Compressed sensing magnetic resonance imaging (CS-MRI), as a traditional model-based method, provides a theoretical framework for MRI reconstruction. The problems with traditional CS-MRI reconstruction are solved by iterative numerical solvers, which have expensive computational cost and are long time consuming, limiting its application in MRI reconstruction. At present, deep learning (DL) as a new data-driven method has made rapid progress in the field of computer vision and image processing. In this article, we apply DL network model based on CS (CSDL-Net), a DL method with compressed sensing thory, for MR image reconstruction to generate high-quality images in a short time with a lower sampling rate. It includes three blocks: data fusion block, image reconstruction block, and image fusion block. We utilize single-coil part from FastMRI dataset, a widely known public database. Results show that the validation set. The average peak signal-to-noise ratio (PSNR) was 32.31 dB, and the structural similarity (SSIM) was 0.729, with a processing time of 1.21 s on the graphics processing unit (GPU). Therefore, the proposed method has a shorter reconstruction time and better image quality than the existing methods.