Examining mcDESPOT Mar 12, 2013 Jason Su MRM 2012: Lankford and Does. On the Inherent Precision of mcDESPOT. Results Summary • Good – A well done analysis of the unconstrained situation • Bad – Very different constraint scenario from the one used in practice with Stochastic Region Contraction (SRC) – Some doubts about step size and forward finite difference • Take-home message – Exchange rate and MWF could not both be estimated well – Additional phase cycles may provide benefit SRC vs. Unbiased Estimator • SRC produces a biased estimate but the coefficient of variation is well under Lankford’s 10% cut-off pcMCDESPOT.c • We have access to an old version of Sean’s source code – Results produced with both the binary provided to us (though this itself is old) match those produced from this source, so it has likely the same core fitting • However, there are some bugs in the code: – DESPOT2-FM implements an incorrect signal equation, the off-resonance estimate from this is used in mcDESPOT fits – The mcDESPOT SSFP signal equation models the magnetization before RF excitation, which is not measured what is in experiment Problem: Model is Before RF Fit w/ Data Before RF Fit w/ Data After RF Problem: “Gaussian” Sampling • The code uses a Taylor approximation of the Gaussian CDF which is fairly inaccurate – In addition, discrete uniform samples are drawn from a set of 999 bins – Not well understood how the sampling affects SRC convergence but this is definitely not Gaussian Problem: Cyclic Phase • SRC needs to be properly adapted to handle cyclic parameters, i.e. off-resonance/phase Problem: Mean Normalization • Mean normalization of SSFP data is used to reduce the fitting problem, but produces a fundamental ambiguity in the phase – At cross-over points, phase0 = phase180: the most important information is the amplitude – But this is thrown away with mean normalization Mean Normalization -> Ambiguity With Mean Normalization No Mean Normalization Idea: 3 Phase Cycles • We can still do mean normalization as long as the collected data provides a unique “signature” – With 3 phase cycles, all signals will never be equal at the same time, so the combined set of data is not degenerate after normalization 3-phase mcDESPOT • Is 3 phase cycles the future? • We can use some CRLB theory to examine how it would benefit an unbiased estimator – There is a huge improvement in estimating the off-resonance – There is some but little improvement elsewhere – SNR is matched here for constant acquisition time Current & Future Work • 3pc could be critical in scenarios with high banding – Acquiring a phase90 SSFP may not be a common option on all scanners • Re-implementing mcDESPOT fitting code in Python/Cython – Fix implementation bugs – Nearly eliminate the cost in processing addition phase cycles by taking advantage of redundancies in the signal equations – A general, open source, parameter fitting framework • What is the optimal way to sample the free parameter space? – Flip angle, TR, phase cycle, etc.