Doctoral Thesis: On Improving the Acquisition and Reconstruction of Spatio-Temporal Magnetic Resonance Imaging

Friday, August 19
12:00 pm - 1:00 pm

Siddharth Srinivasan Iyer


Magnetic Resonance Imaging (MRI) is a non-invasive but slow imaging modality that provides unparalleled flexibility in acquiring multiple forms of soft-tissue contrast. Recently, there has been a lot of interest in mapping the inherent magnetization properties of the underlying human tissue and in temporally resolving the acquired data. Broadly classified as spatio-temporal MRI, these methods yield unprecedented details of the human anatomy and function, improving clinical diagnostic performance and prognosis. However, such methods are inherently high-dimensional, resulting in encoding-intensive data acquisition processes and computationally-intensive reconstructions. This begets long acquisition and reconstruction times, making such methods difficult to integrate into clinical workflows. This thesis aims to improve the acquisition and reconstruction times of spatio-temporal MRI to enable its use in a clinical setting.

From the MRI reconstruction angle, the iterative proximal gradient methods typically used to solve the formulated regularized linear inverse problems are studied, and a generalizable polynomial-based preconditioner is proposed to accelerate iterative convergence, resulting in faster reconstruction times. Further reduction in reconstruction times is then achieved by using deep-learning to jump-start the above polynomial-preconditioned iterative algorithm. This initialization approach is synergystic in that the iterative algorithm guards against deep-learning hallucinations while still leveraging the fast processing times of deep-learning models to reduce the required number of iterations of the iterative algorithm.

With respect to MRI acquisition, the goal is to improve the encoding-ability of the imaging process to reduce MRI examination times. First, the spatio-temporal MRI method T2-Shuffling is studied, which reconstructs multiple T2-weighted images from a single volumetric fast spin-echo (3D-FSE) scan. The T2-Shuffling model is augmented with the Wave-CAIPI parallel imaging method that improves encoding by better utilizing MRI receive-channel information using additional sinusoidal gradients. The resulting “Wave-Shuffling” approach is applied to 3D-FSE and Magnetization-Prepared Rapid Gradient-Echo (MPRAGE) to achieve rapid, 1 mm-isotropic resolution, time-resolved structural imaging. For brain imaging application with 32-channel coil at 3T, Wave-Shuffling MPRAGE provides comparable reconstruction at approximately 8 times faster acquisition compared to Shuffling MPRAGE. Finally, an ultra-fast unified, rapid calibration sequence termed Physics Calibration (PhysiCal) is proposed that uses a carefully designed mix of full and variable density sampling acquisitions across echoes for robust and accurate recovery of whole-brain B0, B1+, and coil sensitivity maps. This detailed calibration will improve the accuracy of spatio-temporal MRI and aid the quantification of tissue parameters.

The proposed methods demonstrate how the acquisition and reconstruction times in spatio-temporal MRI can be reduced by encoding-design and algorithmic-design respectively, and provide a framework for the clinical translation of such high-dimensional imaging methods.


  • Date: Friday, August 19
  • Time: 12:00 pm - 1:00 pm
  • Category:
Additional Location Details:

The defense is taking place remotely at Stanford University, David Packard Hall (9 am PST).

Thesis Supervisor: Dr. Kawin Setsompop

If you wish to attend remotely, please contact the doctoral candidate for details,