Challenges in Resting State Functional Connectivity Analysis: Removal of Head Motion Artifacts and Machine Learning-based Disease Classification

Abstract

In recent years, the use of resting-state functional Magnetic Resonance Imaging (rs-fMRI) for examining the brain function in healthy and clinical populations has increased drastically. Simultaneous modulations in neural activity between remote brain regions when the subject does not perform an explicit task, also called resting state functional connectivity (RSFC) are associated with the presence or absence of neurological disorders. However, some challenges remain to be addressed for widespread use of RSFC as a tool for disease classification. In this thesis, we address two crucial issues associated with RSFC. In the first part of this thesis, we examine how in-scanner head motion can cause artifactual changes in RSFC and evaluate the utility of an image based prospective motion correction in reducing the head motion artifacts in RSFC derived metrics. Our results indicate that the use of prospective motion correction combined with commonly used retrospective motion correction methods was able to visibly reduce the artifactual changes in RSFC. In the second part of this thesis, we examine the issues associated with the use of RSFC for disease diagnosis. Specifically, we evaluate how variations in age ranges and the data acquisition site of the sample can affect the performance of machine learning classifiers especially in heterogeneous disease populations with small sample sizes. We observe that the use of small, homogenous subject samples might give inflated measures of accuracy possibly due to overfitting. Finally, we recommend the use of a hold-out test data or a replication dataset to reproduce the classification performances to ensure good generalization across the disease population.