UCSF

Magnetic Resonance Imaging (MRI) offers exceptional diagnostic capabilities but faces limitations in acquisition speed and reconstruction complexity. While deep learning methods show promise for accelerating MRI, their clinical translation is hindered by limited diverse raw training data and high computational demands.

This dissertation addresses these challenges through two complementary approaches. First, we develop generative frameworks that synthesize realistic complex-valued MRI data from widely available magnitude-only images (typical clinically). We demonstrate that reconstruction networks trained on synthetically generated data perform comparably to those trained on real acquisitions. By implementing latent space exploration and an RF coil sensitivity library, we enable the creation of diverse, realistic multi-coil k-space data from magnitude images, bridging the gap between limited research datasets and vast clinical repositories.

Second, we investigate hardware and algorithmic efficiency improvements. We implement deep learning reconstruction models on custom RISC-V architectures with specialized accelerators, exploring quantization and parallelism strategies. Additionally, we develop mixed-precision implementations of iterative reconstruction methods, particularly leveraging 8-bit floating-point formats to reduce memory requirements and computational complexity while preserving image quality.

The contributions presented collectively address barriers to clinical translation of accelerated deep learning based MRI by enabling the use of existing clinical archives for training data generation and reducing computational requirements for deployment.