Model-based reconstruction methods for CT perfusion imaging
M1.2 | Under-sampled MRI for percutaneous intervention
Magnetic resonance imaging (MRI) is an inherently slow process turning the real-time monitoring of a patient during interventions into a challenging task. Discarding image signal parts (i.e. undersampling) during data acquisition might be one way to shorten scan times, however negatively affecting image quality.This sub-project focuses on the reconstruction of highly undersampled MR data, which equals solving an enormous underdetermined system of equations with an infinite number of solutions.
To cope with this task, it is useful to take additional information into account by, for instance, integrating prior information from planning datasets or clinical scans acquired on a daily basis.
Machine learning algorithms provide means to efficiently make use of those already existing information, not least allowing for feeding pre-existing data into a neural network - the latter representing a computational model being based on a biological network of neurons like the human brain.
In contrast to conventional reconstruction software, artificial neural networks are "able to learn or autonomously adjust” relevant parameters from training datasets, which can in turn be used to support the reconstruction of the undersampled image data.
The application of this smart method in interventional MRI will significantly speed up image acquisition, moreover facilitating real-time, minimal-invasive interventions of e.g. liver metastases.