Mitigating errors in external respiratory surrogate-based models of tumor position.

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Accurate tumor position monitoring is vital for effective radiotherapy.
• External marker-based models offer a non-invasive approach to infer tumor location.
• Variability in tumor and marker positions presents challenges for predictive modeling.

Purpose of the Study:

• To assess how various factors influence the accuracy of external marker-based tumor position models.
• Investigate the impact of measurement precision, data selection, and model design on predictive accuracy.
• Determine the role of interpatient and interfraction variations in tumor tracking errors.

Main Methods:

• Analysis of Cyberknife Synchrony system log files from 167 treatment fractions.
• Evaluation of Synchrony, ordinary-least-squares (OLS), and partial-least-squares (PLS) regression models.
• Systematic variation of training data quantity/timing and introduction of noise to assess data precision effects.

Main Results:

• Tumor position prediction errors increased throughout treatment fractions.
• Increased training data did not consistently improve model accuracy.
• Noise in external marker data significantly degraded model performance (16% PLS, 57% OLS).
• Prediction errors were linked to patient variability but not tumor site or fraction number.

Conclusions:

• All investigated factors, except tumor site and fraction index, impacted surrogate-based tumor position model accuracy.
• Partial-least-squares models demonstrated accuracy comparable to Synchrony and superior to OLS.
• Optimizing model design and ensuring data precision are key for accurate tumor tracking.