Abstract: Electron paramagnetic resonance imaging (EPRI) is an advanced method for imaging tumor oxygen concentration. The current bottleneck of EPRI is the extremely long scanning time. Sparse reconstruction, i.e., image reconstruction from projection data obtained from a sparse perspective, is an effective rapid-imaging method. However, sparse-reconstructed images using the classical filtered back-projection (FBP) algorithm contain severe streak artifacts, which affects their subsequent processing. Therefore, this study investigates the EPRI sparse-reconstruction method based on deep learning. We propose a residual, attention-based, dense U-network (RAD-UNet) to suppress streak artifacts. The input of the network is an image with streak artifacts from FBP sparse reconstruction, whereas the label is a high-quality image from FBP dense reconstruction. Through large-scale data training, the network suppresses streak artifacts. In image reconstruction, an EPRI image with streak artifacts is first sparse reconstructed from the FBP algorithm. Subsequently, it is input to the RAD-UNet for artifact suppression. Finally, a high-quality reconstructed image is obtained. The network couples advanced strategies such as residual connection, attention mechanism, dense connection, and perceptual loss, thus improving the nonlinear fitting ability of the network and the network’s ability to suppress streak artifacts. Experimental results of actual data reconstruction show that compared with the existing three representative deep convolution networks, the investigated network offers a higher sparse-reconstruction accuracy and greater ability in suppressing streak artifacts.