An artificial neural network for predicting the incidence of radiation pneumonitis

Abstract

A method to predict radiation-induced pneumonitis (RP) using an artificial neural network (ANN) was investigated. A retrospective study was applied to the clinical data from 142 patients who have been treated with three-dimensional conformal radiotherapy for tumors in the thoracic region. These data were classified, based on their treatment outcome, into two patient clusters: with RP (Np=26) and without RP (Np= 116). An ANN was designed as a classifier. To perform the classification, a patient-treatment outcome with RP was assigned a value of 1, and a patient treatment outcome without RP was assigned a value of -1. The input of the ANN was limited to the patient lung dose-volume data only. A volume vector (VD) that describes patient lung subvolumes receiving more than a set of threshold doses was used as the network input variable. A zero value was used as the threshold to set the output value into -1 or 1. Three ANNs (ANN_1, ANN_2, and ANN_3), each with three layers, were trained to perform this classification function and to show the effect of training data on the ANN performance. Radial basis function was applied as the hidden layer neuron activation function and a sigmoid function was selected as the output layer neuron function. Backpropagation with a conjugate gradient algorithm was used to train the network. ANN_1 was trained and tested by using the leave-one-out method. ANN_2 was trained by randomly selecting 2/3 of the patient data, and tested by the remaining 1/3 of the data. ANN_3 was trained by a user selecting 2/3 of the patient data, and tested by the remaining 1/3 of the data. The predictive accuracy was verified as the area under a receiver operator characteristic (ROC) curve. The correct classification rates of 73% for RP cases, and 99% for non-RP cases were obtained from ANN_1. The corresponding correct classification rates of 44% for RP cases, and 89% for non-RP cases were obtained from ANN_2. From the ANN_3 test phase, the corresponding correct classification rates of 55% for RP cases, and 95% non-RP cases were achieved. The area under ROC curve was 0.85+/-0.05, 0.68+/-0.10, and 0.81+/-0.09 for ANN_1, ANN _2, and ANN_3, respectively, within its asymmetric 95% confidence interval. The sensitivity was 95%, 57%, and 71%, and the specificity was 94%, 88%, and 90% for ANN_1, ANN_2, and ANN_3, respectively. Preliminary results suggest that the ANN approach provides a useful tool for the prediction of radiation-induced lung pneumonitis, using the patient lung dose-volume information.

The relationship between pulmonary function metrics and radiation-induced lung injury

“…The importance of these clinical factors have led to the development of complex neural network modeling algorithms that incorporate dosimetric as well as clinical parameters in order to most accurately predict RILI risk [9][10][11] , but these models seem too complicated and unwieldy to be clinically useful or practicable for most radiation oncologists. A recent study proposes a new metric, TFS5, which combines the volume of lung spared from receiving more than 5 Gy and the transfer coefficient for carbon monoxide (KCO) [14] .…”

“…A number of predictive models have been posited to ascertain which patients are likely to develop RILI given similar radiation doses and treatment parameters [2][3][4][5][6][7][8][9][10][11][12][13][14] .…”

A neural network model to predict lung radiation‐induced pneumonitis

“…Neural networks, unlike simpler models, have the potential to model the synergistic interaction between variables using a flexible nonlinear relationship. 28 Two prior analyses from our group have considered neural networks and concluded that their predictive capabilities are equivalent 29 or better 30 than other commonly used dosimetric models. 31,32 However, these earlier works were limited, since the network was constructed with a fixed number of nodes in the hidden layer, 29 they lacked a selection process for input features, 30 and input features did not include non-dose variables.…”

A Clinical Application of Fuzzy Logic

“…In this chapter fuzzy model structure and different steps of model performance were explained graphically and finally we compared fuzzy model performance with two different correlation models based on Artificial Neural Network and State model [Procházka & Pavelka, 2007;Robert et al, 2002;Ruan et al, 2008;Seppenwoolde et al, 2007;Sharp et al, 2004;Su et al, 2005]. The state model was implemented as a linear/quadratic correlation between external marker motion and internal tumor motion.…”