The shapes of the multiple sclerosis of magnetic resonance imaging images are usually very different, it leads to poor performance of traditional segmentation methods of multiple sclerosis automatic. In terms of this issue, a segmentation method of medical images based on attention dilated U-Net is proposed. Firstly, the dense dilated residual block is used instead of traditional convolutional layer in the cascade connection structure of the U-Net, in order to adjust the receptive field of each scale adaptively, so that the important information in each scale is extracted. Secondly, an attention model is add between the encoder and the decoder of same scale feature maps, in order to increase the feature map weight of the interest of region in each scale, and reduce the weight of the rest region, so that the information redundancy is prevented. Finally, a mixed loss function is used to train the networks on multi-scale feature maps, in order to solve the imbalanced classification problem of multiple sclerosis segmentation. Compared experimental results demonstrate the precision of proposed method increases 10.8% compared with U-Net.