I recently had the wonderful opportunity to present our deep learning work in magnetic resonance imaging at the ML4AU CoP showcase event to more than 65 participants. I showed one example on efficiently solving an ill-posed inverse problem for quantitative susceptibility mapping that was trained with purely synthetic data describing the forward problem and the network learns an efficient inversion of this. I showed the challenges of the first prototypes and how we are addressing these in the current works to enable a translation of this technique to a clinical MRI system. A second example shows the use of deep learning in predicting calibration data of the MRI system and how we inform a physics calculation with this initial prediction to make sure that the results are plausible and deliver a robust imaging result. Finally, I highlighted some of the open-source aspects of this work and show how we shared the data, model and training code with the MRI community to make it easier to build on our work and improve it further.
This is the talk recording:
and here the slides: