Forecasting respiratory motion with accurate online support vector regression (SVRpred).

Area of Science:

• Medical Physics
• Robotics
• Signal Processing

Background:

• Accurate radiation delivery in image-guided robotic radiosurgery necessitates precise prediction algorithms for patient motion.
• Human respiratory motion presents a significant challenge in achieving targeted radiation therapy.

Purpose of the Study:

• To introduce and evaluate a novel prediction method, SVRpred, for human respiratory motion.
• To assess the performance of SVRpred against established prediction algorithms in radiosurgery contexts.

Main Methods:

• Developed SVRpred, a prediction method utilizing support vector regression (SVR).
• Tested SVRpred using simulated respiratory motion data and real-world data from CyberKnife treatments.
• Compared SVRpred performance against MULIN and Wavelet-based multi scale autoregression (wLMS) prediction methods.

Main Results:

• SVRpred demonstrated superior performance over MULIN and wLMS on real and noise-corrupted simulated data.
• The algorithm showed significant improvements of 15-48 percentage points compared to other methods.
• On noise-free simulated data, SVRpred matched MULIN but was outperformed by wLMS.

Conclusions:

• SVRpred is a viable tool for predicting human respiratory motion, outperforming prior algorithms.
• The primary limitation is its computational complexity, leading to slower prediction speeds.
• Real-time application of SVRpred requires high-performance computing for high-resolution signal sampling.