Automatic Brain Tumor Segmentation by Exploring the Multi-modality Complementary Information and Cascaded 3D Lightweight CNNs

Accurate segmentation of brain tumors is critical for clinical quantitative analysis and decision making for glioblastoma patients. Convolutional neural networks (CNNs) have been widely used for this task. Most of the existing methods integrate the multi-modality information by merging them as multiple channels at the input of the network. However, explicitly exploring the complementary information among different modalities has not been well studied. In fact, radiologists rely heavily on the multi-modality complementary information to manually segment each brain tumor substructure. In this paper, such a mechanism is developed by training the CNNs like the annotation process by radiologists. Besides, a 3D lightweight CNN is proposed to extract brain tumor substructures. The dilated convolutions and residual connections are used to dramatically reduce the parameters without loss of the spatial resolution and the number of parameters is only 0.5M. In the BraTS 2018 segmentation task, experiments with the validation dataset show that the proposed method helps to improve the brain tumor segmentation accuracy compared with the common merging strategy. The mean Dice scores on the validation and testing dataset are (0.743, 0.872, 0.773) and (0.645, 0.812, 0.725) for enhancing tumor core, whole tumor, and tumor core, respectively.