Abstract: In order to accelerate the acquisition of Magnetic Resonance Imaging (MRI), a multi-scale feature fusion network is proposed. The network consists of two feature fusion blocks (FFB) based on dilated convolution with different dilation scale, the dilated convolution is employed for expanding the receptive field with less network parameters. Furthermore, the FFB groups several dilated convolution modules together and directly propagates the features on each local module branch through skip connections. The feature fusion blocks are used to fuse local residual features and further to produce more representative features, at the same time, the local residual learning is introduced to extract more details. Compared with other algorithms, the comprehensive experiments on different acceleration rates show that the proposed network can reconstruct high quality MR image with higher peak signal noise ratio (PSNR) and structural similarity index (SSIM).