Sci-Thur AM: YIS – 05: Prediction of lung tumor motion using a generalized neural network optimized from the average prediction outcome of a group of patients

Authors

  • Teo Troy,

    1. University of Manitoba/ CancerCare Manitoba, University of Manitoba, University of Manitoba, University of Manitoba / CancerCare Manitoba
    Search for more papers by this author
  • Alayoubi Nadia,

    1. University of Manitoba/ CancerCare Manitoba, University of Manitoba, University of Manitoba, University of Manitoba / CancerCare Manitoba
    Search for more papers by this author
  • Bruce Neil,

    1. University of Manitoba/ CancerCare Manitoba, University of Manitoba, University of Manitoba, University of Manitoba / CancerCare Manitoba
    Search for more papers by this author
  • Pistorius Stephen

    1. University of Manitoba/ CancerCare Manitoba, University of Manitoba, University of Manitoba, University of Manitoba / CancerCare Manitoba
    Search for more papers by this author

Abstract

Purpose:

In image-guided adaptive radiotherapy systems, prediction of tumor motion is required to compensate for system latencies. However, due to the non-stationary nature of respiration, it is a challenge to predict the associated tumor motions. In this work, a systematic design of the neural network (NN) using a mixture of online data acquired during the initial period of the tumor trajectory, coupled with a generalized model optimized using a group of patient data (obtained offline) is presented.

Methods:

The average error surface obtained from seven patients was used to determine the input data size and number of hidden neurons for the generalized NN. To reduce training time, instead of using random weights to initialize learning (method 1), weights inherited from previous training batches (method 2) were used to predict tumor position for each sliding window.

Results:

The generalized network was established with 35 input data (∼4.66s) and 20 hidden nodes. For a prediction horizon of 650 ms, mean absolute errors of 0.73 mm and 0.59 mm were obtained for method 1 and 2 respectively. An average initial learning period of 8.82 s is obtained.

Conclusions:

A network with a relatively short initial learning time was achieved. Its accuracy is comparable to previous studies. This network could be used as a plug-and play predictor in which (a) tumor positions can be predicted as soon as treatment begins and (b) the need for pretreatment data and optimization for individual patients can be avoided.

Ancillary