Machine learning for multi-subject fMRI analysis

Abstract

Functional magnetic resonance imaging (fMRI) data analysis plays an essential role in helping researchers understand how the brain works. As a non-invasive brain imaging technique, fMRI records brain activities with high spatial resolutions while the subjects were performing specific cognitive tasks. fMRI data analysis typically aims to exploit task-related patterns in the data. Since fMRI data is high-dimensional and noisy, such analysis can be challenging. This thesis introduces new machine learning methods to alleviate this data scarcity issue in fMRI data analysis by aggregating information from multiple subjects in multiple fMRI datasets.

We primarily explore three methods to aggregate information from multi-subject fMRI datasets. We first introduce a family of factor models that extract the shared representations in a multi-subject fMRI dataset. Combined with searchlight analysis or regularizers, the factor models can also find the location of the shared information.

We then introduce a method to perform transfer learning between fMRI datasets. By jointly learning a probabilistic model using multiple datasets, we improve prediction accuracy on all the datasets involved. We assume the datasets are directly or indirectly linked through sets of partially shared subjects. This method is particularly useful when one or more of the datasets are small. We demonstrate how this method can be applied to neuroscience problems, such as handling missing data and generating text from fMRI data.

We then look into a method to utilize the covariance structure of the stimuli the subjects received during data collection to extract task-related shared representations. The stimuli structure is computed by treating the stimuli in various formats as feature vectors. This method can be applied to separate information that can or cannot be explained by the stimuli and identify brain regions that are more task-related.