Optimization of an adaptive neural network to predict breathing

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy Technology

Background:

• Respiratory motion significantly impacts the accuracy of external beam radiation therapy.
• Effective compensation requires precise prediction of breathing patterns.
• Adaptive feed forward neural networks offer potential for real-time motion prediction.

Purpose of the Study:

• To determine the optimal configuration and performance of an adaptive feed forward neural network filter.
• To predict breathing patterns for respiratory motion compensation in radiation therapy.
• To evaluate the filter's accuracy across diverse breathing histories and prediction intervals.

Main Methods:

• Trained a two-layer feed forward neural network on 27 breathing histories.
• Optimized sampling frequency, input samples, training rate, and epochs for prediction intervals (100-500 ms).
• Compared standard filter configuration accuracy against individually optimized setups and breathing signal autocorrelation.

Main Results:

• A standard filter configuration was found to be near-optimal for most breathing histories.
• For a 300 ms prediction interval, the standard filter achieved accuracy within a few percent of individually optimized filters for 24 of 27 histories.
• The standard filter was 5%-15% less accurate for the remaining three histories, indicating a dependency on breathing pattern variability.

Conclusions:

• A standardized adaptive neural network filter setup can achieve near-optimal breathing prediction for various patterns.
• The prediction accuracy is strongly correlated with the stability of the patient's breathing.
• This approach facilitates robust respiratory motion compensation in external beam radiation therapy.