Abstract:Because the feature extraction in MRI brain tumor segmentation can not adapt to the arbitrary and differential of the brain images, this paper proposed a novel segmentation method based on multimodal 3D convolutional neural networks (CNNs) features. As a supervised features extraction method, CNNs can extract the texture, shape and structure characteristics automatically according to the characteristics of the image itself in the training process. By combining the 2D multimodal MR images into 3D primitive characteristics and utilizing 3D-CNNs to extract features, it is more conducive to extract the different information between different modal, remove the interference information among the different modal, shrinkthe neighborhood size of the original characteristics, and adapt to different tumor size change image layer on the same and further improve the accuracy of MRI brain tumor segmentation. The experimental results showed that compared with the current unsupervised feature extraction method and 2D-CNNs, this method can adapt to the diversity and dynamics of different patients and modals, and improve the segmentation accuracy of brain tumors.