the performance of the next best traditional method, TVSAR (0.462).
Discussion and Conclusion
The success of OPPRED lies in its treatment of a patient's breathing history as a predictive library and its emphasis on the "right end" of data segments. By reconstructing signals using lower-order polynomials, the system effectively removes high-frequency noise without overfitting.
While highly effective, the system requires an initial monitoring period to build a robust pattern library. Future work may involve optimizing computational overhead for real-time library searches and incorporating additional physiological signals to further enhance prediction accuracy during sudden changes in breathing dynamics.
Summary updated