Dissertation Talk: Low Dimensional Methods for High Dimensional MRI
Seminar | December 5 | 12-1 p.m. | 510 Soda Hall
In this talk, I will present methods to reconstruct 3D dynamic magnetic resonance images (MRI) with unprecedented spatiotemporal resolution. The problem considered is vastly underdetermined and computational demanding (trying to reconstruct ~100 GBs of image from ~2 GBs of measurements). I will first introduce a multi scale low rank matrix model to compactly represent general dynamic images. Then using this model with explicit factorization and stochastic optimization, I will present a computational and memory efficient algorithm that can reconstruct high-res 3D dynamic images from continuous non-gated MRI acquisition. Finally, I will present other enhancements to accelerate reconstruction convergence using k-space preconditioner, and to leverage external datasets using convolutional sparse coding.