SU-D-BRA-01: Accurate Real-Time Tumor Motion Estimation from Respiratory Surrogates via Memory-Based Learning

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Respiratory tumor motion poses a significant challenge in radiation therapy, impacting treatment accuracy.
• Real-time tumor motion knowledge is essential for effective beam gating and tracking.
• Invasive fluoroscopic tracking with fiducial markers increases patient imaging dose.

Purpose of the Study:

• To develop and evaluate a noninvasive method for estimating real-time tumor motion using respiratory surrogates.
• To investigate the efficacy of a memory-based learning approach for capturing complex tumor motion dynamics.
• To reduce reliance on invasive tracking methods in radiation therapy.

Main Methods:

• Utilized a memory-based learning approach with locally weighted functions to model tumor motion from respiratory surrogates.
• Trained the model using short pretreatment data (5 seconds) and evaluated on extended unseen data (120 seconds).
• Assessed the method's robustness to outliers and efficiency in adapting to new data.

Main Results:

• Achieved an average 3D error of 0.37 mm and a 95th percentile error of 0.86 mm on unseen test data.
• Errors were significantly below the average tumor motion amplitude (-10 mm).
• Demonstrated a 21% average reduction in prediction error compared to the best linear model.

Conclusions:

• The memory-based learning technique accurately captures nonlinear tumor and surrogate motion relationships.
• The method is efficient (∼1 ms per prediction), making it suitable for real-time applications.
• This approach is an ideal candidate for accurate and robust tumor gating and tracking using respiratory surrogates.