The imaging of quantitative MRI parameters enables an objective comparison of the investigated tissues on the basis of physical properties and is therefore considered an invaluable component of precision medicine. Recent work shows that long scan times for qMRI can be reduced to a fraction through data subsampling and model-based reconstructions. However, reconstruction with nonlinear models can be time-consuming and thus makes them an ideal candidate for deep learning methods. There have only been very few approaches of mapping MRI parameters by means of deep learning, and only one, where a model-augmented neural network is used to estimate M0 and T2 maps. In this master’s thesis, a U-Net is proposed that estimates M0 and T1 maps from a corresponding set of subsampled Variable Flip Angle (VFA) images, comprising the physical model consistency term that incorporates the signal model into the objective function. The acceleration potential is shown on numerical brain phantoms and on retrospectively subsampled in vivo measurements via transfer-learning for cartesian subsampling of R = 1.89, R = 3.43 and R = 5.84. It is further shown that prior knowledge of B1 can be included in the signal model but it is even possible to estimate B1 inhomogeneity maps in a separate output channel of the neural network. A comparison between learning the parameter maps in image domain or in k-space was performed, showing that training in image domain yields significantly better results without any backfolding artifacts.
|Qualification||Master of Science|
|Publication status||Published - 2020|
- quantitative MRI
- Deep Learning
- accelerated MRI