3-D RPET-NET: Development of a 3-D PET Imaging Convolutional Neural Network for Radiomics Analysis and Outcome Prediction

Abstract

Radiomics is now widely used to improve the prediction of treatment response and patient prognosis in oncology. In this paper, we propose an end-to-end prediction model based on a 3-D convolutional neural network (CNN), called 3-D RPET-NET, that extracts 3-D image features through four layers. Our model was evaluated for its ability to predict the response to radio-chemotherapy in 97 patients with esophageal cancer from positron emission tomography (PET) images. The accuracy of the model was compared to five other methods proposed in the literature for PET images, based on 2-D CNN and random forest algorithms. The role of the volume of interest on the accuracy of 3-D RPET-NET was also evaluated using isotropic margins of 1-4 cm around the tumor volume. After segmentation of the lesion using a fixed threshold value of 40% of the maximum standard uptake value, the best accuracy of 3-D RPET-NET reached 72% and outperformed the other methods tested. We also showed that using an isotropic margin of 2 cm around the tumor volume improved the performances of 3-D RPET-NET to reach an accuracy of 75%.