Artificial Intelligence in Magnetic Resonance: A Focus on Enhanced NMR Shimming and RASER MRI Artefact Correction

This thesis explores artificial intelligence (AI) methods, such as deep learning (DL), for two applications in nuclear magnetic resonance (NMR) spectroscopy and magnetic resonance imaging (MRI). Specifically, an AI-driven approach for enhancing the tedious and time-consuming shimming process in NMR spectroscopy was initiated. Additionally, the image quality of a new MRI technique, namely RASER (Radio-frequency Amplification by Stimulated emission of Radiation), was improved through DL-based artefact removal.

A highly homogeneous magnetic field, up to parts-per-billion (ppb), is crucial to achieve accurate NMR/MRI results. This can be achieved by using "shim coils" to compensate for magnetic field gradients and make the field as uniform as possible, which is a time-consuming and tedious process. Thus, proper and fast shimming is critical, which is hampered by the non-bijectivity between field distortions in the sample volume and one-dimensional NMR signals. As the primary contribution of this thesis, four studies have been developed to accelerate shimming by incorporating deep learning (DL). These AI-driven shimming studies cover the full DL pipeline from data acquisition, preprocessing, architecture design, training, and deployment.
... mehr

The first study demonstrated how an ensemble of convolutional neural network (CNN) could rapidly predict three linear shim currents from a batch of four measured spectra, significantly improving spectral quality across various substances. This approach simplified and accelerated shimming, either as a standalone method or in conjunction with traditional methods.
The second study further enhanced AI-driven shimming. It employed a temporal history through recurrent connections, combining spectra and past shim actions for fast, quasi-iterative shimming on four shims simultaneously. The approach also introduced efficient data collection through randomized dataset acquisition, allowing scalability and significant enhancements in both the speed and performance of shimming algorithms. On average, this method reduced linewidths from 4 to 0.72 Hz, demonstrating its efficiency and effectiveness in avoiding the local minima of traditional methods.
A further study addressed the challenges in parallel detection for high-throughput NMR, presenting a substantial opportunity in upscaling detection sites within the magnet bore. This approach introduced a custom probehead design allowing for parallel spectroscopy, and centred a separate shim system over each detector. Deep learning was employed to handle their overlapping, non-orthogonal shimming fields, which enabled rapid calibration of parallel detectors.
Finally, deep reinforcement learning (DRL), a subfield in AI for strategic decision-making combined with DL, was studied to improve shimming. Using DRL can eliminate the need for the acquisition of a dataset, and can optimize for a criterion directly. The first experiments in simulation and on real hardware revealed that DRL can work for shimming, but inherit further challenges.
In summary, AI introduces a powerful tool to overcome traditional challenges in achieving homogeneous magnetic fields.

Complementing these advancements in NMR, a novel MRI technique, namely RASER, was tackled. RASER MRI images are (currently) acquired along projections, but artefacts arise from non-linear interactions among the image projections and prevent qualitative statements in medical applications. These artefacts were removed by leveraging a DL approach dedicated to reconstructing distorted projections with a CNN, and further denoising the assembled 2D RASER-MRI images with the U-Net architecture. The networks were successfully trained and tested on pure synthetic and random image data that underwent a RASER simulation. The models also demonstrated generalization to real measurements.

In conclusion, integrating AI and DL into NMR shimming and RASER MRI demonstrates these technologies' enhanced capabilities and paves the way for future innovations in the field of magnetic resonance.