Ensembles of Neural Networks for Tumour Motion Prediction

Abstract

Dynamic tumour-tracked radiotherapy is a promising method for delivering conformal doses to tumours that exhibit a large degree of motion as a result of patient respiration. However, there exists an inevitable latency between the acquisition of an image of a moving tumour and the adaptation of the therapeutic beam to match its observed position and contour. This must be addressed by predicting respiration-induced tumour motion so that the requisite mechanical adjustments can be initiated sufficiently in advance. For MR-based tracking, this latency is relatively long compared to other imaging techniques. Accurate motion prediction therefore requires a more sophisticated approach than those used for short-latency hardware. A novel application of long short-term memory recurrent neural networks for respiration-induced tumour motion is presented in this thesis. It consists of three main components: (1) acceleration of training using super-convergence regularization with intelligent early stopping; (2) mitigation of overfitting and instability through homogeneous network ensembles; and (3) improvement of the reaction to changing respiratory patterns during treatment through a novel adaptation method called intermittent retraining. Compared to previous studies, this approach reduces the amount of time required for network training by several orders of magnitude while simultaneously improving the accuracy and consistency of predictions. This work represents a step toward bringing linac-MR based dynamic tumour-tracked radiotherapy into clinical relevance by making it both more practical and more precise.