Premium
WE‐D‐M100J‐03: A Neural Network Model to Predict Lung Radiation‐Induced Pneumonitis
Author(s) -
Chen S,
Zhou S,
Zhang J,
Marks L,
Das S
Publication year - 2007
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.2761564
Subject(s) - receiver operating characteristic , lung volumes , dlco , radiation therapy , artificial neural network , nuclear medicine , lung cancer , medicine , mathematics , dose volume histogram , lung , statistics , computer science , radiation treatment planning , diffusing capacity , radiology , artificial intelligence , lung function
Purpose: To build and test a feed‐forward neural network model to predict the occurrence of lung radiation‐induced Grade 2+ pneumonitis. Method and Materials: The database comprised 235 patients with lung cancer treated using radiotherapy (34 diagnosed with pneumonitis). The neural network was constructed using a unique algorithm that alternately grew and pruned it, starting from the smallest possible network, until a satisfactory solution was found. The weights and biases of the network were computed using the error back‐propagation approach. The network was tested using ten‐fold cross‐validation, wherein 1/10 th of the data were tested, in turn, using the model built with the remaining 9/10 th of the data. Results: The network was constructed with input features selected from dose and non‐dose variables. The selected input features were: lung volume receiving > 16 Gy (V 16 ), mean lung dose, generalized equivalent uniform dose (gEUD) for the exponent a=3.5, free expiratory volume in 1s (FEV 1 ), diffusion capacity of Carbon Monoxide (DLCO%), and whether or not the patient underwent chemotherapy prior to radiotherapy. With the exception of FEV1, all input features were found to be individually significant (p < 0.05). The area under the Receiver Operating Characteristics (ROC) curve for cross‐validated testing was 0.76 (sensitivity: 68%, specificity: 69%). To gauge the impact of non‐dose variables on model predictive capability, a second network was constructed with input features selected only from lung dose‐volume histogram variables. The area under the ROC curve for cross‐validation was 0.67 (sensitivity: 53%, specificity: 69%). The network constructed from dose and non‐dose variables was statistically superior (p=0.020), indicating that the addition of non‐dose features significantly improves the generalization capability of the network. Conclusions: The neural network constructed from dose and non‐dose variables can be used to prospectively predict radiotherapy‐induced pneumonitis and, thereby, appropriately alter radiotherapy plans to reduce this possibility.