Real-time tumor motion estimation using respiratory surrogate via memory-based learning.

Area of Science:

• Medical Physics
• Machine Learning
• Radiation Oncology

Background:

• Respiratory tumor motion poses a significant challenge in radiation therapy for thoracic and abdominal cancers.
• Accurate real-time tumor motion tracking is crucial for effective motion management.
• External respiratory monitoring offers a practical, non-ionizing approach to capture respiratory signals.

Purpose of the Study:

• To develop and evaluate a memory-based learning method for accurately inferring internal tumor motion from external respiratory surrogates.
• To address the complex and nonlinear relationships between tumor and surrogate motion.
• To provide a robust solution for real-time tumor motion management in radiation therapy.

Main Methods:

• A memory-based learning approach was employed to model the relationship between 3D tumor motion and 1D abdominal surface surrogates.
• The method stores training data and assigns relevance weights to nearby data points for accurate prediction.
• Local model fitting was used to capture nonlinear dynamics, enhancing robustness to outliers.

Main Results:

• The memory-based learning method achieved an average 3D error of 1.5 mm with only 5 seconds of pretreatment training data.
• The 95th percentile error was 3.4 mm, further reduced to 1.4 mm with optimal tuning.
• The method demonstrated robustness, achieving a ~50% error reduction compared to linear models in the presence of outliers.

Conclusions:

• Memory-based learning accurately captures complex, nonlinear tumor motion dynamics from respiratory surrogates.
• The method is efficient, robust to outliers, and suitable for real-time adaptation in radiation therapy.
• This technique offers a promising solution for accurate and reliable tumor gating and tracking.