Kernel density estimation-based real-time prediction for respiratory motion.

Area of Science:

• Medical Physics
• Radiotherapy
• Image-guided Therapy

Background:

• Adaptive radiotherapy requires precise real-time target localization.
• System latency in radiotherapy necessitates accurate prediction of target motion.
• Respiration-induced motion in thoracic and abdominal tumors poses significant prediction challenges due to complex, unpredictable patterns.

Purpose of the Study:

• To develop a statistical approach for predicting the future position of mobile tumors, accounting for inherent motion uncertainties.
• To move beyond deterministic prediction models towards a probabilistic framework for target motion.
• To provide a more robust and flexible method for real-time target prediction in adaptive radiotherapy.

Main Methods:

• Proposed a statistical treatment of prediction by estimating the joint probability distribution of historical and future target positions.
• Utilized kernel density estimation (KDE) for efficient estimation of the joint probability density function (pdf).
• Derived estimators based on the estimated conditional distribution, enabling a nonparametric description of prediction uncertainty.

Main Results:

• The kernel density estimation-based prediction method demonstrated universally significant improvement over benchmark methods (most recent sample, adaptive linear filter).
• The proposed probabilistic approach showed particular value for long lookahead times where alternative methods failed.
• Evaluated on ten patient RPM traces using normalized root mean squared difference as the error metric.

Conclusions:

• The statistical, probabilistic approach using kernel density estimation offers a more accurate and robust method for predicting mobile tumor motion in adaptive radiotherapy.
• This method is compatible with inconsistent training data and does not require prior structural assumptions about motion patterns.
• The nonparametric nature of the method provides a full description of prediction uncertainty, enhancing treatment planning and delivery.