Flexible FMRI Data Analysis Using Machine Learning

Abstract

Functional magnetic resonance imaging (FMRI) is a non-invasive neuroimaging tool for capturing brain neural activations. This dissertation proposes various machine learning methods for fMRI data analysis focusing on estimating the neural activation maps and analyzing their relations to brain functions and disorders. First, a novel dictionary learning (DL) method is proposed for estimating brain neural activation maps by exploiting the sparse nature of brain activations. The subject group attributes are incorporated into a supervised DL formulation to characterize the activation maps that are shared across subject groups and those that can explain the group differences. The proposed method is tested with real task fMRI data sets from schizophrenic subjects and healthy controls. The benefits of our method are reflected on generating more stable results, finding novel maps that are not estimated from benchmark methods, and estimating task-related maps that are able to significantly discriminate the subject groups. The results are further validated using a correlation analysis with neuropsychological test scores. Then, the aforementioned DL method is extended for analyzing multisubject multiset fMRI data sets. The subject group attributes are again incorporated. The multiple fMRI data sets, such as the task fMRI data sets acquired from different tasks, are analyzed jointly. From the analysis, four types of maps emerge. Namely, the maps that are either shared or showing group differences across subject groups are estimated. Also, the maps that are common across data sets and those that are unique to a data set are identified. Importantly, the algorithm can flexibly determine the map types without rigid allocation of them in the factors. Both synthetic and real data experiments show the benefits of the proposed approach. Finally, a deep object-centric learning-based fMRI data analysis method is proposed to estimate the variabilities of brain neural activation maps over fMRI volumes. The matrix factorization-based approaches such as the DL models assume a common subspace for the neural activation maps, thereby limiting the ability of capturing rich variabilities. The proposed method regards each map as an “object” with latent variables learned using an autoencoder with an attention mechanism. Learning efficient representations in the latent space encourages the model to learn a set of consistent maps across the fMRI volumes. The 3D convolutional neural network is utilized as the main building block of the proposed architecture. Experiments using both synthetic and real fMRI data sets verify the advantages of the proposed approach compared to existing matrix factorization-based methods.