Scattering is a fundamental phenomenon in nature. The imaging with large depth-of-field through a scattering medium is significant and valuable. In recent years, with the wide application of deep learning in computational imaging, it is urgent to study and further extend the depth-of-field in a scattering imaging system. In the paper, based on DenseNet and combined with the UNet architecture, a deep convolutional neural network model, namely DUNet, with good mobility and depth-of-field expansion ability is proposed. Moreover, the network is trained with speckle images passing through frosted glasses of different mesh, and the depth-of-field can be generalized to 50 mm away from the focal plane. The preliminary results on a rat brain slice demonstrate that the DUNet can be further implemented in the tomographic scanning of deep tissues.