Meeting Banner
Abstract #2833

Statistical Local SAR Analysis by Latin Hypercube Sampling for 11.7 Tesla Brain MRI

Yu Shao1, Peng Zeng2, Joseph Murphy-Boesch3, Jeff H. Duyn3, Alan P. Koretsky3, Shumin Wang1

1Electrical and Computer Engineering, Auburn University, Auburn, AL, United States; 2Mathematics and Statistics, Auburn University, Auburn, AL, United States; 3LFMI/NINDS/NIH, Bethesda, MD, United States

Local specific absorption rate analysis is critical to the safety of high-field human MRI studies. In order to address the inter-subject variability in head dimensions and the variability in the relative position of the human body to the RF coil, applying the conventional Monte Carlo method would require a fairly large number of simulations. In order to dramatically improve the efficiency of statistical simulations, we propose a new approach based on the Latin Hypercube Sampling (LHS). The LHS can achieve the same accuracy with much smaller run size than conventional Monte Carlo sampling because it guarantees that the selected runs uniformly spread across the domain of each input variable. We demonstrate that with a few sampling points (17 samples), the expectation, the standard deviation and sensitivity to changes in conditions, such as the head geometry and its relative position, can be accurately computed when six random variables were considered. This approach appears uniquely suited for RF safety assessment.

Abstract PDF Poster PDF

Keywords

accelerated accuracy accurately achieve address appears applied assess assumed auburn axis ball brain breadth called chosen close coil coils computed computer conditions considerably conventionally convergence convergent correlation coupled critical degree deposition describe determined deviation deviations devices domain duke efficient electrical eliminate engineering equal ergonomics evaluating exclusive expectation falls field finite fitted fixation fixed geometries geometry gram guarantees head help human hypercube hypothesized importance improved inductively influences influential input intended inter internals interval landmarks largest least length local longitudinal mathematics mesh meshes mode model modeled modeling models much mutually near novel numerical often pairwise partitioned peak position positioning possibility power practical precisely probability propose proposed quicker random reduce reducing regression related represent required respectively response rotation runs safe safer safety sampling selected sense sensitivity shielded sigma significantly simplified simulated simulations smaller statistical statistics studies subject subjects sufficient surface table tabulated transmit transmitter tuned uniform uniformly unique unit variability variable variables verified watt