Rapid approximate Bayesian parameter inference is investigated for $$$T_2$$$ analysis of multi spin-echo signals. We demonstrate that rapid inference using Rician noise distributions is possible using maximum likelihood estimation (MLE) if the MLE procedure is initialized with samples drawn from an approximate Bayesian posterior. This posterior is learned using a conditional variational autoencoder (CVAE) network, trained on exclusively on simulated data. Nevertheless, we show good generalization to three diverse datasets, including improved inference accuracy compared to standard nonnegative least squares-based methods which implicitly assume Gaussian noise.
This abstract and the presentation materials are available to members only; a login is required.