Brain Tumor Segmentation for Missing Modalities by Supplementing Missing Features

Abstract

Brain tumor segmentation in multi-modal magnetic resonance images is an essential step in brain cancer diagnosis and treatment. Despite the recent success of multi-Modalities fusion network for brain tumor segmentation, we usually confront the situation that some acquired modalities are not available beforehand during clinical practices. In this paper, we propose an advanced fusing network which robust to the absence of any modality in brain tumor segmentation. The network we proposed consists of two modules, the first named Cascade Supplement Module (CSM) uses an advanced cascade operation to generate shared features for missing modalities and the second named Modality Fusion Module (MFM) utilizes squeeze and excitation to fuse the generated features and real features. We evaluate this network on BraTS2018 using subsets of the imaging modalities as input. The experimental results show that our method could achieve better segmentation accuracy than HeMIS, TS and Fusion methods.