Ensemble of Deep Convolutional Neural Networks for Prognosis of Ischemic Stroke

Abstract

We propose an ensemble of deep neural networks for the two tasks of automated prognosis of post-treatment ischemic stroke, as imposed by the ISLES 2016 Challenge. For lesion outcome prediction, we employ an ensemble of three-dimensional multiscale residual U-Net and a fully convolutional network, trained using image patches. In order to handle class imbalance, we devise a multi-step training strategy. For clinical outcome prediction, we combine a convolutional neural network (CNN) and a logistic regression model. To overcome the small sample size and the need for whole brain image, we use the CNN trained using patches as a feature extractor and trained a shallow network based on the extracted features. Our ensemble approach demonstrated an appealing performance on both problems, and is ranked among the top entries in the Challenge.