Meeting Banner
Abstract #0058

Fast DSI Reconstruction with Trained Dictionaries

Berkin Bilgic1, Itthi Chatnuntawech1, Kawin Setsompop2, 3, Stephen F. Cauley4, Lawrence L. Wald2, 5, Elfar Adalsteinsson, 56

1EECS, Massachusetts Institute of Technology, Cambridge, MA, United States; 2A. A. Martinos Center for Biomedical Imaging, Dept. of Radiology, MGH, Charlestown, MA, United States; 3Harvard Medical School, Boston , MA, United States; 4A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, United States; 5Harvard-MIT Division of Health Sciences and Technology, MIT, Cambridge, MA; 6EECS, MIT, Cambridge, MA, United States

Significant benefit in Compressed Sensing (CS) reconstruction of Diffusion Spectrum Imaging (DSI) data from undersampled q-space was demonstrated when a dictionary trained for sparse representation was utilized rather than wavelet and Total Variation (TV). However, computation times of both dictionary-based and Wavelet+TV methods are on the order of days for full-brain processing. We present two algorithms that are 3 orders of magnitude faster than these CS methods with reconstruction quality comparable to the previous dictionary-CS approach.

Abstract PDF Presentation Video

Keywords

abstract acceleration achieves alluding applied audience available benefit biomedical bottom brain chosen clinicians closed coefficients collected columns comparable component compressed computation computational content cost dataset days denote density depicts dept described determined dictionaries dictionary diffusion dimensionality division employed error errors example expressed extracted fast faster find fold form full fully function generalize generating give ground health healthy in vivo investigators isotropic letting location magnitude maps matrix maximal medical minimize minimizers multiplication noise novel operator optimal optimization outermost particular penalty place previous previously principal prior probability problems processing produce property proposed quality radiology rather reconstruct reconstruction reconstructions reduce reduced regularization regularized relies relying represent representation resolution respect reveals sampled school sciences scientists seeking sensing shell simple since slew slice slightly smaller solutions solve solved solving space sparse spectrum speed subject subjects substantially subtracting successfully support system target technology trained training transform truth typical underestimates utilized variation volunteer wavelet workstation