We introduced a novel reconstruction framework by combining deep learning (DL) neural network with the Projections Onto Convex Sets (POCS) algorithm, termed DL-POCS. The image restoration from undersampled images was first performed by a convolutional encoder-decoder network. Then the output from deep learning was used as initialization and extra constraints were imposed to promote the POCS reconstruction. We evaluated this approach on vastly undersampled knee MR data and found that this combined approach is superior to each of individual components alone. Our study suggests that deep learning regularized image reconstruction will have a substantial impact on data-driven accelerated MR imaging.