Learning motion artefacts in non - Cartesian magnetic resonance imaging

Type of publication:  Artikel
Zeitschrift: Biomedical Technology/Biomedizinische Technik
Band: 63
Nummer: S1
Jahr: 2018
Seiten: S280
DOI: 10.1515/bmt-2018-6051
Abriss: As magnetic resonance imaging is a relatively slow imaging method, motion artefacts are a major problem in many clinical applications. Motion sensitivity and appearance of image artefacts can vary considerably between imaging sequences and experimental conditions. If the raw data collection is performed in a radial fashion or using a stripe in kspace that is subsequently rotated about the origin (PROPELLER sequence), the sensitivity to motion is low. Although these sequences proved to be relatively insensitive to motion, the motion-correcting techniques built into the image reconstruction are not always robust: they fail in some situations. In addition, motion during this type of acquisition can lead to artefacts that are less easy to recognize in the final images when compared to the so-called Cartesian order of acquisition in k-space. Here, learning algorithms are used to detect motion degrading image quality. The detection algorithm can be used to inform motion correction approaches or to trigger data rejection and re-acquisition. The algorithm is trained and tested using a home-built phantom with MR-visible parts that can perform computer-controlled, reproducible movements in a clinical MR system. The motion is induced with an MR-compatible setup. The motion traces are known exactly and can be repeated accurately. These traces – together with the corresponding motioncorrupted k-space sequences as well as their non-corrupted counterparts – are used to train a deep convolutional neural network.
Autoren: Koch, Martin A.
Hagenah, Jannis
Wattenberg, Maximilian
Ernst, Floris