Detecting mild traumatic brain injury from MEG data using machine learning

This repository contains the code used for the Master's thesis by Veera Itälinna.

Repository structure

• dicom2nifti

  Contains a MATLAB script for converting DICOM files to NIfTI format.

• pipeline-tbi

  Contains the steps of the source modelling pipeline (adapted from Rantala, 2020).

• model

  Contains the code for training and evaluating the model and visualizing the results.

The source modelling pipeline

1. Run FreeSurfer on the MRI images (pipeline-tbi/freesurfer)
2. Make BEM meshes with Watershed algorithm (pipeline-tbi/watershed)
3. Perform ICA on the raw MEG measurements, and apply the ICA solutions also to the empty room recordings (pipeline-tbi/ica)
4. Calculate noise covariance matrices from the empty room recordings (pipeline-tbi/noisecov)
5. (Optional: Increase the dataset size using a sliding window approach on the MEG measurements (pipeline-tbi/window))
6. Compute the coordinate transformations (pipeline-tbi/trans)
7. Compute source PSDs and morph to an average brain (pipeline-tbi/psd)

Creating the dataset used in the thesis

1. Calculate mean and standard deviation matrices from the normative samples (pipeline-tbi/averages)
2. Calculate Z-score maps (pipeline-tbi/zmap)
3. Parcellate the Z-maps (pipeline-tbi/parc)
4. Collect all parcellated Z-maps into a single dataset (pipeline-tbi/zmap/create_zmap_dataset.py)

Training and evaluating the model

Training and evaluating the model is done by running the Python script model/run_zmap_svm.py with appropriate options.

usage: run_zmap_svm.py [-h] [-f FIT_PARAMS] [-g] [-r] [-n] [--fs]
                       [--norm-data {full,age,random}] [-v] [-p]

optional arguments:
  -h, --help            show this help message and exit
  -f FIT_PARAMS, --fit-params FIT_PARAMS
                        Parameters to pass to the classifier (JSON string)
  -g, --grid            Use grid search for model selection
  -r, --repeat          Use repeated cross-validation
  -n, --nested          Use nested cross-validation for model selection and
                        validation
  --fs                  Apply feature selection
  --norm-data {full,age,random}
                        Select what normative data to use
  -v, --visualize       Visualize the results
  -p, --perm-test       Use permutation test

How to obtain the results presented in the thesis:

• Nested and repeated cross-validation with full normative data and all features: run_zmap_svm.py -n -r
• Nested and repeated cross-validation with full normative data and feature selection: run_zmap_svm.py -n -r --fs
• Nested and repeated cross-validation with age-matched normative data and all features: run_zmap_svm.py -n -r --norm-data age
• Nested and repeated cross-validation with age-matched normative data and feature selection: run_zmap_svm.py -n -r --fs --norm-data age
• Nested and repeated cross-validation with random normative data and all features: run_zmap_svm.py -n -r --norm-data random*
• Nested and repeated cross-validation with random normative data and feature selection: run_zmap_svm.py -n -r --fs --norm-data random*

* repeated three times with different random seeds and averaged the results

References

Rantala, A. (2020). Creating a normative database of resting-state brain activity from a large number of MEG recordings [Aalto University School of Science]. http://urn.fi/URN:NBN:fi:aalto-202101311732