Ensemble of CNNs for Segmentation of Glioma Sub-regions with Survival Prediction

Abstract

Gliomas are the most common malignant brain tumors, having varying level of aggressiveness, with Magnetic Resonance Imaging (MRI) being used for their diagnosis. As these tumors are highly heterogeneous in shape and appearance, their segmentation becomes a challenging task. In this paper we propose an ensemble of three Convolutional Neural Network (CNN) architectures viz. (i) P-Net, (ii) U-Net with spatial pooling, and (iii) ResInc-Net for glioma sub-regions segmentation. The segmented tumor Volume of Interest (VOI) is further used for extracting spatial habitat features for the prediction of Overall Survival (OS) of patients. A new aggregated loss function is used to help in effectively handling the data imbalance problem. The concept of modeling predictive distributions, test time augmentation and ensembling methods are used to reduce uncertainty and increase the confidence of the model prediction. The proposed integrated system (for Segmentation and OS prediction) is trained and validated on the Brain Tumor Segmentation (BraTS) Challenge 2019 dataset. We ranked among the top performing methods on Segmentation and Overall Survival prediction on the validation dataset, as observed from the leaderboard. We also ranked among the top four in the Uncertainty Quantification task on the testing dataset.