Magnetic Resonance Imaging (MRI) is a non-invasive imaging modality frequently used in clinic. MRI acquisition, particularly for dynamic imaging, is often time intensive. To mitigate this challenge, undersampling techniques are employed, wherein only a subset of k-space data is collected to reduce scan time. Advanced reconstruction algorithms are then applied to recover high-quality images. A common method is compressed sensing, but its reconstruction requires a large number of iterations, resulting in long reconstruction time.
Recently, deep learning MR reconstruction has shown advantages over conventional reconstruction methods. In deep learning reconstruction, a neural network will be trained with a large set of training data, trying to output high-quality images similar to the fully sampled labels. Once trained, it can directly process undersampled input data and produce the corresponding reconstructed images. This direct approach substantially reduces reconstruction time while maintaining or improving image quality.
Despite the great progress, there were some limitations in previous deep learning reconstruction. Some works did not include prior MR physics knowledge into their model, leading to suboptimal results. For quantitative or dynamic MRI, non-Cartesian acquisition is common because it can reduce motion artifacts, but there was no efficient non-Cartesian data-consistency layer for that. Another limitation is the mismatch between training data and “true” inference data, known as “implicit data crime”. MR deep learning training data are often physically unrealistic since many of them are simulated from fully sampled images. The network’s performance may substantially decline when those methods are applied to real-world data.
This dissertation focuses on physics-based deep learning for MR reconstruction and enhancement. The first goal is to create a relaxation-aware super-resolution neural network for fast, high-resolution TSE MRI, using physics-based resolution degradation to avoid “implicit data crime”. The second goal is to develop a physics-based deep subspace learning reconstruction for MR Multitasking, a recently developed motion-resolved quantitative MR technique. The development includes a non-Cartesian deep subspace data-consistency layer and advancements in SMS and AI-assisted reconstruction. Lastly, a self-supervised learning reconstruction method was developed to improve quantification repeatability for quantitative MRI without needing reference reconstruction for training.
