Data Processing using Artificial Neural Networks to Improve the Simulation of Lung Motion

To optimize the delivery in lung radiation therapy, a better understanding of the tumor motion is required. On the one hand to have a better tumor-targeting efficiency, and on the other hand to avoid as much as possible normal tissues. The 4D-CT allows to quantify tumor motion, but due to artifacts it introduces biases and errors in tumor localization. Despite of this disadvantage, we propose a method to simulate lung motion based on data provided by the 4D-CT for several patients. To reduce uncertainties introduced by the 4D-CT scan, we conveniently treated data using artificial neural networks. More precisely, our approach consists in a data augmentation technique. The data resulting from this processing step are then used to build a training set for another artificial neural network that learns the lung motion. To improve the learning accuracy, we have studied the number of phases required to precisely describe the displacement of each point. Thus, from 1118 points scattered across 5 patients and defined over 8 or 10 phases, we obtained 5800 points of 50 phases. After training, the network is used to compute the positions of 40 points from five other patients on 10 phases. These points allow to quantify the prediction performance. In comparison with the original data, the ones issued from our treatment process provide a significant increase of the prediction accuracy: an average improvement of 16% can be observed. The motion computed for several points by the neural network that has learnt the lung one exhibits an hysteresis near the one given by the 4D-CT, with an error smaller than 1 mm in the cranio-caudal axis.