Reconstruction of strongly under-sampled MREG in the presence of field inhomogeneities

Abstract: Fast functional magnetic resonance imaging (fMRI) technique, such as MR-Encephalography (MREG), provides a magical power to see the rapid activity within our brain. However, the main challenges of MREG are the extra-high computational cost and the strong sensitivity to off-resonance effects during reconstruction. Therefore, the core goal of this thesis is to solve these challenges.
To reduce the computational cost, a time-domain principal component reconstruction (tPCR) method is developed. It contains three steps: (i) decomposing the k-t-space fMRI datasets into time-domain principal component space using singular value decomposition, (ii) reconstructing each principal component with redistributed computation power according to their weights, (iii) combining the reconstructed principal components back to image-t-space. This operation significantly improves the reconstruction efficiency, allowing higher integrated reconstruction precision with much less computational cost when compared with the traditional reconstruction.
To reduce the dynamic off-resonance artifacts caused by physiological noise and motion, a dynamic field map estimation technique based on two-shot reversed-trajectory design and deep learning is developed. The field map is estimated from the two images with reversed artifacts. It is more difficult to estimate field maps in a non-Cartesian trajectory using analytical methods, so the deep learning technique is introduced. A convolutional neural network was trained using simulated data, then the field map was estimated using the trained network at each time point. With the corrected field map, both the image quality and the sensitivity for functional analysis are improved.
In conclusion, the two techniques in this thesis make MREG reconstruction more efficient and accurate, promoting its broader applications

Angaben zum Objekt

• Standort: Deutsche Nationalbibliothek Frankfurt am Main
• Umfang: Online-Ressource
• Sprache: Englisch
• Anmerkungen: Universität Freiburg, Dissertation, 2020

Klassifikation und Themen

• Schlagwort: Deep learning

Beteiligte, Orts- und Zeitangaben

• Ereignis: Veröffentlichung
• (wo): Freiburg
• (wer): Universität
• (wann): 2020
• Urheber: Wang, Fei
• Beteiligte Personen und Organisationen: Hennig, Jürgen ; Albert-Ludwigs-Universität Freiburg. Fakultät für Mathematik und Physik

Weitere Informationen

• DOI: 10.6094/UNIFR/169784
• URN: urn:nbn:de:bsz:25-freidok-1697848
• Rechteinformation: Der Zugriff auf das Objekt ist unbeschränkt möglich.
• Letzte Aktualisierung: 25.03.2025, 13:46 MEZ