Funding period: May 2017 to Jul 2021

Researcher: Mario Breitkopf

Wrap-up

Keywords:
MRI, undersampling, reconstruction, deeplearning, unblackboxing

Background:
Undersampling MR images leads to an insufficient amount of data for conventional reconstruction techniques, making it an ill posed inverse problem. Deep neural networks provide promising solutions to the problem, but lack explainability.

Objective:
>> MRI acceleration, especially golden angle radial sampling, in the process making real time MRI possible 

Methods:
>> Utilizing and improving data driven neural network approaches and their analysis 

Results:
>> Up to date deep learning reconstruction methods for undersampled radial MR signal data in image and signal domain with competitive results in that field of research 

Conclusions:
Current methods still mark the starting point since they are still missing key points like holoporphic activation functions for computing complex gradients throughout neural nets.

Originality:
>> Problem specific methods that are tailored to the underlying complex valued MR problem