Effective Multipath Feature Extraction 3D CNN for Multimodal Brain Tumor Segmentation

In recent years, in order to improve the segmentation accuracy of the tumor, the researchers gradually from single mode state brain image segmentation turned to the study of multimodal brain image segmentation, usually with Flair, Tl, Tlc, T2 four modal for brain image segmentation, the four modal images can be more comprehensive information shows that the tumor of brain tumors, splits out the results of precision is higher. Due to the rapid development of deep learning in recent years, deep learning has made remarkable achievements in the field of biomedicine. However, most of brain tumor segmentation algorithms based on deep learning directly input multimodal data into the network at the same time. When multi-modal input data is fused in the first convolution layer, there will be a large loss of information, resulting in the relatively weak ability of the network to distinguish the differences between different modal data. To overcome this shortcoming, we propose an effective framework:(I) delay the fusion of different modal data;(ii) multipath processing of data with different modes;(iii) add attention mechanism to the coding part to improve the feature extraction ability of the network; The model has good universality and can be easily used in other fields of multimodal data processing. We use the brain tumor segmentation data of Brats2019 to verify the effectiveness of our algorithm through strict comparative experiments.