Patient motion in MRI is a significant problem that can result in ghosting, ringing and blurring artifacts possibly hindering the diagnostic quality of acquired images. Acquisition of MRI images remains a lengthy process despite many improvements over the years to devise faster pulse sequences and motion correction methods. It is unlikely that a universal solution can be found.

This thesis focuses on the problem of head motion in MRI, which can be approximated as rigid-body. Existing techniques can be classified as prospective or retrospective. These approaches will be presented in a state-of-the-art, focusing on the latter methods. Deep learning approaches have demonstrated a great potential to solve many diverse and complex problems. Deep neural networks taking a motion-corrupted MRI acquisition as input, processing it through its hidden layers, and producing a corrected version as output have been proposed. This type of approach is akin to a blackbox system to some extent and can be prone to ''deep hallucinations''. The work proposed in this thesis tackles the problem in two steps. We rely on the assumption that motion information is embedded in k-space and can be at least partially retrieved. A deep convolutional neural network analyses the k-space of a motion-corrupted acquisition and estimates motion parameters. Subsequently, a model-based approach leveraging the non-uniform Fourier transform is used to correct motion artifacts given the network's predictions.

A novel protocol for generating in vivo datasets of brain scans corrupted by choreography-controlled (ChoCo) motion artifacts is proposed. It provides a way to build test sets for assessing generalization performance of the deep learning model. It also served as a basis for simulating realistic motion artifacts, which was needed to build large training sets for deep learning. Other contributions of this thesis entail a spatio-temporal upsampling method useful for obtaining more accurate estimates of fast motion events performed by volunteers in the scope of the ChoCo project. A novel k-space quality metric is also proposed. It proved useful for performing sanity checks of in vivo motion-corrupted data and also to provide insights on how to accurately simulate motion artifacts.

The k-space quality metric was evaluated qualitatively and quantitatively by means of receiver operator characteristic curves and is able to discriminate between three classes of motion degradation. The deep neural network model was trained on large datasets of brain images corrupted with simulated motion. Its generalization performance was evaluated on in vivo brain datasets acquired with the proposed ChoCo protocol. Following motion estimations, in vivo corrupted brain scans were corrected with a model-based approach, and motion artifacts were significantly reduced.