This study tests a super-resolution deep learning reconstruction (SR-DLR) algorithm to enhance magnetic resonance (MR) myelography images for lumbar spinal stenosis (LSS) diagnosis. MR myelography is a non-invasive method of obtaining images of spinal canal contents, such as the cauda equina and filum terminale, for LSS assessment. Currently, zero-filling interpolation (ZIP) is the standard practice to gain high-resolution spinal canal images, though the presence of artifacts limits its resolution. Deep learning-based reconstruction (DLR) methods have been developed to limit noise, though the SR-DLR demonstrates an even more powerful technique to offer high spatial resolution images like ZIP without substantial noise or artifacts. Retrospective data was analyzed from 40 lumbar MR myelography images and subsequently reconstructed into three types of images (SR-DLR, DLR, and ZIP images) for comparison. The SR-DLR algorithm incorporates two neural networks: the first trained to reduce noise and the second to suppress artifact levels. Three radiologists of varying experience (two, three, seven years) were blinded to patient information and image type and tasked to evaluate the images based on six criteria (number of levels at which LSS is observed, spinal canal structure resolution, image sharpness, noise, artifacts, overall quality). SR-DLR demonstrated superior ratings in all criteria (p < 0.001) except for artifacts, which was relatively the same compared to the DLR and ZIP images (p greater than or equal to 0.037). Furthermore, while there were no significant signal-to-noise ratio differences between the three groups, interobserver agreement (kappa value = 0.819) was best in SR-DLR, which indicates improved image consistency and possible diagnostic reliability.