Generation of Synthetic MRI with Deep Optical Flow Field Estimation for Faster Imaging

Magnetic Resonance Imaging (MRI) is an effective, non-invasive, and revolutionary imaging technique used to diagnose, study, and analyze chemical and physical structures inside the body. MR image acquisition suffers from two significant problems. First of all, prolonged scanning sessions are inconvenient and costly for patients. Secondly, the respiratory motion of the patients or external noise cause artifacts. Addressing these two issues is a strong motivation for making MRI procedures faster and more accurate. A wide variety of methods have been used to shorten the MR image acquisition process by optimizing current techniques or improving the mechanical and computational performance of the scanners. An appealing solution for the mentioned problems is to scan fewer MR images and generate in-between images to make the MR image acquisition faster. Also, in the presence of motion artifacts, we can reconstruct the imperfect images, if such a technique is available. In this thesis, we trained and applied a deep learning model to synthesize an arbitrary number of intermediate MR images by estimating optical flow vectors of two consecutive MRI slices or frames. We investigated the performance of synthetic MR images produced by this method and compared them with one of the available-to-use related methods. The evaluation results show that this technique produces high quality multiple intermediate images and outperforms the related method.