Two stages CNN-Based segmentation of gliomas, uncertainty quantification and prediction of overall patient survival

Abstract

This paper proposes, in the context of brain tumor study, a fast automatic method that segments tumors and predicts patient overall survival. The segmentation stage is implemented using two fully convolutional networks based on VGG-16, pre-trained on ImageNet for natural image classification, and fine tuned with the training dataset of the MICCAI 2019 BraTS Challenge. The first network yields to a binary segmentation (background vs lesion) and the second one focuses on the enhancing and non-enhancing tumor classes. The final multiclass segmentation is a fusion of the results of these two networks. The prediction stage is implemented using kernel principal component analysis and random forest classifiers. It only requires a predicted segmentation of the tumor and a homemade atlas. Its simplicity allows to train it with very few examples and it can be used after any segmentation process.