 |
 |
Dear Colin, > I was trying to come up with a working implementation of the "eddy" using the paper "Non-parametric representation and prediction of single- and multi-shell diffusion-weighted MRI data using Gaussian processes". I am using MATLAB for implementation, the Gaussian process regression (fitrgp) is an already inbuilt function in MATLAB. My intuition is that this function is extensible enough to be used in place of you Guassian process functionality, but your opinion would be highly insightful and useful. > > Fitrgp documentation: https://www.mathworks.com/help/stats/fitrgp.html from a quick look at the documentation for fitgrp it certainly looks like it can do everything you need to implement the GP predictor in eddy. You would need to write your own ‘KernelFunction’, but I am sure you have seen that. In the paper I describe “Marginal Likelihood Maximisation” for hyperparameter optimisation, but I have since found that Leave One Out is more robust, so that is what eddy is currently using. So, if you want to build a GP model to make predictions for diffusion data I think that would be quite straightforward. Mimicking all the functionality in eddy for movement/distortion correction would be more challenging. Good luck Jesper