Prediction of lung tumor motion extent through artificial neural network (ANN) using tumor size and location data

Abstract

The aim of this study is to assess the possibility of developing novel predictive models based on data mining algorithms which would provide an automatic tool for the calculation of the extent of lung tumor motion characterized by its known location and size. Data mining is an analytic process designed to explore data in search of regular patterns and relationships between variables. The ultimate goal of data mining is prediction of the future behavior. Artificial neural network (ANN) data-mining algorithm was used to develop an automatic model, which was trained to predict extent of the tumor motion using the data set obtained from the available 4D CT imaging data. The accuracy of the designed neural network was tested by using longer training time, different input values and/or more neurons in its hidden layer. An optimized ANN best fit the training and test datasets with a regression value (R) of 0.97 and mean squared error value of 0.0039 cm 2. The maximum error that was recorded for the best network performance was 0.32 cm in the craniocaudal direction. The overall prediction error was largest in this direction for 70% of the studied cases. In this study, the concepts of neural networks were discussed and an ANN algorithm is proposed to be used with clinical lung tumor information for the prediction of the tumor motion extent. The results of optimized ANN are promising and can be a reliable tool for motion pattern calculation. It is an automated tool, which may assist radiation oncologists in defining the tumor margins needed in lung cancer radiation therapy.

Feasibility of predicting tumor motion using online data acquired during treatment and a generalized neural network optimized with offline patient tumor trajectories

“…In the present work 27 trajectories, each with approximately 60 s of data sampled at an interval of 0.133 s, provided a total of 12,180 data samples for this study. To further validate our approach, additional data with various motion patterns were required to yield a more robust generalized neural network . The sensitivities of the parameters with respect to the calculated errors requires further study.…”

“…The design of a neural network has often been considered an art rather than an exact science because of the numerous design considerations, the trial-and-error approach to determining the optimal parameters, and a lack of universal design guidelines. [36][37][38][39] To minimize the problems associated with designing an optimal neural network model, a model averaging approach which combines a diverse range of predictors has been used to realize a model that is less susceptible to bias and significant errors. [31][32][33][34] By combining predictors, the prediction performance may be better than using an individually optimized model.…”

“…To further validate our approach, additional data with various motion patterns were required to yield a more robust generalized neural network. 39 The sensitivities of the parameters with respect to the calculated errors requires further study. With irregular tumor trajectories, an ensemble of neural networks can be used to predict the mean and uncertainty of the predicted tumor positions.…”

“…The number of hidden neurons is a trade‐off between, the inability to learn with few neurons vs. poor generalization capability with an excessively large number of neurons. The design of a neural network has often been considered an art rather than an exact science because of the numerous design considerations, the trial‐and‐error approach to determining the optimal parameters, and a lack of universal design guidelines …”

Machine learning applications in radiation oncology

“…A large body of work has shown neural network, SVM, manifold learning and kernel density estimation can efficiently predict respiratory motion based on previously measured motion traces [153][154][155][156][157][158][159]. Other applications of machine learning involve prediction of motion extent based on tumour size and location in the lungs, automatic diaphragm motion trajectory assessment and incorporation of lung tumour motion into patient setup and prediction of tumour baseline shifts in the short term (approximately 5 s) [160][161][162][163].…”