Machine learning for fMRI brain data

The goal of this thesis is to explore different Machine Learning techniques in order to improve fMRI decoding on a specific dataset. This is part of the more global research aiming to understand how the human brain works. The examined dataset contains motion direction information induced by visual and auditory stimuli for 23 subjects. The perception of motion direction stimuli is handled by the V5/hMT+ area of the visual cortex, while motion direction for auditory stimuli is encoded by the Planum Temporale (PT). The aim of the trained classifier is to predict motion direction (i.e. left, right, up or down) using spherical Regions Of Interest (ROI) to select useful data. We analysed the effect of various methods such as feature selection, data augmentation as well as the results obtained with different classifiers. In addition to this, we examined the implications of a change in the size of said ROIs and the type of data used (beta-maps or t-maps). Even though we observed some significant accuracy changes in our results, we did not observe major improvements for decoding. However, we noticed both the LR and SVM classifiers behave in similar fashions when trained on our data, LR being the best.