Evolutionary training of artificial neural networks for radiotherapy treatment of cancers

Radiotherapy treatment of cancers involves the need to orient a set of X-ray beams directed at a tumour in such a way as to maximise the dosage of X-rays to the tumour itself, while minimising the effect on surrounding healthy tissue and organs. This process is either planned by a radiotherapist, which is a time-consuming process often with suboptimal results, or rarely, where expensive computing equipment is available, by direct optimisation using sophisticated search techniques, which is similarly time-consuming. This paper explores an alternative, whereby a neural network is trained on good medical advice on beam orientations and other treatment factors for a set of cases. The resulting neural network provides the ability to instantly generate good treatment plans based on CT scan data. In particular, when generated with an evolutionary algorithm the neural network's generalisation capability is significantly improved. The system has been tested in a hospital and found very acceptable, occasionally producing better treatment plans than trained radiotherapists.