Motion Correction in Brain MRI using Deep Image Prior

Motion correction in MRI has been successful with deep learning techniques to reduce motion artifacts. However, these methods were based on supervised learning, which have required a large dataset to train neural networks. To overcome this issue, we propose a new technique to correct motion artifacts without any training dataset, which optimizes a neural network only with a single motion-corrupted image. We adopt Deep Image Prior framework to capture image priors from a convolutional neural network using motion simulation based on MR physics. Using the deep image prior, our method finely reduces motion artifacts without any dataset.