Controlling motion prediction errors in radiotherapy with relevance vector machines.

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Medicine

Background:

• Robotic radiotherapy requires precise targeting of moving tumors, necessitating compensation for time latencies.
• Relevance vector machines (RVM), a probabilistic regression technique, have shown promise in predicting respiratory motion.
• The inherent advantage of RVM is modeling predictions with associated uncertainty (variance).

Purpose of the Study:

• To investigate the utility of predicted variance from RVM for controlling prediction errors in robotic radiotherapy.
• To develop and evaluate hybrid algorithms combining RVM with other methods for improved respiratory motion compensation.
• To assess the clinical translatability of variance-controlled prediction algorithms.

Main Methods:

• Correlated treatment duty cycle and precision with variance, interrupting treatment if variance exceeded a threshold.
• Developed two hybrid algorithms: multiple RVMs (HYB(RVM)) and a wavelet-based least mean square (wLMS) with RVM (HYB(wLMS-RVM)).
• Analyzed variance patterns across different motion traces to understand prediction error characteristics.

Main Results:

• Limiting variance increased treatment precision while decreasing the duty cycle.
• All developed hybrid algorithms demonstrated improved prediction accuracy over individual components.
• The HYB(RVM) algorithm reduced the mean Root Mean Square Error (RMSE) from 0.18 mm to 0.17 mm across 304 motion traces.

Conclusions:

• Predicted variance is an effective metric for controlling prediction errors, enhancing the robustness of radiotherapy.
• The HYB(RVM) algorithm shows potential for clinical translation due to its simplicity and parallelizability.
• Variance-based error control contributes to more accurate and reliable treatment delivery for moving tumors.