Physics-Driven Deep Learning Techniques for High-Resolution MRI

Abstract

Magnetic resonance imaging (MRI) is a non-invasive diagnostic tool used in clinics to evaluate the functional properties of the human body with superior soft-tissue contrast. Scan duration is a major issue in MRI that requires trade-offs between signal-to-noise ratio (SNR), spatio-temporal resolution, and coverage leading to numerous challenges. The need for faster MRI acquisition is particularly important in cardiac imaging and functional MRI (fMRI), where improved spatio-temporal resolution is essential for better coverage and evaluation. To tackle these issues, accelerated MRI techniques have been developed, such as parallel imaging, simultaneous multi-slice (SMS) imaging and compressed sensing (CS). Although advanced image reconstruction techniques are applied to reduce scan time while maintaining high-quality images, these techniques are limited in certain ways. Hand-crafted sparsity assumptions, blocky artifacts due to reconstruction errors, time-consuming parameter tuning, and long reconstruction times due to the iterative nature of the algorithms are the main limitations of CS. Recently, physics-guided deep learning (PG-DL) or physics-driven deep learning (PD-DL) reconstruction has gained immense interest in fast MRI. PD-DL is particularly useful because it combines the benefits of MRI physics with advanced neural network-based regularization techniques. On the other hand, PD-DL has already shown improved image quality compared to parallel imaging and CS and has led to unprecedented acceleration rates. Yet, PD-DL has its own limitations some of which are, limited training raw-data availability, overregularization or artificially hallucinated image features, generalizability issues across different SNRs, and sensitivity to noise. In this thesis, novel reconstruction methods were introduced to address these challenges using parallel imaging, cutting-edge SMS techniques, and state-of-the-art PD-DL reconstruction. First, we introduced an SMS reconstruction technique that was applied to cardiac MRI (CMR) to achieve faster heart coverage without compromising the image quality. This method addressed noise amplification and inter-slice leakage problems in accelerated SMS imaging using an extended k-space approach where SMS acceleration was characterized as a uniform sub-sampling in the readout direction. Second, we proposed to encode signal intensity variations across image phases into the forward operator of the MRI inverse problem without altering coil sensitivities to tackle the generalizability issue across different SNRs/contrast. This led to a more uniform contrast across the image series, which in turn facilitated generalizability for PD-DL methods. Third, we proposed to use a subject-specific self-supervised physics-guided deep learning reconstruction that exploits spatio-temporal correlations by using a 3D convolutional neural network. This network was trained on a subject of interest without a database to overcome the challenging database learning process of cardiac motion patterns for free-breathing CMR. Fourth, we extended a self-supervised PG-DL reconstruction to 20-fold accelerated 7T fMRI to show functional precision and temporal effects in the subsequent fMRI analysis were not altered by deep learning reconstruction leading to ultra-high acceleration rates with SMS and in-plane acceleration. Lastly, we re-envisioned the conventional fMRI computational imaging pipeline. Instead of performing reconstruction followed by denoising, we achieved improved image quality by employing PG-DL reconstruction after denoising the raw k-space leading to a synergistic combination of thermal noise suppression followed by deep learning reconstruction which leveraged the best of both worlds.