Optimized order estimation for autoregressive models to predict respiratory motion.

Area of Science:

• Medical Physics
• Robotic Surgery
• Signal Processing

Background:

• Robotic radiosurgery requires precise compensation for respiratory-induced organ motion.
• CyberKnife Synchrony system uses body surface tracking and time series prediction to minimize tracking errors.
• Estimating the optimal model order for autoregressive (AR) prediction algorithms is challenging.

Purpose of the Study:

• To evaluate the effectiveness of Akaike information criterion (AIC), corrected AIC, and Bayesian information criterion (BIC) for estimating AR model order.
• To assess these criteria for least mean square (LMS) and wavelet-based LMS (wLMS) predictors in respiratory motion compensation.

Main Methods:

• The study analyzed the first minute of respiratory signals.
• AIC, corrected AIC, and BIC were evaluated for model order selection.
• LMS and wLMS prediction algorithms were tested using these criteria.

Main Results:

• AIC demonstrated the highest prediction accuracy for both LMS and wLMS predictors across 12 motion traces.
• Using AIC with wLMS on 304 real motion traces reduced prediction error by 85.1% compared to the original implementation.
• AIC-based model order estimation proved more efficient than grid search.

Conclusions:

• AIC is a valid and efficient method for estimating model order in AR-based prediction algorithms for robotic radiosurgery.
• This approach enhances prediction accuracy and reduces computational load, improving tumor ablation during respiration.
• The findings support the use of AIC for optimizing motion compensation in real-time treatment delivery.