Detection of step displacements in fMRI head motion data based on machine learning

Accurate artifacts detection in functional magnetic resonance imaging (fMRI) data is important in clinical applications and research. Subjects head motion remains the major source of fMRI artifacts, and retained head motion in the pre-processed fMRI data could also predict anthropomorphic, behavioral and clinical factors. However, an accurate characterization of subject head motion artifacts is lacking. We searched for step displacements in fMRI head motion data using machine learning approaches. Head motion data were defined using conventional six motion parameters as produced by fMRI realignment procedure. We created the semi-automatic markup tool to prepare head motion data for classification. This preparation was done using the sliding-window statistical anomaly detection and manual refinement of characteristic step artifacts. Marked up training dataset was used to train various classifiers to classify step-like head displacements. The best accuracy was achieved using neural and k-nearest neighborhood classifiers. Proposed approach could be used for an accurate detection of specific fMRI artifacts associated with head motions.