Improving Brain Tumor Segmentation with Multi-direction Fusion and Fine Class Prediction

Convolutional neural networks have been broadly used for medical image analysis. Due to its characteristics, segmentation of glioma is considered to be one of the most challenging tasks. In this paper, we propose a novel Multi-direction Fusion Network (MFNet) for brain tumor segmentation with 3D multimodal MRI data. Unlike conventional 3D networks, the feature-extracting process is decomposed and fused in the proposed network. Furthermore, we design an additional task called Fine Class Prediction to reinforce the encoder and prevent over-segmentation. The proposed methods finally obtain dice scores of 0.81796, 0.8227, 0.88459 for enhancing tumor, tumor core and whole tumor respectively on BraTS 2019 test set.