A comparison of neural network approaches for on-line prediction in IGRT

John H. Goodband, Olivier C.L. Haas, J.A. Mills

    Research output: Contribution to journalArticle

    32 Citations (Scopus)
    28 Downloads (Pure)

    Abstract

    Image-guided radiation therapy aims to improve the accuracy of treatment delivery by tracking tumor position and compensating for observed movement. Due to system latency it is sometimes necessary to predict tumor trajectory evolution in order to facilitate changes in beam delivery. Neural networks (NNs) have previously been investigated for predicting future tumor position because of their ability to model non-linear systems. However, no attempt has been made to optimize the NN training algorithms, and no mention has been made of potential errors which can be caused by using NNs for extrapolation purposes. In this work, after giving a brief explanation of NN theory, a comparison is made between 4 different adaptive algorithms for training time-series prediction NNs. New error criteria are introduced which highlight error maxima. Results are obtained by training the NNs using previously published data. A hybrid algorithm combining Bayesian regularization with conjugate-gradient backpropagation is demonstrated to give the best average prediction accuracy, whilst a generalized regression NN is shown to reduce the possibility of isolated large prediction errors.
    Original languageEnglish
    Pages (from-to)1113
    JournalMedical Physics
    Volume35
    Issue number3
    DOIs
    Publication statusPublished - 2008

    Keywords

    • backpropagation
    • biomedical imaging
    • conjugate gradient methods
    • dosimetry
    • Gaussian processes
    • generalisation (artificial intelligence)
    • mean square error methods
    • medical computing
    • multilayer perceptrons
    • radial basis function networks
    • radiation therapy
    • regression analysis
    • tumours

    Fingerprint Dive into the research topics of 'A comparison of neural network approaches for on-line prediction in IGRT'. Together they form a unique fingerprint.

  • Cite this