Application and Parametric Studies of a Sliding Window Neural Network for Respiratory Motion Predictions of Lung Cancer Patients

Abstract

In real-time adaptive image-guided radiotherapy (IGRT), the beam delivery position is changed to follow the tumor motion. Most systems cannot respond instantaneously, and compensation for system lag is required. Typically, future tumor positions are predicted based on the respiratory motion tracked from an external surrogate. In the current work, a sliding window of time series data taken from the respiratory cycle is input into a neural network to predict a future position. The finite past history of the respiratory position is used to train the model. A nonlinear autoregressive neural network with exogenous inputs was used to simultaneously predict future positions. Patient data from the Respiratory Trace Generator (RTG) [1] was used for the training, validation and testing of the model. Parametric studies involving the number of input nodes (length of sliding window), number of hidden nodes and prediction horizon were performed. Tradeoffs between under-learning, training rate and over-learning were identified. While training error decreases as the number of hidden nodes increases, the validation error increases beyond 20 nodes. Large errors occur during transitions between inhale and exhale as well as when the prediction horizon increases.