A Tri-Attention fusion guided multi-modal segmentation network

In the field of multimodal segmentation, correlation between different modalities can be considered for improving segmentation results. Considering MR modalities, in this paper, we propose a multi-modality network guided by novel tri-attention fusion. Our includes N model-independent encoding paths with image sources, fusion block, dual-attention and decoding path. The model independent capture modality-specific features from modalities. that not all extracted encoders are useful to use dual attention based re-weight along modality space paths, which suppress less informative emphasize ones each at positions. Since there exists strong on module form block. module, description block is first used learn then constraint guide latent correlated more relevant segmentation. Finally, obtained fused feature representation projected decoder obtain experiment results tested BraTS 2018 dataset brain tumor demonstrate effectiveness our proposed method.