Meeting Banner
Abstract #2457

Accelerating T2 Mapping Via Under-Sampled Low-Dimensional-Structure Self-Learning and Thresholding (LOST) Reconstruction

Tri Minh Ngo1, Haiyan Ding2, Mehmet Akakaya3, Elliot R. McVeigh1, Daniel A. Herzka4

1Biomedical Engineering, Johns Hopkins School of Medicine, Baltimore, MD, United States; 2Biomedical Engineering, Tsinghua University, Beijing, China; 3Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, United States; 4Biomedical Engineering, Johns Hopkins University, Baltimore, MD, United States

Myocardial 3D T2 mapping is useful for differentiating between infarct, edema and normal tissue but currently requires prohibitively long acquisition times to be clinically useful. We apply Low-dimensional-structure self-learning and thresholding (LOST) to reconstruct an under-sampled T2 mapping dataset and compare with SENSE reconstruction of equivalent acceleration rates. For rates R3, R3.9, the mean T2 error is lower for LOST than SENSE. At rate R3 SENSE reconstructs edge details better than LOST even though SENSEs mean T2 error is higher. At rate R3.9, SENSE exhibits high noise amplification while LOST exhibits blurring of high frequency details.

Abstract PDF

Keywords

accelerated accelerating acceleration acquisition aliasing amplification artifacts auto basis beth better biomedical blurring bottom calibrated channel china clinical clinically coil coils combined computed conditioned consistency dataset datasets deaconess decay degrading details deviation differentially differentiating dimensional ding distribution edema edge enable encodes encoding enforced engineering entire equivalent error errors even exhibits exponential features fidelity frequency fully function functions generalized generating goodness head heart improved in vivo incorporate incorporated incorporating increasing infarct inverse iterations johns kaiser knowledge learning linear lists locations long lost making many mapping maps matrix medical medicine model multiplying myocardial needed noise original performance pixel pixels poor poorly prep preserves prior profile prospectively pseudo quantitative randomly reconstruct reconstructed reconstruction reconstructions reconstructs reduce reducing regression regularly rejected remaining removed representative retrospectively sampled school selected self sense sensing sensitivity similarly simulated slice space stages steps strategy structure suitable table though threshold thresholding tissue transformed typically validate varying viability volume volumes wavelets weightings window wise