A Tri-Attention fusion guided multi-modal segmentation network

•A novel correlation description block is introduced to discover the latent multi-source correlation between modalities.•A constraint based on the correlation using KL divergence is proposed to aide the segmentation network to extract the correlated feature representation for a better segmentation.•A tri-attention fusion strategy is proposed to recalibrate the feature representation along modality-attention, spatial-attention and correlation-attention paths.•The first 3D multi-modal brain tumor segmentation network guided by tri-attention fusion is proposed.