To acquire MR signals, detectors (or “coils”) have to be put around the imaged object in the magnet. The extremely tiny signal sensed by these coils must be amplified and this receiver chain should introduce as low noise as possible.

Low-field MRI typically means low-frequency signals (few MHz, or “RF” range); in such cases, the noise is mainly generated by the coil itself, contrary to the noise from the sample in higher-field scanners. Therefore, optimizing the RF coils can greatly improve the quality of the acquired MR images and allow faster scans. 

At the AMT Center, we simulate, build and test our own coils with different geometries and features depending on the desired application. Moreover, we build electronics for low-noise amplifiers at 4.3 MHz in-house for our receiver chains.

Compared to other imaging modalities, MRI is a rather slow technique for which the data sampling can require a longer acquisition time. This is especially true in the case of low field MRI, which is characterised by a lower SNR, in light of the reduced net magnetisation (proportional to B₀). To circumvent this issue, signal averaging is usually performed, resulting in a further increase in scan time. While reducing the extent of the sampled k-space can speed up the acquisition process, this also impacts the resolution of the acquired images, hence hindering the detectability of small anatomic features and affecting their clinical relevance.

Fast acquisition sequences (spoiled gradient-echo, turbo spin-echo, FLASH sequences, ...) that maximise k-space acquisition can be found in the literature. K-space undersampling represents another popular acceleration method, although an accurate definition of the portion of k-space to be retained must be carefully planned, to avoid incurring into artifacts. Developments on the hardware side can also lead to an increased SNR, hence allowing less signal averaging.

At the AMT centre, we combine all these aspects in order to make the most of the reduced signal. Ultimately, we aim to develop MRI protocols characterised by acquisition times comparable with those achieved when using conventional scanners.

Despite promising benefits for a more accessible diagnosis, the low signal to noise ratio per unit time of low-field MRI leading to long time acquisitions is challenging its relevance at clinical level. Acquiring less data by sampling the k-space at frequency lower than Nyquist level (undersampling) is an effective way to speed up the acquisition, however, this causes aliasing when image is reconstructed with Fourier Transform. Recently, deep learning has shown promising results in reconstructing alias-free undersampled images.

At the AMT center, we investigate deep learning approaches that best fit the current needs at low-field regime. Additionally, we validate these approaches on acquired undersampled images.