Fitting matrix MRI data with shallow neural network

Hello, very newbie basic question. I have only been using Matlab for a couple of months, so still unfamiliar with a lot of the high-level functions and technical details. As a basic project to introduce me to biomedical applications of Matlab, I have been give some MRI data consisting of 3 matrices, B0 and B1+ field maps and amplitude data, which is subject‐specific structural information from a quick scan.

The AMP data is a 4D double matrix, B0 a 3D double matrix, and B1+ a 4D complex double matrix. Setting aside the issues of the slices, resizing the matrices to the same dimensions etc, I am trying to use basic machine learning, to see if the AMP data can be used to predict the field maps, similar to this paper. The known information on which the training would happen is the existing field data.

From my initial quick search, I've decided to try to use Matlab's Machine Learning toolbox, and am trying to figure out how to manipulate the data in a way that can be used by basic ML. My question is this: is it even possible to put this data in such a data structure, that it can be used with Matlab's neural fitting app, and if so how? If it only accepts 2D input, and it is impossible to convert the matrices to 2D ones without losing information, could you give me some hints as to where I should head, for a very basic, quick and proof of concept approach to generating field distribution maps from the patient data? Thanks in advance.

Answer 1

[Bernhard Suhm]

You'll need to "serialize" your multidimensional matrices, but that's standard practice in machine learning and MATLAB has tools to resize and reshape matrices. Given you have an image application, I suggest you start experimenting with deep learning, not shallow neural networks, since that's the state of the art for that type of application. A recent similar question provides you with links how to get started with that. You can train a deep learning network in MATLAB with just a few lines of code.