MR-double-zero
![Diagram of the proposed sequence development workflow called MR-double-zero. The core principle is a closed loop sequence parameter optimization running on a real MR scanner. A derivative-free optimizer ("nevergrad") sends the parametrized sequence to the MR scanner. The acquired raw data are mapped to a contrast prediction, which is compared to the given target contrast. The mean squared error between prediction and target informs the optimizer how to update the sequence parameters for the next iteration.](/717930/original-1702973889.jpg?t=eyJ3aWR0aCI6MjQ2LCJvYmpfaWQiOjcxNzkzMH0%3D--ebf87f230ba8d0b622e82a028bd3c67932fe013e)
Traditionally, the discovery of new MRI contrasts was often a trial-and-error process, typically involving laborious human interaction with the MR scanner or sophisticated simulations. For the latter, a theoretical description of the underlying MR physics is required to capture the desired contrast effects. In this context, a framework for MR sequence parameter optimization based on differentiable Bloch simulations has recently been introduced as "MRzero". However, such a theoretical model inevitably requires limiting assumptions that may neglect concomitant effects and imperfections that would occur in reality.
Here, we investigate whether novel contrasts can be found by performing numerical optimization directly on a real MR scanner instead of a simulation. To this end, a derivative-free optimization algorithm is set up to repeatedly update and execute a parametrized sequence on the scanner and map the acquired signals to a given target contrast. This approach requires neither a model nor human interaction with the scanner, therefore we call it "MR-double-zero".
To mimic a real contrast discovery, as a proof-of-principle, we pretend that the chemical exchange saturation transfer (CEST) effect of creatine guanidine protons is unknown and to be discovered autonomously with MR-double-zero. For this purpose, we provided a phantom consisting of appropriately prepared sample solutions of different creatine concentrations, which were made indistinguishable in conventional T1 and T2 contrasts. It was demonstrated that MR-double-zero jointly finds optimized RF preparation parameters and image post-processing that allow quantification of absolute creatine concentration. The optimization was completed after 300 iterations for 6 optimizable parameters, which took ~3h.