Recent advances for silent MRI have shown that spatial encoding can be achieved using RF rather than linearly varying static magnetic field gradients. This has been demonstrated using homogeneous transmit (B1) fields with linearly varying phase gradients. Similar results can be achieved with linear B1 amplitude gradients with homogeneous phase. The efficacy of either method is limited by a maximum B1 gradient strength (phase or magnitude) per specific absorption rate. Here we demonstrate a novel approach to relieve this restriction where highly nonlinear B1 gradients can be used for combined amplitude and phase modulation with reconstruction using state-of-the-art machine learning models.
Magnetic resonance imaging (MRI) methods use non-ionizing radiation in combination with a large static magnetic field to probe the magnetic properties of biological tissue. Current MRI methods rely on the use of linearly varying static magnetic fields generated by gradient coils to spatially encode the MRI signal. The utility of MRI is limited by these gradients causing high levels of acoustic noise and peripheral nerve stimulation (PNS). This work reports a novel MRI method that entirely eliminates both acoustic noise and PNS by transferring the burden of spatial encoding from acoustically loud gradient coils to silent radio frequency (RF) devices. Achieving spatial mapping of the MR signal using highly parallel radio frequency transmitters and receivers, rather than gradient fields, lies at the core of this approach.
Previous work in silent MRI1 has shown that a homogeneous transmit with linearly varying B1 phase can be used to traverse k-space. Analogously, linearly varying B1 magnitude fields can achieve a similar result. Spatial encoding can be achieved using a surface coil’s spatial B1 amplitude profile to create a relationship between transmit power and duration to the flip angle of the magnetization. The result is a transverse magnetization with banding artifacts (Figure 1D) that act as encoding functions. The issues that arise in reconstruction are primarily due to the non-orthogonality of these encoding functions, prohibiting analytic reconstruction. It can be shown that particular subsets of power levels can be chosen to increase orthogonality between transmitters and receiver pairs. Unfortunately, this subset quickly exceeds SAR limitations and can vary between receivers. As such, it would be beneficial to create a reconstruction method that can use signal modulation due to small changes of power for image reconstruction. Here we show this can be achieved using deep machine learning for reconstruction.
A representative image reconstruction is shown in Figure 2. The encoding of higher spatial frequencies in the periphery do not translate to higher classification accuracy. The high anatomical variations in cortical folding patterns appears to drive the miss-classification, whereas deep brain CSF appears to have higher classification accuracy than expected. The confusion matrix from each method is shown in Figure 3. The decrease in classification accuracy for the proton density data is a minor setback compared to the gain of 2-3 orders of magnitude on the temporal efficiency.
Current assumptions that were used in this study, but which require further investigation include wavelength effects for improved B1 estimation and safety limits, and the introduction of noise. The addition of more subjects in the training set would also improve the classification of cortical structures known to have a high degree of anatomical variation.