A System for Brain Image Segmentation and Classification Based on Three-Dimensional Convolutional Neural Network

Abstract

We consider the problem of fully automatic brain tumor segmentation in MR images containing glioblastomas. We propose a three-Dimensional Convolutional Neural Network (3D-CNN) approach that achieves high performance while being extremely efficient, a balance that existing methods have struggled to achieve. Our 3D-Brain CNN is formed directly on raw image modalities and thus learn a characteristic representation directly from the data. We propose a new cascading architecture with two pathways that each model normal details in tumors. Fully exploiting the convolutional nature of our model also allows us to segment a complete cerebral image in one minute. In experiments on the 2013 and 2015 BRATS challenge dataset; we exhibit that our approach is among the most powerful methods in the literature, while also being very effective.