Prediction of the motion of chest internal points using a recurrent neural network trained with real-time recurrent learning for latency compensation in lung cancer radiotherapy

Prediction of the Position of External Markers Using a Recurrent Neural Network Trained With Unbiased Online Recurrent Optimization for Safe Lung Cancer Radiotherapy

“…The prediction error decreased as the number of hidden units increased. That fact has previously been observed in the case of RTRL [32].…”

“…10). Similarly, it has been reported in [32] that increasing the number of hidden units of a vanilla RNN with a single hidden layer trained with RTRL to predict breathing signals led to a decrease of the prediction MAE. Fig.…”

“…2), and to achieve this, various prediction methods have been proposed [41,18,11]. Among those methods, artificial neural networks (ANNs) with different architectures and training algorithms have been studied in the context of radiotherapy [28,36,9,13,26,6,27,19,15,20,4,37,12,43,40,21,46,22,45,17,42,32]. ANNs are efficient for performing prediction with a high response time, which is the time interval in advance for which the prediction is made, also called the look-ahead time or horizon, and for non-stationary and complex signals.…”

“…They are able to efficiently learn features and long-term dependencies from sequential and time-series data [34]. As a result, almost all of the recent research about ANNs applied to time series forecasting for motion management in radiotherapy focuses on RNNs [12,43,46,21,22,45,17,42,32]. RNN models such as long short term memory (LSTM) networks have also been used in related medical data processing problems such as cardiorespiratory motion prediction from X-ray angiography sequences [1] and next-frame prediction in medical image sequences, including chest dynamic imaging [29,33].…”

“…Adaptive or dynamic learning has been applied many times to radiotherapy, and several studies demonstrated the benefit of that approach in comparison with static models [15,40,22]. Real-time recurrent learning (RTRL) [44,8] is one of these dynamic approaches that has already been used for predicting tumor motion from the Cyberknife Synchrony system [22] and the position of internal points in the chest [32].…”