Automatic brain image analysis based on multimodal deep learning scheme

Abstract

In this paper, we propose a new approach for brain image segmentation based on 3D U-Net deep learning architecture. The proposed approach takes into consideration both the neural network's optimizer as well the biological context of the segmentation tissue, by modeling the structured nature of glioma and edematous tissue around the enhancing and non-enhancing tumor within the U-net model. By training multiple deep neural networks based on 3D U-Nets, with a two-stage design, with whole tumor segmentation as the first stage, followed by segmentation of enhancing and non-enhancing tumors in the second stage, along with data augmentation, it was possible to build sparse deep learning model with few images, and achieve better tumor detection performance as compared to other deep learning models reported for BraTS 2018 challenge task, involving the usage of large dataset for building the models.