hide
Free keywords:
-
Abstract:
Introduction The discovery of new MRI contrasts often happened hitherto by trial-and-error. Here we consider whether this can be formulated as an optimization problem, still making use of a real MR scanner. Whereas traditionally an analytical description of the contrast mechanism (a model) is required, thus having to make limiting assumptions, the presented approach requires neither a model nor human interaction with the scanner; thus, we call this approach MR-double-zero following the previously published model-based approach termed MR-zero [1]. Methods Samples of different creatine concentrations (cCr=0…120mM) are created, T1 and T2 relaxation times are adjusted to in vivo like values [2] and glucose is added as a confounding CEST pool. The MR scanner is controlled by an optimizer using Pulseq files [3] sent via network to the host PC (Fig. 1). Data flow back to the optimizer ([4] implemented in [5]) on a local PC for reconstruction. For each iteration the parameterized sequence gets updated by the optimizer and the data (MRI; up to 3 images) are mapped to the target (T=cCr) by linear regression as T=[MRI, MRI², …] β. Higher order powers of the pixel intensities (e.g. MRI²) are included to enable more flexible mapping. The sequence consists of a 2D readout with an RF preparation pulse train parameterized by B1,i, Δωi (off-resonance) , np,i (number of pulses) for i=1…3 images. Pulse duration tp,i and delays td,i are fixed. It is pretended that relaxation effects are known but CEST is not. Results The proposed framework learned to map creatine concentration using off-resonant RF preparation. A direct mapping based on T1 and T2 is not possible but the optimizer makes use of the “unknown” CEST mechanism. Within as little as 300 iterations (duration ~3h), decent mapping independent of confounding glucose concentration is achieved by designing both acquisition (B1,i, Δωi, np,i,) and mapping to the target (β coefficients) as shown in Fig. 2. Discussion In contrast to a previously published approach [6], both acquisition and evaluation are jointly optimized on a real MR scanner without any predetermined model or human interaction besides providing the target and suitable samples. The proposed method is intended as a paradigm shift towards autonomous, model-free and target-driven sequence design. Besides sequence design, the framework may be used to calibrate system imperfections or for testing hypotheses as to whether and how arbitrary targets could be accessed with MRI applied as a tool.
