Feasibility Study of the Use of Artificial Neural Networks in Predicting Acute Rectal and Urinary Bladder Toxicity Following Prostate Cancer Radiotherapy

Abstract

The role of radical dose radiation therapy (RT) in the management of non-metastatic prostate cancer is established. To prevent secondary effects at organ at risk (OAR) is a primary condition. The dose-volume relationship has been studied in many Institutions and the application of dose volume histogram (DVH) constraints is currently the most reliable way to prevent rectal and bladder toxicity. In this work we studied the ability of artificial neural networks (ANN) to predict acute rectal and urinary bladder toxicity induced by radiation. Data were obtained from 159 patients, who were treated at Istituto Europeo di Oncologia (IEO, Milan IT) between 2006 and 2008 with radical radiotherapy using a 3D dynamic arc conformal technique, supported by image guidance devices. We designed the network architecture as a feed-forward multilayer perceptron net, with a single hidden layer. The aim of this preliminary study was to minimize false negative predictions. The trained ANNs were applied on a dataset of 7 testing cases. In both acute rectal and bladder toxicity, the prediction score was 100% for positive values (medium/severe toxicity) , 60% for negative values (no or light toxicity), giving raise to an overall performance of the network of 71.43% of the testing cases.