3:45 PM - 4:00 PM
[2M04] Prediction of the position of external markers on the chest and abdomen using a recurrent neural network trained with real-time recurrent learning for accurate and safe lung cancer radiotherapy
Keywords:Lung cancer radiotherapy, Latency compensation, Recurrent neural network, Real-time recurrent learning, External markers
During lung cancer radiotherapy, targeting the tumor with the treatment beam is difficult due to patient breathing motion. External markers on the chest may be used to infer the tumor position. Their prediction is necessary to compensate for delays inherent to the data acquisition and radiation delivery process, and therefore avoid excessive irradiation of healthy tissues.
In the proposed study, we use observation records of the three-dimensional position of external markers in 9 time series sequences. Prediction is performed using a recurrent neural network (RNN) trained with the real-time recurrent learning (RTRL) algorithm, the least mean-square algorithm, and linear regression for a horizon value of 2.0s.
When performing prediction with the RNN, we found a mean absolute error of 0.9mm, a root-mean-square error of 1.6mm, and a maximum error of 12.7mm of the test set. These errors are lower than with the other methods considered, except the LMS method which has a corresponding maximum error equal to 10.2mm.
This research shows that RNNs trained with RTRL are efficient for predicting respiratory motion and is thus a step towards safer radiotherapy.
In the proposed study, we use observation records of the three-dimensional position of external markers in 9 time series sequences. Prediction is performed using a recurrent neural network (RNN) trained with the real-time recurrent learning (RTRL) algorithm, the least mean-square algorithm, and linear regression for a horizon value of 2.0s.
When performing prediction with the RNN, we found a mean absolute error of 0.9mm, a root-mean-square error of 1.6mm, and a maximum error of 12.7mm of the test set. These errors are lower than with the other methods considered, except the LMS method which has a corresponding maximum error equal to 10.2mm.
This research shows that RNNs trained with RTRL are efficient for predicting respiratory motion and is thus a step towards safer radiotherapy.