UCLA

Magnetic resonance imaging (MRI) is widely adopted to assess anatomy and function of human bodies. Current clinical MRI mainly relies on contrast-weighted images and the clinicians make diagnosis by observing the relative intensity differences between tissues. In contrast, quantitative MRI directly measures the tissue physical parameters, thus providing biomarkers with enhanced objectivity, reproducibility, and sensitivity. Considering the clinicians’ reliance on conventional weighted MRI and the advantages of quantitative MRI, obtaining both is promising to provide a comprehensive tissue characterization and to benefit patient care. However, the acquisition of a full array of contrast-weighted images and parameter maps demands a prolonged scan time, which has been a major challenge.In this dissertation, we developed deep learning approaches to obtain both multi-contrast weighted images and multi-parametric maps within short scan times. Based on the physical relationships between the contrast-weighted images and the tissue parameter maps, we proposed to only acquire one of them (i.e., either weighted images or quantitative maps) and to retrospectively derive the other from the same acquired data. Both directions were explored. First, an efficient multi-parametric quantitative MRI technique was employed for acquisition, and multiple conventional T1-weighted and T2-weighted images were synthesized from the quantitative MRI data using deep learning. The method was evaluated on healthy volunteers and multiple sclerosis patients, achieving comparable image quality with true acquired images and better performance than the traditional Bloch-equation-based synthesis. In the second direction, multi-contrast weighted images were acquired following current clinical practices and deep learning approaches were developed to retrospectively estimate T1 and T2 maps from them. Both supervised and physics-guided self-supervised learning methods were developed. The supervised learning method requires a set of label parameter maps paired with the input weighted images as training data, whereas the self-supervised learning framework eliminated such a demand by incorporating physics models and formulating the optimization problem in the weighted image space. Both approaches generated results well matched with the conventionally acquired T1 and T2 maps. The self-supervised learning approach was further applied to a public glioblastoma dataset and the resulting retrospective T1 and T2 values showed potential in brain tumor characterization.