Abstract
To obtain high-quality magnetic resonance (MR) images, it is necessary to increase the size of the phase-encoding matrix (Mpe) and the number of signal averages (NSA). However, doing so increases the imaging time. In this study, we sought to reduce the imaging time by using deep learning to improve the image quality. The input image was an MR image with a short imaging time, and the training image was a high-quality MR image with a long imaging time. We used a deep denoising super resolution convolutional neural network for image improvement. Each image was divided into small patches and subjected to super-resolution processing. The optimum conditions for the input image were examined by adjusting the Mpe, NSA, and patch size. Furthermore, we examined the clinical conditions and high-quality imaging conditions for the training images. Image improvement was evaluated both objectively by using the peak signalto-noise ratio and structural similarity and subjectively by 25 radiological technologists. It was found that the images with Mpe 256 and NSA 2 and those with Mpe 256 and NSA 1 had the same quality as images obtained under clinical conditions. These results suggest that imaging time can be reduced from 90.5 s to 31.5 s and 59.5 s, respectively, by this method.
