MRspecNET

This repository contains a collection of Python codes to set up and train a convolutional neural network for denoising in magnetic resonance spectroscopy (MRS). The network is trained using a set of simulated metabolite spectra, which are loaded from a file, randomly varied and decorated with Gaussian noise. This provides a pair of noisy (simulated) data and ground truth. Besides the metabolite basis sets the simulation tool offers the opportunity to add a macromolecular baseline to the simulated spectrum, which is commonly present in the human brain (for certain MRS sequences).

The main components of our library are:

• MMBG_basis(PATH,kwargs): a class that loads and stores the basis sets for the macromolecular baseline located at PATH. Accepts a dict of keyword arguments that define how the basis sets are combined into a single spectrum;
• Metab_basis(PATH,kwargs): a class that loads and stores the basis sets for the metabolites located at PATH. Accepts a dict of keyword arguments that define how the basis sets are combined into a single spectrum;
• Simulator(metab_basis,mmbg_basis): a class that contains the main simulation tool. The method Simulator.simulate() combines the data from MMBG_basis and Metab_basis into a realistic MR spectrum and adds a degree of random variation;
• MRspecNET(kernel_size,n_channels,n_layers): The neural net (a convolutional autoencoder) used for denoising. Written in PyTorch;
• train.py: The main training script. Adds noise to simulated spectra, combines them into a batch and performs stochastic gradient minimization with MSE Loss.

The concentration values we use to generate realistic simulated spectra are those found in [Hoefemann, Maike, et al. "Parameterization of metabolite and macromolecule contributions in interrelated MR spectra of human brain using multidimensional modeling." NMR in Biomedicine 33.9 (2020): e4328.]

Any comments or queries are welcome at https://frank-roesler.github.io/contact/