Denoising of motion artifacted MRI scans using conditional generative adversarial network

Patient motion causes image distortion during Magnetic Resonance Imaging (MRI) capture. These distorted motion artifact induced MRI scans are difficult to read and sometimes lead to a faulty diagnosis. The simplest solution to remove these artifacts in motion-blurred scans is to re-scan the patient. But this method is costly, time-consuming, and cannot guarantee a successful scan even after re-scanning the MRI, because the patient can still move involuntarily. Hence, correction in the motion artifact induced images is an important part of the medical imaging domain. Here we have modified a well-known conditional Generative Adversarial Network called Pix2Pix for removing motion artifacts from MRI scans. We have modified the structure of the original network and fine-tuned the parameters with the help of a database that we have created. The database was collected from a local hospital consisting of 436 images that include motion artifact induced scans and their corresponding artifact-free scans obtained by re-scanning the patients. The proposed method could achieve RMSE of 0.004 and PSNR of 27.97 dB with accuracy greater than 96 %. We expect that this method will help radiologists to save time and cost of re-scanning and will eventually help the doctors in diagnosis.