Imaging and Computing Seminar

Misha Kilmer , Mathematics, Tufts

Hybrid and generalized hybrid regularization for pMRI reconstruction

This is joint work with Dr. W. Scott Hoge, Brigham and Women's Hospital.

The reconstruction of MR images from accelerated parallel MR data presents as a familiar inverse problem, and system regularization techniques are often employed to ensure robust solutions. To be clinically viable, however, regularized solutions must be computed efficiently and in a manner consistent with the speed of data acquisition. Image quality in the regularized solution depends directly on the value of the regularization parameter, and therefore the regularization parameter(s) must be selected in a stable and efficient manner while simultaneously constructing candidate solutions. We present an LSQR-hybrid iterative algorithm that allows us to quickly and efficiently select a regularization parameter value for a class of parallel MRI image reconstruction problems. To achieve higher acceleration rates, parallel MRI can be combined with partial-Fourier acquisition techniques. For this special case, we propose a two-parameter constrained reconstruction problem. A generalization of the LSQR-Hybrid approach for this special case allows us to efficiently select both parameters via information generated during various runs of the LSQR-Hybrid algorithm. In both versions of the reconstruction problem, our algorithm gives high quality results for high acceleration factors in a fraction of the time compared to the naive approach. Phantom and in-vivo results will be presented.