Abstract: In the field of infant brain tissue segmentation， infant brain tissue has problems such as low contrast and uneven gray scale. These problems lead to the unsatisfied accuracy of the existing methods. A brain MRI image tissue segmentation method was proposed based on a three-dimensional U-Net network(3D U-Net)， which combines the attention mechanism module and the pyramid structure module， to better provide model information at different levels and positions. The contextual information of the image is fully applied to reduce the loss of image information. It can also mine the interdependence and feature mapping between channel mappings to improve the feature representation of specific semantics.The DICE index results of WM (white matter) and GM (gray matter) of the algorithm proposed in the Iseg2017 dataset have increased by 0.7% compared with the previous optimal results， and the CSF (cerebrospinal fluid) index is with comparability. In the evaluation of the Iseg2019 cross-dataset challenge， the segmentation results of WM， GM in DICE ratio and ASD achieved first place，while the CSF index won the second place.

Key words: infant brain MR images; brain tissue segmentation; multimodality data; 3D deep learning