Deep Feature Learning with Discrimination Mechanism for Brain Tumor Segmentation and Diagnosis

Abstract

Brain tumor segmentation is one of the main challenging problems in computer vision and its early diagnosis is critical to clinics. Segmentation needs to be accurate, efficient and robust to avoid influences caused by various large and complex biases added to images. This paper proposes a multiple convolutional neural network (CNNs) framework with discrimination mechanism which is effective to achieve these goals. First of all, this paper proposes to construct different triplanar 2D CNNs architecture for 3D voxel classification, greatly reducing segmentation time. Experiment is conducted on images provided by Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) organized by MICCAI 2013 for both training and testing. As T1, T1-enhanced, T2 and FLAIR MRI images are utilized, multimodal features are combined. As a result, accuracy, sensitivity and specificity are comparable in comparison with manual gold standard images and better than state-of-the-art segmentation methods.