UCLA

MRI plays an important role in abdominal and cardiac imaging due to its excellent soft tissue contrast and high image resolution. Despite the benefit of excellent image quality, MRI acquisition is intrinsically slow, causing patient discomfort and slowing down the clinical workflow, which hinders its broad clinical use. For decades, undersampling reconstruction techniques have been investigated to accelerate MRI acquisition. Traditional parallel imaging and compressed sensing methods either have limited acceleration capability or require extensive computational and time resources. While the recent development of deep learning achieved unprecedented performance in image reconstruction and image enhancement tasks, there are challenges remaining to be solved. One challenge is the potential loss of image details due to network over-smooths or over-regularization. Another challenge is that networks may struggle to generalize well to diverse MRI data acquired under different conditions. In medical imaging, high-quality diverse datasets are challenging to acquire, especially for rare or specialized MRI applications. Lastly, for non-Cartesian sampling, the reconstruction can be challenging due to the need for time-consuming interpolation of non-Cartesian k-space onto a Cartesian basis.

The overall goal of the dissertation is to contribute to the development of deep learning-based accelerated image reconstruction techniques and investigate the challenges in network development as mentioned above. Specifically, we aim to develop deep learning networks to improve image quality and reduce artifacts and noise for the application of (1) undersampled radial MRI reconstruction in the abdomen (aims 1 and 2), and (2) ferumoxytol-enhanced cardiac cine MRI reconstruction (aim 3). In aim 1, I developed a generative adversarial network using paired undersampled and ground truth images to reduce streaking artifacts and preserve image sharpness. In aim 2, I developed a radial k-space prediction framework by training an attention-based transformer network on k-space data. By combining the acquired and predicted k-space data, the reconstructed images will have an improved signal-to-noise ratio and fewer streaking artifacts. In aim 3, I developed an unrolled spatiotemporal deep learning network for ferumoxytol-enhanced cardiac cine MRI reconstruction. The network was trained using non-contrast-enhanced bSSFP cine images and can be successfully generalized to ferumoxytol-enhanced images.